Short explanations of key AI concepts: what they mean, how they are used and where their limits lie.
List of terms
- A/B test
- Action
- ad asset
- ad auction
- ad group
- ad impression
- Ad Rank
- Agent evaluation
- Agent memory
- Agent role
- AI agent
- AI model
- anchor text
- Application programming interface (API)
- artificial intelligence
- attribution model
- audience segment
- Authentication
- Authorization
- Automation
- average daily budget
- backlink
- Baseline
- Bid
- campaign
- canonical URL
- Chunking
- click-through rate
- Condition
- Context engineering
- context window
- conversion
- conversion action
- conversion rate
- Conversion tracking
- conversion value
- conversion window
- Core Web Vitals
- cost per acquisition
- cost per click
- Cost per thousand impressions (CPM)
- crawling
- Cumulative Layout Shift
- data-driven attribution
- Dataset
- distillation
- embedding
- enhanced conversions
- entity
- Event
- few-shot prompting
- fine-tuning
- generative AI
- GPU
- groundedness
- Guardrail
- hallucination
- heading hierarchy
- hreflang
- HTTP status code
- Human in the loop
- Idempotency
- image alt text
- indexing
- inference
- Integration
- Interaction to Next Paint
- internal link
- JavaScript rendering
- JSON
- Key performance indicator (KPI)
- keyword match type
- keyword research
- Knowledge base
- landing page
- large language model
- Largest Contentful Paint
- Latency
- machine learning
- meta description
- Metric
- Model Context Protocol
- Model parameter
- Model training
- Multi-agent system
- multimodal model
- negative keyword
- Neural processing unit
- nofollow
- noindex
- Observability
- offline conversion import
- On-page SEO
- Orchestration
- Organic search result
- Parallelization
- pay-per-click advertising
- Performance Max
- Pipeline
- pretraining
- Prompt
- prompt engineering
- prompt injection
- Provenance
- Quality Score
- quantization
- ReAct
- redirect
- remarketing
- reranking
- responsive search ad
- Retrieval-augmented generation
- Retry with backoff
- return on ad spend
- rich result
- robots.txt
- Routing
- Sandbox
- Schema.org
- search engine optimization
- Search engine results page (SERP)
- search impression
- search intent
- search query
- Search ranking
- Semantic search
- Single source of truth (SSOT)
- Smart Bidding
- structured data
- System instruction
- target CPA
- target ROAS
- Technical SEO
- Telemetry
- Throughput
- title link
- token
- tool calling
- TOPS
- Transformer
- Trigger
- vector search
- Webhook
- Workflow
- Workflow step
- Workspace
- XML sitemap
A/B test
An A/B test is a controlled experiment comparing variants on predefined outcomes, typically by randomly assigning eligible users or units to the variants.
Hypothetical example
Eligible visitors are randomly split between two variants and compared using the same predefined conversion metric.
Limits
Check allocation, sample size, uncertainty and comparable observation windows; a before-and-after change alone does not establish a treatment effect.
Reader FAQ
Does any before-and-after comparison qualify?
No. Time changes can confound a simple before-and-after comparison.
Sources: MS-EXPERIMENT. Checked 2026-10-17.
Action
An action is a defined operation performed within a workflow step.
Hypothetical example
One action validates an input file before another processes it.
Limits
An action’s effects depend on its implementation and permissions.
Reader FAQ
Is an action the whole workflow?
Usually it is one operation within the larger process.
Sources: GH-ACTIONS. Checked 2026-10-17.
ad asset
An ad asset is a content component used to build or supplement an ad, such as a headline, description, image, sitelink or call information. The resulting ad format can combine several assets.
A practical example
A business could provide a reviewed headline and a link to its delivery information, with each destination kept current.
Limits
Do not assume every supplied asset appears in every impression. Review supported formats, eligibility and the actual assembled ad.
Reader FAQ
Is an asset the complete ad?
Not necessarily. It is one component; the displayed format can combine multiple components.
Further reading: ad asset.
ad auction
An ad auction determines which eligible ads may appear and their order for an advertising opportunity. In Google Ads, bids, quality, thresholds and context contribute to the decision.
A practical example
Two advertisers can enter the same search opportunity with different bids and relevance.
Limits
This entry uses Google Ads as its documented example. Other platforms have their own auction rules.
Reader FAQ
Does the highest bid always win?
No. Google Ads considers more than the bid, including ad and landing-page quality and search context.
Further reading: ad auction.
ad group
An ad group is a campaign component containing one or more ads with similar targeting. It helps organize related advertising messages and their targeting.
A practical example
A retailer separates two product themes into different ad groups within one campaign.
Limits
An ad group is distinct from the campaign. The documented structure here is Google Ads; do not assume identical controls on every platform.
Reader FAQ
Is an ad group the same as an individual ad?
No. It can contain multiple ads; the grouping organizes their shared advertising context.
Further reading: ad group.
ad impression
An ad impression is a recorded showing of an ad. In Google Ads it is counted when an ad appears on a search results page or another eligible Google Network location.
A practical example
An illustrative campaign report could show many impressions but fewer clicks; those are different stages of the interaction.
Limits
An impression is not a distinct person or a conversion. Use a separate reach measure when the question concerns unique people.
Reader FAQ
Can several impressions come from the same person?
An impression counts a showing, not a unique individual. Do not substitute impression totals for unique audience reach.
Further reading: ad impression.
Ad Rank
Ad Rank is Google’s set of auction values used to decide whether an ad can appear and its position relative to other eligible ads. Bids, ad and landing-page quality, thresholds, competition, search context and expected asset impact contribute.
A practical example
An illustrative advertiser could improve the relevance of its ad and destination while reviewing bids, rather than treating a higher bid as the only lever.
Limits
Ad Rank is auction dependent. Do not reduce it to a universal fixed formula or infer a guaranteed placement from one account diagnostic.
Reader FAQ
Does the highest bid always win the best position?
No. Google considers several factors besides the bid, including auction-time quality and eligibility thresholds.
Further reading: Ad Rank.
Agent evaluation
Agent evaluation tests an agent system against defined tasks and criteria. It can examine the final result and the path used to obtain it. OpenAI’s evaluation guidance describes traces and structured grading for workflow-level issues.
A practical example
For a document-review agent, prepare cases with a correct source, conflicting sources and a missing document. Check its answer, selected tool, source identity and response to missing evidence.
Limits
A successful demonstration is one observation. Broader readiness requires representative cases, explicit failure criteria and review of the exact system version. A grader’s score is an assessment method, not an independent guarantee of truth or safe permissions.
Reader FAQ
Why test the path as well as the answer?
A plausible answer can hide an unsupported source choice or an inappropriate action. Inspecting the trajectory can reveal the first failure.
When should evaluation be repeated?
Revisit affected cases after material changes to prompts, models, tools, sources or permissions.
Agent memory
Agent memory is information stored and retrieved by an agent system beyond the immediate interaction, for later use.
Hypothetical example
An assistant retains an approved project preference for a later session.
Limits
Storage, access, correction and retention depend on the application; memory is not necessarily model training.
Reader FAQ
Is memory the same as the context window?
No. Stored information must be made available to the model when needed.
Sources: ANTH-MEM. Checked 2026-10-17.
Agent role
An agent role describes the responsibilities, task boundaries and expected outputs assigned to an agent within an application or team.
Hypothetical example
A source-checking agent reviews evidence while an editor revises prose.
Limits
A role description alone does not enforce permissions or guarantee useful specialization.
Reader FAQ
Is a role a different model?
No. The same underlying model can be used in different assigned roles.
Sources: ANTH-ROLE. Checked 2026-10-17.
AI agent
In the language-model systems discussed here, an AI agent is a system in which a model helps choose the next step toward a goal, uses permitted tools and responds to what those tools return. It operates through a sequence of decisions rather than one isolated answer. Anthropic’s architecture discussion distinguishes model-directed agents from workflows with predefined paths.
A practical example
A research assistant could search an approved document collection, inspect a source, notice that it describes the wrong edition and choose another search. The model selects a next action; surrounding software controls the available tools and permissions.
What the term does not establish
Calling something an agent says little about reliability or authority. Specify its allowed actions, stopping condition and failure response separately. A repeatable fixed process may need no model-directed loop.
Reader FAQ
Can an agent work only in read-only mode?
Yes. It can select searches and inspections without receiving tools that change records. Read-only access still needs privacy boundaries.
Is an agent the same as a chatbot?
A chat interface can contain an agent, but the interface alone does not establish an agent architecture.
AI model
In machine learning, an AI model is a learned representation used to produce predictions or generated content from input data.
Hypothetical example
A trained classifier assigns a topic to a request.
Limits
A model is one component; an agent application also needs surrounding logic and tools.
Reader FAQ
Is every AI model a language model?
No. Models can address different prediction or generation tasks.
Sources: GOOGLE-MODEL. Checked 2026-10-17.
anchor text
Anchor text is the visible text of a link. It gives readers and search engines context about the destination before they follow it.
A practical example
In a sentence explaining a measurement, ‘conversion rate’ could be the link text pointing to its relevant definition.
Limits
Useful link text is descriptive and fits the surrounding sentence. Repeating keywords or using misleading descriptions does not make a link more helpful.
Reader FAQ
Does a linked image have text that serves a similar purpose?
Yes. Google uses the image’s alt text as anchor text when the image acts as a link, so its function and destination matter.
Further reading: anchor text — 1; anchor text — 2.
Application programming interface (API)
An API is a defined interface through which software components request capabilities or exchange information.
Hypothetical example
A reporting application requests campaign data through a documented interface.
Limits
An API is an interface, not automatic permission to access every resource.
Reader FAQ
Is an API a user interface?
It is designed for software interaction rather than direct human interaction.
Sources: MDN-API. Checked 2026-10-17.
artificial intelligence
Artificial intelligence, or AI, is a broad field of technologies for performing tasks associated with intelligence, such as interpreting language, recognizing patterns, reasoning or making predictions. The term covers multiple approaches rather than one model or product.
A practical example
An illustrative team could use a classifier to route requests and a generative model to draft responses; both applications can belong to AI while doing different tasks.
Limits
The AI label does not establish accuracy, autonomy or suitability for a particular job. Evaluate the actual system and task rather than treating the category name as a performance claim.
Reader FAQ
Is AI the same as machine learning?
No. Machine learning is an approach within AI that learns from data; AI is the broader field.
Further reading: artificial intelligence — 1; artificial intelligence — 2.
attribution model
An attribution model assigns conversion credit to eligible interactions in a measured customer path. Different models can allocate the same conversion differently.
A practical example
Last-click attribution can assign the credit to the final eligible ad click; a data-driven model can distribute credit.
Limits
Credit depends on measured interactions and settings. Attribution does not itself prove that an ad caused an incremental sale.
Reader FAQ
Why can two reports disagree?
First compare eligible channels, model and conversion settings. Different measurement boundaries can produce different allocations.
Further reading: attribution model.
audience segment
An audience segment groups people by characteristics such as interests, intent, demographics or prior engagement, as defined or estimated by the advertising system. Its role varies with campaign settings.
A practical example
A team could distinguish visitors who previously engaged with its website from a broader in-market segment.
Limits
A segment label is not proof of an individual person’s circumstances. Check whether it is used for targeting, observation or an automation signal.
Reader FAQ
Does adding a segment always restrict who sees an ad?
No. Its effect depends on the campaign and setting; Performance Max audience signals are not a strict audience boundary.
Further reading: audience segment.
Authentication
Authentication verifies a claimed identity using evidence that the receiving system can check.
Hypothetical example
A service verifies the identity of an application requesting data.
Limits
Proving identity does not by itself grant access to every resource.
Reader FAQ
Does signing in establish all permissions?
No. Access permissions require a separate authorization decision.
Sources: MDN-AUTHN. Checked 2026-10-17.
Authorization
Authorization determines what an identified user or workload is allowed to access or do.
Hypothetical example
An application may read reporting data while lacking permission to change campaigns.
Limits
Permissions are resource- and action-specific and can change.
Reader FAQ
How is authorization different from authentication?
Authentication checks identity; authorization checks allowed access.
Sources: GC-AUTHZ. Checked 2026-10-17.
Automation
Automation uses technology to perform tasks with reduced manual intervention, following configured rules or logic. For a project-specific application, see the DIOS production experiment.
Hypothetical example
A scheduled process retrieves yesterday’s campaign totals.
Limits
Automation does not by itself imply an AI agent, correctness or permission to act.
Reader FAQ
Is automation the same as AI?
No. Rule-based automation can work without an AI model.
Sources: IBM-AUTO. Checked 2026-10-17.
average daily budget
An average daily budget is Google’s per-day average spending setting for a campaign over a month. It is a pacing amount, not a promise that actual spending is identical each day.
A practical example
A team could choose an average amount for its campaign and review the budget report for daily variation and the applicable spending limits.
Limits
Google documents that some days may spend above the average and others below it. Verify campaign-specific limits before treating the setting as a hard daily cap.
Reader FAQ
Is average daily budget the maximum charge on every day?
No. Google distinguishes the average budget from daily and monthly spending limits.
Further reading: average daily budget.
backlink
A backlink is a link from another website to a page on your website. It describes the direction of the link from the receiving site’s perspective.
A practical example
If an independent publication links to a glossary entry as a useful reference, that entry has a backlink from the publication.
Limits
A reported link may no longer be live. Search Console’s Links report is a sample, so its absence from that report does not prove that no backlink exists.
Reader FAQ
Is every link to another page a backlink?
From the receiving site’s perspective, a link from another site is a backlink. A link between pages on the same site is an internal link.
Further reading: backlink.
Baseline
A baseline is a defined reference result or simple solution against which another result is assessed.
Hypothetical example
A new model is compared with a simple classifier on the same evaluation set.
Limits
Comparisons require compatible metrics, data and conditions; an old number may be an unsuitable reference.
Reader FAQ
Must a baseline be a complex model?
No. A simple solution can provide a useful comparison.
Sources: GOOGLE-BASELINE. Checked 2026-10-17.
Bid
A bid specifies what an advertiser offers or is willing to pay for a defined advertising outcome, such as a click or a thousand viewable impressions.
Hypothetical example
An advertiser sets a maximum CPC bid for a click.
Limits
A bid and the actual amount charged are different concepts; bidding rules vary by platform and strategy.
Reader FAQ
Is a bid the campaign budget?
No. A bid concerns an outcome; a budget constrains spending.
Sources: GOOGLE-BID. Checked 2026-10-17.
campaign
In Google Ads, a campaign organizes advertising under shared settings such as budget and location targeting. Its internal structure depends on campaign type; Search campaigns commonly organize ads and keywords into ad groups.
A practical example
An advertiser could separate two markets when each needs its own budget and location settings.
Limits
Campaign is a platform object as well as a broad marketing word. Do not assume every campaign type exposes identical structures or controls.
Reader FAQ
Is a campaign the same as an ad group?
No. An ad group belongs within a campaign; campaign-level settings govern a broader scope.
Further reading: campaign — 1; campaign — 2.
canonical URL
A canonical URL identifies a preferred representative of duplicate or very similar pages. Google can use signals such as redirects, canonical annotations and sitemap inclusion in its selection.
A practical example
An owner indicates a preferred product-page URL among near-duplicate variants.
Limits
A canonical annotation is a signal, not an unconditional instruction. Keep its purpose separate from language alternates.
Reader FAQ
Does declaring a canonical force Google to choose it?
No. Google evaluates canonicalization signals and may select a different representative.
Further reading: canonical URL.
Chunking
Chunking splits material into smaller units for processing, indexing or retrieval.
Hypothetical example
A long guide is divided into sections that can be retrieved separately.
Limits
Size and overlap affect context and retrieval; no single size is best for every collection.
Reader FAQ
Why use overlap between chunks?
It can preserve context that crosses a split, while adding duplication.
Sources: AZ-CHUNK. Checked 2026-10-17.
click-through rate
Click-through rate, or CTR, is clicks divided by impressions, usually expressed as a percentage. In an ad report it describes how often recorded impressions led to clicks.
A practical example
An illustrative ad with 12 clicks and 600 impressions has a CTR of 2%.
Limits
Compare like-for-like placements and reporting definitions. CTR describes clicking, so it does not by itself establish sales, lead quality or profitability.
Reader FAQ
Can a high CTR coexist with poor conversions?
Yes. The metric counts clicks relative to impressions; the later conversion outcome is a separate measure.
Further reading: click-through rate.
Condition
A condition is a test that determines which path a workflow follows.
Hypothetical example
A workflow sends incomplete records for review and complete records for analysis.
Limits
The test is only as reliable as its input and rules.
Reader FAQ
Can a condition pause automatic processing?
Yes. A chosen branch can require human review.
Sources: AWS-WF. Checked 2026-10-17.
Context engineering
Context engineering is the work of selecting and maintaining the information available to a language model at each step. It includes instructions, retrieved evidence, tool descriptions and relevant history. Anthropic’s explanation treats this as an ongoing task across a model’s working context.
A practical example
During a long research task, keep the current objective and evidence references available. Retrieve a detailed document when needed rather than repeating every document in every request. Preserve decisions in durable records so a compact working context can still point back to their sources.
Limits
Removing irrelevant material can help focus, but a summary that drops a qualification or source can damage the task. Review what is retained and what is omitted.
Reader FAQ
How is it different from prompt engineering?
Prompt engineering focuses on instructions; context engineering also manages other information entering the model’s context and its evolution during work.
Is a larger context window sufficient?
Capacity alone does not select authoritative, relevant or current information.
context window
A context window is the token capacity available for a model’s working context in a request. It limits how much information can be considered together; the exact usable allocation is model and application specific.
A practical example
For an illustrative document task, the application could select relevant passages rather than placing every file into the request.
Limits
Capacity does not establish that every included fact is current or useful. Check the actual model’s allocation and reserve before relying on a quoted window size.
Reader FAQ
Does a larger context window replace context engineering?
No. Capacity and information selection are separate questions; the latter still needs an authority and relevance review.
conversion
A conversion is a measured action selected as valuable for an advertising objective, such as a purchase, registration or phone call.
A practical example
A campaign can measure a submitted enquiry as a conversion while a separate sales system records whether it becomes a sale.
Limits
A conversion count is not automatically a count of customers or completed purchases. Define the action and measurement boundary.
Reader FAQ
Must every conversion be a purchase?
No. The advertiser chooses valuable actions; preserve what each action actually measures in reporting.
Further reading: conversion.
conversion action
A conversion action is a specifically defined customer action that an advertiser treats as valuable, such as a purchase or a qualified call. Recorded occurrences are conversions; the action is the configured event definition.
A practical example
A team could keep a newsletter signup and a purchase as separate actions because they serve different purposes.
Limits
Inspect source, counting rules, goals and primary or secondary status. A tracked event is not automatically a campaign’s bidding objective.
Reader FAQ
Are all conversion actions used for bidding?
No. Goal and action settings determine bidding use; Google also documents an exception for secondary actions included in a custom goal.
Further reading: conversion action — 1; conversion action — 2.
conversion rate
Conversion rate is conversions divided by the eligible interaction count, expressed as a percentage. Google’s ad definition uses trackable ad interactions in the same period; a website report may use another denominator.
A practical example
In an illustrative click-based report, 8 conversions from 200 eligible interactions gives a 4% conversion rate.
Limits
State the event, denominator, counting and reporting scope. Comparing an ad interaction rate with a session-based website rate can mislead.
Reader FAQ
Is conversion rate always conversions divided by website sessions?
No. The denominator depends on the reporting system and metric definition; Google’s ad metric uses eligible ad interactions.
Further reading: conversion rate.
Conversion tracking
Conversion tracking measures configured valuable actions after advertising interactions, such as purchases, calls or sign-ups.
Hypothetical example
A configured purchase action records eligible sales attributed under the chosen settings.
Limits
Coverage depends on implementation, consent and settings; reported conversions are not automatically all business revenue.
Reader FAQ
Is tracking the same as choosing a conversion action?
The action defines what matters; tracking implements its measurement.
Sources: GOOGLE-CONV. Checked 2026-10-17.
conversion value
Conversion value is the assigned value associated with a measured conversion. It can represent transaction-specific revenue or another explicitly defined business-value basis.
A practical example
An illustrative store could send the order value for each purchase rather than assigning the same default amount to every order.
Limits
Document currency, value basis and duplicates or adjustments. An assigned value is not automatically verified revenue or profit.
Reader FAQ
Can two actions have the same count but different total value?
Yes. Their assigned values can differ, so conversion count and conversion value answer different questions.
Further reading: conversion value.
conversion window
A conversion window is the interval after an eligible ad interaction during which a conversion can be recorded for that action. In Google Ads, the configured window is expressed in days.
A practical example
An illustrative purchase occurring ten days after a click could be within a thirty-day window but outside a seven-day window.
Limits
State the interaction type and action settings when comparing reports. Do not treat a Google Ads conversion window as identical to every analytics lookback setting.
Reader FAQ
Does changing the window change what a conversion means?
The business action can stay the same, while the eligible timing and reported count differ. Keep those two definitions separate.
Further reading: conversion window.
Core Web Vitals
Core Web Vitals are metrics for real-world loading performance, responsiveness and visual stability. The current set is Largest Contentful Paint, Interaction to Next Paint and Cumulative Layout Shift.
A practical example
A team checks field experience for a page class and investigates slow loading or unstable layout.
Limits
Distinguish field observations from lab tests. Good metric values alone do not guarantee a particular search ranking.
Reader FAQ
Is a single lab score the whole assessment?
No. Record the measurement context and inspect real-user experience as well as diagnostic lab observations.
Further reading: Core Web Vitals.
cost per acquisition
Cost per acquisition describes cost relative to a defined acquisition outcome. In Google Ads, CPA means cost per action: average CPA summarizes conversion cost divided by conversion count, so the chosen action determines what that metric measures.
A practical example
In an illustrative lead report, 300 cost units for 15 qualified lead conversions gives a CPA of 20 units for that defined action.
Limits
A lead is not automatically a new customer. Label the event, cost basis and attribution scope; do not silently call an ad action metric customer acquisition cost.
Reader FAQ
Is CPA necessarily the cost of acquiring one new customer?
No. In Google Ads it may refer to another selected action, such as a lead; the denominator must be stated.
Further reading: cost per acquisition — 1; cost per acquisition — 2.
cost per click
Cost per click, or CPC, is the amount paid for an ad click. Average CPC is total click cost divided by the number of clicks; actual CPC and a maximum CPC bid describe different values.
A practical example
In an illustrative report, 80 currency units spent on 40 clicks gives an average CPC of 2 units.
Limits
A cheaper click does not establish a better business outcome. Compare the same cost basis, currency and reporting scope, then inspect conversion quality.
Reader FAQ
Is average CPC the same as my maximum bid?
No. The average summarizes actual click costs; the maximum bid is a bidding setting.
Further reading: cost per click — 1; cost per click — 2.
Cost per thousand impressions (CPM)
CPM expresses advertising cost per thousand impressions and can describe a pricing model or a reporting metric.
Hypothetical example
For comparable counts, divide spend by impressions and multiply by 1,000.
Limits
Impressions are not unique people, clicks or conversions; viewable CPM uses a different impression basis.
Reader FAQ
Can two CPM figures always be compared directly?
Only when their impression basis, scope and currency are compatible.
Sources: AMAZON-CPM. Checked 2026-10-17.
crawling
Crawling is the process in which an automated crawler discovers and fetches web resources. For Google Search, it is distinct from indexing the retrieved information and selecting results for a search query.
A practical example
A crawler could follow a link from a published guide to a glossary entry and request that entry’s URL.
Limits
Discovery does not guarantee a fetch, and a successful fetch does not guarantee indexing or visibility in search results.
Reader FAQ
Does a crawl prove that a page is indexed?
No. Crawling retrieves resources; indexing is a separate stage that may or may not include them.
Further reading: crawling.
Cumulative Layout Shift
Cumulative Layout Shift, or CLS, is a Core Web Vital measuring unexpected visual movement of page content. Its current calculation uses the largest burst of layout-shift scores within a session window rather than adding every shift across the whole visit.
A practical example
An image appearing without reserved space could push text downward while someone is reading it.
Limits
CLS is a unitless score, not a loading duration. A quick page can still be unstable, and one lab run may miss shifts experienced during real visits.
Reader FAQ
What is the good CLS threshold?
The documented threshold is 0.1 or less at the 75th percentile of page loads, assessed separately for mobile and desktop.
Further reading: Cumulative Layout Shift.
data-driven attribution
Data-driven attribution uses conversion-path data to allocate credit among eligible ad interactions. Google’s model uses account data to assess how interactions contribute within the measured path.
A practical example
A report could assign part of the credit to an earlier ad interaction rather than giving the entire conversion to the final one.
Limits
Attributed credit is a model-based allocation within the observed data. Do not treat it alone as experimental proof of incremental business impact.
Reader FAQ
Is data-driven attribution just last-click attribution?
No. It can distribute credit among interactions rather than assigning all credit to the last eligible interaction.
Further reading: data-driven attribution.
Dataset
A dataset is a collection of data used for analysis, training or evaluation.
Hypothetical example
A collection contains labeled support requests.
Limits
Representativeness and quality require checking.
Reader FAQ
Must a dataset be a spreadsheet?
No. Data can have other formats.
Sources: GOOGLE-ML. Checked 2026-10-17.
distillation
Knowledge distillation transfers a teacher model’s behavior to a student by using the teacher’s outputs as a training signal. The original research explored compressing knowledge from an ensemble into a model that was easier to deploy.
A practical example
A teacher could provide scores for training examples, helping train a smaller student for a narrowly defined task.
Limits
Evaluate the student on the intended workload. Distillation is a training method, not a promise that every teacher capability or quality level is preserved.
Reader FAQ
Is distillation the same as quantization?
No. Distillation trains a student from a teacher’s signal; quantization concerns the numerical representation used in a model.
Further reading: distillation — 1; distillation — 2.
embedding
An embedding is a vector representation of data. Positions and distances in the model’s representation can encode relationships useful for comparison.
A practical example
A search system compares a question’s representation with document representations to select candidates.
Limits
Similarity in that representation does not prove that two claims mean exactly the same thing or are factually correct.
Reader FAQ
Is an embedding a readable summary?
No. Its numbers are a representation; the dimensions may not have an interpretation understandable to a reader.
Further reading: embedding.
enhanced conversions
Enhanced conversions is a Google measurement feature that supplements conversions with hashed first-party customer data. It can apply to website conversions and to imported lead outcomes, depending on the implementation.
A practical example
A permitted measurement setup could use customer-provided information to help associate an eligible conversion with an ad interaction.
Limits
Hashing is a transformation, not blanket permission to collect or disclose data. Review applicable customer-data policies, permissions and implementation separately.
Reader FAQ
Is this the same as creating a new conversion action?
No. It supplements measurement; the business event and its conversion-action definition still need to be specified.
Further reading: enhanced conversions — 1; enhanced conversions — 2.
entity
An entity, in semantic search and structured-data contexts, is a distinguishable thing or concept that can be described and related to other things. Its identity helps distinguish the subject from the words used to name it.
A practical example
Two organizations could share a similar name while remaining different entities with different identities and relationships.
Limits
A matching name alone does not establish identity. This meaning of entity is broader than a legal entity and does not promise a particular Google display or ranking.
Reader FAQ
Is an entity just a keyword?
No. A keyword is a word or phrase; an entity is the subject it may refer to. The same phrase can refer to different subjects.
Further reading: entity.
Event
In automation, an event is a recorded occurrence or change that a system can communicate to other components.
Hypothetical example
A record-approved event can start a reporting workflow.
Limits
An event and the action taken in response to it are different things.
Reader FAQ
Is an event the same as a trigger?
An event may be used as a trigger when a workflow subscribes to it.
Sources: AWS-EVENT. Checked 2026-10-17.
few-shot prompting
Few-shot prompting includes several examples of the desired response within a prompt. It uses the model’s existing abilities to follow a demonstrated pattern rather than performing parameter training.
A practical example
A prompt could show two approved examples of message-to-category output and then ask for the same format on a new message.
Limits
Examples should reflect the intended task, including difficult cases. More examples are not a substitute for checking the result.
Reader FAQ
Does few-shot prompting change model weights?
No. Supplying examples in a prompt does not itself update the model’s parameters.
Further reading: few-shot prompting.
fine-tuning
Fine-tuning adapts a pretrained model through additional training on task-specific data. It changes learned parameters; parameter-efficient methods can update a limited subset rather than all existing weights.
A practical example
A team could train a base model on reviewed examples of its document categories, then evaluate the adapted model on held-out cases.
Limits
Useful training examples do not replace independent evaluation. Editing the wording of a prompt is a different operation from parameter fine-tuning.
Reader FAQ
Does changing a prompt fine-tune the model?
No. Prompt engineering uses the existing model; fine-tuning involves a training step that updates parameters.
generative AI
Generative AI produces content, such as text, images or audio, using patterns learned during training. Generation is a capability category; it does not identify one particular model architecture.
A practical example
For example, a team could ask a model to draft alternative descriptions for an approved product brief, then review them before use.
Limits
Generated content needs factual and rights review. A fluent new description is not evidence that its product claims are true.
Reader FAQ
Is generative AI limited to chatbots?
No. Its outputs can include several kinds of content, and a chat interface is only one way to use it.
Further reading: generative AI.
GPU
A graphics processing unit, or GPU, is a processor associated with graphics and highly parallel computation. GPUs can execute many operations concurrently; supported software also uses them for machine-learning workloads.
A practical example
A training system could run parallel numerical operations on a compatible GPU while other work remains on the CPU.
Limits
Performance depends on the workload, memory and software route. A processor category or headline specification alone does not establish application speed.
Reader FAQ
Does a GPU’s presence prove a model will run well?
No. Check the model’s resource needs and compatible runtime, then measure the actual workload.
Further reading: GPU.
groundedness
Groundedness describes how well an answer’s claims are supported by specified reference material. It evaluates the relationship between the response and those sources rather than just whether the response sounds plausible.
A practical example
A reviewer could check whether each statement in a policy summary is supported by the supplied current policy.
Limits
Source support and real-world truth are separate checks: an answer can accurately repeat an outdated source. Review the source’s authority and freshness as well.
Reader FAQ
Does a citation alone prove groundedness?
No. The cited material must support the particular claim, including its qualifications; a nearby source link is insufficient.
Further reading: groundedness.
Guardrail
A guardrail is a check or constraint applied to an AI system’s inputs, retrieved context, actions or outputs to help enforce a policy.
Hypothetical example
An execution check blocks an action outside an approved scope.
Limits
Guardrails can fail and require evaluation; they are not a guarantee of truthful or safe output.
Reader FAQ
Is a guardrail only an output filter?
No. Checks can apply at several stages of a system.
Sources: NVIDIA-RAILS. Checked 2026-10-17.
hallucination
In generative AI, hallucination means plausible-looking output that makes a factually incorrect assertion about the real world.
A practical example
An answer could name a documentation section that does not exist. Check the actual source before relying on it.
Limits
This entry concerns model output, not a medical use of the word. Fluency is insufficient evidence for a factual claim.
Reader FAQ
Is fictional writing always hallucination?
Clearly requested fiction makes a different kind of claim. Here the term concerns output presented as a real-world assertion.
Further reading: hallucination.
heading hierarchy
Heading hierarchy is the organization of content through ranked headings, such as h1 through h6. The ranks express sections and subsections and help people, including screen-reader users, understand and navigate the page.
A practical example
A guide could use an h2 for ‘Measurement’ and an h3 for the ‘Attribution models’ subsection beneath it.
Limits
Choose ranks to represent the content structure. A larger font alone does not give ordinary text the semantics of a heading.
Reader FAQ
Should heading ranks be chosen for their visual size?
No. Rank expresses structure; styling can control appearance without changing that structure.
Further reading: heading hierarchy.
hreflang
hreflang is an annotation that identifies language or regional variants of a page for Google Search. It helps associate equivalent localized pages so Google can select a suitable version for a user.
A practical example
An English article and its Ukrainian translation could identify each other as language variants.
Limits
Annotations should point to the correct equivalents and include reciprocal return links. hreflang does not translate content or replace editorial checks.
Reader FAQ
Does hreflang create a translation?
No. It describes the relationship between versions that already exist; their content and metadata still need translation and review.
Further reading: hreflang.
HTTP status code
An HTTP status code is a server’s numeric response indicating the outcome of a request. Codes distinguish success, redirection, client errors and server errors, giving browsers and crawlers information about how to handle a resource.
A practical example
A request might return 200 for a successful page response, 301 for a permanent redirect or 404 when the requested resource is not found.
Limits
A successful status does not establish content quality or guarantee indexing. The response body and destination behavior also matter.
Reader FAQ
Does a 200 response guarantee that Google indexes the page?
No. It allows the received content to proceed for consideration; Google can still decide not to index it.
Further reading: HTTP status code.
Human in the loop
Human in the loop means that a person participates in a workflow, for example by reviewing a result or approving a consequential step.
Hypothetical example
A person reviews a draft before the process can publish it.
Limits
The intervention point and decision authority must be explicit.
Reader FAQ
Does human review have to happen after every step?
No. It can be required at selected decision points.
Sources: AWS-WF. Checked 2026-10-17.
Idempotency
Idempotency means repeating the same logical request has no additional intended side effects beyond the first successful application.
Hypothetical example
A repeated request identifier prevents a second copy of the same order.
Limits
The service must define request identity and retention rules; identical responses are not the defining requirement.
Reader FAQ
Why does idempotency matter for retries?
A timeout can leave the caller unsure whether the first request succeeded.
Sources: AWS-IDEM. Checked 2026-10-17.
image alt text
Image alt text is a text alternative that conveys an image’s relevant information or function when the image cannot be perceived. The appropriate alternative depends on the image and its context.
A practical example
For an image that acts as a link to a report, the alternative could describe the link’s destination or action rather than the image’s colors.
Limits
Decorative images can use an empty alternative. Complex diagrams may need a fuller text explanation in addition to brief alt text.
Reader FAQ
Should every image get a detailed visual description?
No. Describe the useful information or function. Decorative and functional images need different treatment.
Further reading: image alt text.
indexing
Indexing is the search-engine stage that analyzes crawled content and may store information about it in the search index.
A practical example
A search engine processes a newly crawled article and assesses its content and relationship to similar pages.
Limits
Crawling, indexing and displaying a result are distinct stages. Google does not guarantee that every processed page is indexed.
Reader FAQ
Does an indexed page appear for every relevant query?
No. Serving results also depends on the query and other relevance and quality signals.
Further reading: indexing.
inference
Inference is using a trained model to produce predictions from an input. A production system may compute predictions in advance or when a request arrives.
A practical example
A classifier predicts a topic for a new document; a language model predicts output tokens for a prompt.
Limits
Inference is distinct from training model parameters. Request-time execution does not itself establish accuracy.
Reader FAQ
Must inference happen live?
No. Batch inference can compute outputs ahead of a later request; dynamic inference computes them at request time.
Further reading: inference.
Integration
An integration connects applications or services so they can exchange information or participate in a shared process.
Hypothetical example
A reporting tool combines an API request with event notifications.
Limits
A connection does not make field meanings, permissions or data quality automatically compatible.
Reader FAQ
Can integrations use both APIs and webhooks?
Yes. They can request data through APIs and react to webhook events.
Sources: GH-INTEGRATION. Checked 2026-10-17.
Interaction to Next Paint
Interaction to Next Paint, or INP, is a Core Web Vital measuring page responsiveness from click, tap and keyboard interactions during a visit. It represents a slow interaction’s delay until the browser can present the next frame, with an allowance for outliers.
A practical example
A delayed response after pressing an accordion button could contribute to poor INP.
Limits
INP is not the mean of all interaction delays and does not measure the full completion time of a later network request. A visit without qualifying interactions may have no INP value.
Reader FAQ
What is the good INP threshold?
The documented threshold is 200 milliseconds or less at the 75th percentile of field page loads, assessed separately for mobile and desktop.
Further reading: Interaction to Next Paint.
internal link
An internal link connects one page to another within the same site. Contextual links and descriptive anchor text help readers and search engines find related content.
A practical example
An explanatory article links to the site’s definition of a technical term at a useful mention.
Limits
Link to an actual relevant destination. Repeating links or inventing destinations does not add useful context.
Reader FAQ
How many internal links should every page have?
There is no universal ideal number. Choose links that help the reader understand the content and reach relevant pages.
Further reading: internal link.
JavaScript rendering
JavaScript rendering is the processing of a page’s scripts to produce or update its displayed document. For Google Search, rendering can reveal content and links that were not present in the initial HTML response.
A practical example
A glossary interface could receive an HTML shell and add its visible entries after a script runs.
Limits
Rendering depends on accessible resources and executable behavior. Google’s ability to render does not prove that every crawler or user receives the same content.
Reader FAQ
Is initial HTML always identical to the rendered page?
No. Scripts can add or change content, so examining only the initial response may miss what is rendered.
Further reading: JavaScript rendering.
JSON
JSON is a text format for exchanging structured data using objects, arrays, strings, numbers, booleans and null.
Hypothetical example
A response contains a campaign name and a list of daily totals.
Limits
Valid JSON does not guarantee that its values are complete or correct.
Reader FAQ
Is JSON executable code?
JSON is a data format; it does not encode functions as native values.
Sources: MDN-JSON. Checked 2026-10-17.
Key performance indicator (KPI)
A KPI is a defined measurement used to assess progress or performance in an activity important to an objective.
Hypothetical example
A team tracks qualified leads against an agreed target.
Limits
Specify the formula, unit, time window and scope; not every available metric is a useful KPI.
Reader FAQ
How is a KPI different from a metric?
A KPI is selected because it matters to a particular objective.
Sources: IBM-KPI. Checked 2026-10-17.
keyword match type
A keyword match type controls how a search keyword can match the queries that make an ad eligible. Google Ads uses broad, phrase and exact match.
A practical example
A phrase-match keyword can cover qualifying searches containing its meaning rather than a single fixed string.
Limits
Matching establishes eligibility, not guaranteed display. Negative keywords are a separate exclusion control.
Reader FAQ
Does exact match require identical characters?
No. Google Ads exact match can include searches with the same meaning or intent.
Further reading: keyword match type.
keyword research
Keyword research is the process of investigating the words, questions and topics an audience uses when seeking information, then using that evidence to plan useful content. It considers differences in wording and familiarity with the subject.
A practical example
A publisher could compare beginner questions with specialist terminology before deciding which explanations a glossary should provide.
Limits
A list of terms does not establish demand, intent or ranking potential by itself. Tool estimates need their own source and date; these drafts do not claim measured search volumes.
Reader FAQ
Does a page need every exact keyword variation?
No. Google can relate a page to different queries. Write clearly for the audience rather than inserting every possible variation.
Further reading: keyword research.
Knowledge base
A knowledge base is an organized collection of information that an application can search or retrieve to support its responses.
Hypothetical example
An assistant retrieves an approved procedure before answering a question.
Limits
Stored information can be stale or incomplete; retrieval does not automatically verify truth.
Reader FAQ
Is a knowledge base the model’s learned weights?
No. A retrieval-based application consults separately stored information.
Sources: AWS-KB. Checked 2026-10-17.
landing page
In advertising, a landing page is the webpage reached after an ad click. It should be considered together with the ad’s destination and the purpose of the visit.
A practical example
An illustrative ad for a specific course could lead to that course’s information rather than a generic site homepage.
Limits
Check the actual destination and user journey. An attractive page cannot compensate for unsupported claims or a mismatch with the ad.
Reader FAQ
Does a landing page have to be the homepage?
No. It can be a specific destination suited to the advertised offer.
Further reading: landing page.
large language model
A large language model is a trained language model that predicts tokens or token sequences using patterns learned from substantial training data. In text generation, those predictions form a response to supplied context.
A practical example
A writing assistant can propose a paragraph from a brief; that paragraph still needs factual review.
Limits
The label does not establish truth, current knowledge, permissions or a universal minimum parameter count.
Reader FAQ
Does an LLM search the web whenever it answers?
The label alone does not establish retrieval or browsing. Check the actual system and its enabled tools.
Further reading: large language model.
Largest Contentful Paint
Largest Contentful Paint, or LCP, measures when the largest eligible visible content element is rendered in the viewport, relative to the start of navigation. It is a Core Web Vital concerned with perceived loading speed.
A practical example
A large article image could become the LCP element when it appears; another eligible element could have been the largest earlier in loading.
Limits
LCP does not mean every resource has loaded. Field and lab results can differ because real navigation, devices and users differ.
Reader FAQ
What is the good LCP threshold?
The documented threshold is 2.5 seconds or less, assessed at the 75th percentile of page loads, separately for mobile and desktop.
Further reading: Largest Contentful Paint.
Latency
Latency is the elapsed time between a request or event and the response or outcome being measured.
Hypothetical example
A test measures the time from submitting a query to receiving results.
Limits
State the start and end points; one timing does not describe all users or requests.
Reader FAQ
Is latency the same as throughput?
No. Latency measures delay, while throughput measures work completed per unit of time.
Sources: GOOGLE-RANK. Checked 2026-10-17.
machine learning
Machine learning is a way to train a model from data so it can make predictions or generate outputs. Learned relationships help determine its behavior rather than a developer writing a separate rule for every case.
A practical example
An illustrative classifier could learn from labeled support messages to suggest which team should receive a new request.
Limits
Training data and evaluation cases need to fit the intended task. A convincing prediction does not establish accuracy on unfamiliar inputs.
Reader FAQ
Is every automated rule machine learning?
No. A fixed rule can automate a decision without learning a model from data.
Further reading: machine learning.
meta description
A meta description is a short page summary supplied in an HTML meta tag. Google may use it to create a search-result snippet when it provides a suitable description for the query.
A practical example
A glossary page could summarize the term’s meaning and the distinction its explanation covers.
Limits
Google primarily generates snippets from page content and may select different text for different queries. A supplied description is not a guaranteed displayed snippet.
Reader FAQ
Is the meta description the same as the visible search snippet?
No. It is one possible source for the snippet; Google can use other page text.
Further reading: meta description.
Metric
A metric is a defined quantitative measure used to describe behavior, performance or outcomes.
Hypothetical example
A system tracks request count or error rate over a specified interval.
Limits
State units, calculation and scope before comparing values.
Reader FAQ
Is every metric a KPI?
No. A KPI is selected for its importance to a specific objective.
Sources: OTEL. Checked 2026-10-17.
Model Context Protocol
Model Context Protocol, or MCP, is an open standard for connecting AI applications to external systems. Connections can expose data, tools or reusable prompts through a common protocol. The official introduction describes these integration roles.
A practical example
An application could use an MCP connection to search an authorized knowledge collection. Another server might expose an action that changes a business record. A shared connection format does not make those operations equivalent in consequence.
Limits
Treat the server, tool descriptions, data boundary and permitted actions as separate review subjects. MCP identifies the integration standard; it is not a certification of a server or permission to perform every exposed action. Those controls belong to the application’s governed design.
Reader FAQ
Does using MCP make a tool safe?
No safety conclusion follows from protocol use alone. Review the particular server and operation before granting access.
Is MCP an AI model?
No. It defines how applications connect to capabilities and information; it does not supply model weights.
Model parameter
A model parameter is a learned value, such as a weight or bias, that contributes to the model’s behavior.
Hypothetical example
Training updates a weight.
Limits
Parameter count alone does not establish quality.
Reader FAQ
Are parameters prompts?
No. Prompts are input; parameters are model values.
Sources: GOOGLE-ML. Checked 2026-10-17.
Model training
Model training adjusts a model’s parameters using examples.
Hypothetical example
A classifier learns from labeled records.
Limits
Training differs from generating an answer with an already trained model.
Reader FAQ
Is inference training?
No. Inference applies the trained model.
Sources: GOOGLE-ML. Checked 2026-10-17.
Multi-agent system
A multi-agent system uses multiple agents with defined roles or tasks and a way to coordinate their work.
Hypothetical example
A coordinator delegates independent subtasks and combines the returned results.
Limits
More agents add coordination and failure modes; they do not guarantee better answers.
Reader FAQ
Can agents review one another’s work?
Yes, but useful independence and clear acceptance criteria still matter.
Sources: ANTH-PATTERNS. Checked 2026-10-17.
multimodal model
A multimodal model works with more than one type of data, such as text and images. The supported input and output combinations depend on the particular model; they are not implied by the label alone.
A practical example
An illustrative assistant could receive a chart image and a text question, then produce a written explanation.
Limits
Check the actual supported modalities and the quality of the task result. Being able to accept an image does not guarantee accurate interpretation of every chart.
Reader FAQ
Does multimodal mean every media format is supported?
No. A model may support selected combinations, and its input and output capabilities can differ.
Further reading: multimodal model.
negative keyword
A negative keyword excludes eligible matching searches from ad targeting. In Google Search ads, its negative match behavior differs from the behavior of a positive keyword.
A practical example
An illustrative retailer could exclude a phrase about free repairs when it only sells new products, after reviewing the actual query intent.
Limits
Exclusions can remove useful traffic too. Negative keywords do not automatically cover every close variant; match type and campaign context matter.
Reader FAQ
Will one negative word automatically exclude its singular and plural variants?
Not necessarily. Google’s negative keywords do not match close variants, so inspect the intended exclusions.
Further reading: negative keyword.
Neural processing unit
A neural processing unit, or NPU, is hardware specialized for computations used by neural networks. An application needs a compatible software route to use it. Microsoft’s NPU guide explains this requirement for Windows devices.
A practical example
An application may run a supported model on an NPU while performing other work on the CPU or GPU. The presence of an NPU does not establish that a particular language model, runtime or operator is supported.
Limits
Check the actual model format, execution provider, memory requirements and fallback behavior. Device marketing cannot substitute for that compatibility evidence. Microsoft’s Copilot+ eligibility rules are platform-specific; they are not the definition of an NPU.
Reader FAQ
Can every model use a computer’s NPU?
No. Software and hardware compatibility determine the usable path; some models require conversion or a different execution device.
Does an NPU guarantee that all data stays local?
No. Review the complete application’s network, logging and synchronization behavior before making a privacy claim.
nofollow
nofollow is a value in a link’s rel attribute that qualifies the relationship with its destination for search engines. Google uses it when the other qualification values do not fit and the publisher prefers not to associate the site with the destination or have Google crawl it from that link.
A practical example
A publisher could qualify a link to an untrusted reference with nofollow.
Limits
Paid links and user-generated links have more specific sponsored and ugc values. nofollow is not an instruction to remove the destination from the index.
Reader FAQ
Does nofollow mean the linked page cannot be indexed?
No. Google may discover it through other links or a sitemap. Excluding a resource from indexing requires an appropriate index instruction.
Further reading: nofollow — 1; nofollow — 2.
noindex
noindex is an instruction asking a supporting search engine to exclude a resource from its search index. It can be supplied through a robots meta tag or an X-Robots-Tag HTTP response header.
A practical example
A public utility page could remain accessible to visitors while requesting exclusion from Google Search.
Limits
Google must be allowed to crawl the resource to discover the instruction. noindex does not make the content private.
Reader FAQ
Is a robots.txt disallow equivalent to noindex?
No. Disallow restricts fetching; noindex addresses index inclusion. Blocking the fetch can keep the noindex instruction unseen.
Further reading: noindex — 1; noindex — 2.
Observability
Observability is the ability to investigate a system’s behavior using the signals it exposes.
Hypothetical example
A team follows a trace to locate a failed workflow step.
Limits
Useful diagnosis depends on adequate instrumentation and context.
Reader FAQ
Is observability just collecting logs?
No. Metrics and traces can also be needed to explain behavior.
Sources: OTEL. Checked 2026-10-17.
offline conversion import
Offline conversion import sends outcomes measured outside the immediate online conversion flow into the advertising measurement system. It connects later business events, such as a completed phone sale, with eligible earlier ad interactions.
A practical example
An illustrative business could record a qualified sale after a web inquiry and import the final outcome using an approved measurement route.
Limits
Match quality, timestamps, duplicates and permitted data matter. Supported import methods and API eligibility can change, so verify the current route before implementation.
Reader FAQ
Does every imported offline sale become an attributed ad conversion?
No. The outcome must satisfy the route’s matching and eligibility rules; a business record alone does not establish successful attribution.
Further reading: offline conversion import — 1; offline conversion import — 2.
On-page SEO
On-page SEO improves a page’s content and descriptive elements so readers and search engines can understand its purpose.
Hypothetical example
An editor clarifies a title, headings and relevant internal links.
Limits
Readable, useful and original content matters; repeating keywords is not a guarantee of ranking.
Reader FAQ
Is on-page SEO only writing a meta description?
No. It also includes the page’s content, organization and descriptive elements.
Sources: GOOGLE-SEO. Checked 2026-10-17.
Orchestration
Orchestration coordinates multiple tasks, services or workflows, including their dependencies and execution order.
Hypothetical example
A coordinator waits for data collection before starting analysis and review.
Limits
Coordinating tasks does not establish that their outputs are correct.
Reader FAQ
How does orchestration differ from automation?
Automation performs tasks; orchestration coordinates how tasks work together.
Sources: IBM-ORCH. Checked 2026-10-17.
Organic search result
An organic search result is an unpaid listing selected for its relevance to a search rather than purchased as an advertisement.
Hypothetical example
A guide appears as a regular search listing.
Limits
Unpaid placement does not mean producing or optimizing content has no cost.
Reader FAQ
Can buying search ads buy an organic position?
No. Paid and organic placements are separate.
Sources: GOOGLE-ORGANIC. Checked 2026-10-17.
Parallelization
Parallelization executes independent tasks or multiple attempts concurrently, sometimes combining their outputs for comparison.
Hypothetical example
Two independent checks examine sources and terminology at the same time.
Limits
Dependent steps still need ordering; shared resources can limit gains.
Reader FAQ
Does parallel work always finish faster?
No. Coordination overhead and dependencies can offset gains.
Sources: ANTH-PATTERNS. Checked 2026-10-17.
pay-per-click advertising
Pay-per-click advertising, or PPC, is an advertising payment model in which an advertiser pays for recorded ad clicks. It describes how advertising is charged; cost per click, or CPC, is the amount or average amount paid for those clicks.
A practical example
An illustrative advertiser could receive 20 chargeable clicks at an average CPC of 1.50 currency units, giving a click cost of 30 units.
Limits
A paid click is not a sale or a qualified lead. Check the actual billing model, budget and measured outcomes; not every advertising format is charged per click.
Reader FAQ
Is PPC the same thing as CPC?
PPC names the payment model; CPC describes the click cost. They are closely related terms but serve different explanatory purposes.
Further reading: pay-per-click advertising — 1; pay-per-click advertising — 2; pay-per-click advertising — 3.
Performance Max
Performance Max is a goal-based Google Ads campaign type that can access inventory across several Google channels from one campaign. It uses the supplied goals and automation to seek conversions or conversion value.
A practical example
A business could use one campaign to reach eligible inventory across Search, YouTube and Display, with reviewed assets and measurement.
Limits
Cross-channel access does not establish that every channel or placement will receive delivery. Inspect actual reporting, assets and goal quality.
Reader FAQ
Does an audience signal restrict Performance Max to that audience?
No. Google documents that delivery can extend beyond the audiences supplied as signals.
Further reading: Performance Max — 1; Performance Max — 2.
Pipeline
A pipeline is a sequence of processing stages in which work or data moves toward a result, with dependencies between stages. The DIOS case shows this sequence in video production, from scripts and media to publishing and feedback.
Hypothetical example
Collect data, validate it, transform it and produce a report.
Limits
A failed stage may require a stop or recovery; later stages should not silently consume invalid output.
Reader FAQ
Is every pipeline an autonomous agent?
No. A pipeline can follow predefined steps.
Sources: GH-ACTIONS. Checked 2026-10-17.
pretraining
Pretraining is an initial stage of model training on a dataset, often broad in scope, before further adaptation. In language-model discussions, it commonly establishes a base model that may later undergo additional training.
A practical example
A base language model could first learn from a large text collection and later be adapted to a specific classification task.
Limits
Pretraining does not itself establish suitability for the final application. Describe the dataset and later adaptation separately.
Reader FAQ
Is a pretrained model already fine-tuned for my task?
Not necessarily. Pretraining and task-specific fine-tuning describe different stages; inspect the actual model history.
Further reading: pretraining.
Prompt
A prompt is input that asks or guides a model to produce a response, potentially including instructions, context and examples.
Hypothetical example
A prompt asks for a source-supported explanation for beginners.
Limits
A clear prompt cannot guarantee factual accuracy.
Reader FAQ
Is a prompt only a question?
No. It can contain instructions, examples and contextual information.
Sources: GC-PROMPT. Checked 2026-10-17.
prompt engineering
Prompt engineering is designing and refining the instructions, context and examples supplied to a model to guide its response.
A practical example
An editor states the audience, source constraints and required format, then checks the result on representative briefs.
Limits
Changing a prompt is distinct from updating model weights. A clear instruction still needs output evaluation.
Reader FAQ
Can a prompt guarantee factual accuracy?
Treat it as guidance to evaluate, not a guarantee. Keep source verification in the workflow.
Further reading: prompt engineering.
prompt injection
Prompt injection is an attempt to redirect an AI system through adversarial instructions in content it processes. In indirect attacks, those instructions can appear in a webpage, document or tool result rather than the user’s legitimate request.
A practical example
An illustrative retrieved page could contain text asking an assistant to ignore its task and send data to an unrelated destination.
Limits
Treat external content as evidence to inspect, not as permission to act. The presence of safeguards does not establish that every attack will be blocked.
Reader FAQ
Is an instruction in a retrieved document an authorized user command?
No authority follows from being retrieved. The application must distinguish external content from the user’s permitted task.
Further reading: prompt injection.
Provenance
Provenance records where information or an artifact came from and the people, activities and transformations involved in producing it.
Hypothetical example
A claim record links an original source to a dated editorial revision.
Limits
A documented origin helps assessment but is not proof that the claim is true.
Reader FAQ
Does provenance mean only a source URL?
No. It can also record derivation, responsibility and processing history.
Sources: W3C-PROV. Checked 2026-10-17.
Quality Score
Quality Score is Google Ads’ reported estimate of ad, keyword and landing-page quality, shown on a 1–10 scale.
Hypothetical example
An advertiser examines expected CTR, ad relevance and landing-page experience.
Limits
The reported score is not used at auction time to determine Ad Rank.
Reader FAQ
Is Quality Score the same as Ad Rank?
No. It is a reported quality estimate, not the auction ranking itself.
Sources: GOOGLE-QS. Checked 2026-10-17.
quantization
Model quantization uses lower-precision representations, such as for model weights, to reduce memory requirements while aiming to retain useful model quality.
A practical example
A deployment compares a lower-precision model with the original on its own tasks and hardware.
Limits
Memory savings do not by themselves establish better latency or unchanged output quality. Method and hardware support matter.
Reader FAQ
Is quantization lossless?
Do not assume it is. Evaluate quality and execution behavior for the chosen method and workload.
Further reading: quantization.
ReAct
ReAct is a language-model pattern that interleaves reasoning, actions and observations. An action obtains information or changes an allowed environment; its result informs the next step. The original ReAct research studied this combination in question answering and interactive tasks.
A practical example
To check a software claim, a system could choose a documentation search, read the returned page and discover an edition mismatch. It then revises its search instead of continuing with the original assumption.
Limits
A loop can repeat an unhelpful action, accept misleading tool output or stop too early. Specify allowed tools, evidence checks and a completion condition. The pattern itself cannot guarantee a correct answer.
Reader FAQ
How does an observation change the next step?
It supplies new evidence, including errors or missing information, that can change which action is appropriate.
Must a production system publish its internal reasoning?
No. An operational audit can retain observable tool calls, sources and outcomes without exposing private internal reasoning.
redirect
A redirect sends a request for one URL to another destination. HTTP redirects use response status codes and a destination; permanent and temporary redirects communicate different intentions.
A practical example
When an article genuinely moves, its old URL could send visitors to the corresponding new page.
Limits
The destination must serve the intended content. Loops, broken destinations and unnecessary chains can interrupt access; selecting a code does not validate the destination.
Reader FAQ
Which HTTP codes signal a permanent move?
301 and 308 are permanent redirects. The appropriate choice also depends on request-method semantics; Google treats both as strong signals for the target.
Further reading: redirect — 1; redirect — 2.
remarketing
Remarketing means re-engaging people who have already interacted with a business. Google Ads now uses the term your data for many of the segments formerly labeled remarketing.
A practical example
An illustrative campaign could address prior website visitors with an appropriate follow-up offer using a permitted data segment.
Limits
Prior engagement does not guarantee a purchase or authorize every use of personal data. Apply the relevant policies and data permissions before implementation.
Reader FAQ
Is remarketing the same as finding entirely new prospects?
No. Its defining context is previous engagement; reaching new people has a different audience basis.
Further reading: remarketing.
reranking
Reranking reorders an existing set of search candidates using an additional relevance assessment. It comes after initial retrieval and can focus the final list on which candidates better answer the query.
A practical example
A retrieval pipeline could first collect twenty passages and then rerank those candidates before selecting a smaller answer context.
Limits
A reranker can only assess the supplied candidate set. Check retrieval coverage separately from the order of returned results.
Reader FAQ
Can reranking repair a missing source?
It cannot select a document absent from its candidate set. Improve or broaden retrieval when the required source was never retrieved.
Further reading: reranking.
responsive search ad
A responsive search ad is a Google Ads search format that combines supplied headlines and descriptions into different ad arrangements.
A practical example
An advertiser supplies compatible messages about delivery and product choice; the platform assembles combinations.
Limits
Review messages both individually and in combination. Do not assume every asset appears on every impression or that the format guarantees better results.
Reader FAQ
How should required wording be handled in responsive search ads?
Use the format’s supported placement controls when wording must appear, and verify the actual requirements and preview.
Further reading: responsive search ad.
Retrieval-augmented generation
Retrieval-augmented generation, or RAG, combines retrieved external information with language generation. An application finds relevant material and makes it available when producing an answer. The original RAG paper combined a language model with retrieved passages from a document index.
A practical example
A policy assistant could retrieve a current, approved policy section before answering a question. The retrieved passage provides evidence the answer can reference.
Limits
Wrong documents, stale versions or incomplete passages can make the final answer misleading. Check whether retrieval found the right evidence separately from whether the generated answer used it accurately. A citation is useful only if its source supports the claim.
Reader FAQ
Does RAG update model weights?
Retrieving material for a request does not itself change the model’s weights. A RAG system may nevertheless include separately trained components: the original research included fine-tuning. RAG and training are not mutually exclusive.
Is adding more context always RAG?
No. The term includes an information-retrieval stage; merely enlarging a prompt does not establish that stage.
Retry with backoff
Retry with backoff repeats an operation after a delay, often increasing the delay between attempts, to handle temporary failures.
Hypothetical example
A data request waits before retrying after a transient service error.
Limits
Retries need limits and safe side effects; a permanent rejection is not a reason to retry indefinitely.
Reader FAQ
Why delay a retry?
Repeated immediate requests can increase contention or overload a service.
Sources: AWS-RETRY. Checked 2026-10-17.
return on ad spend
Return on ad spend compares attributed conversion value with advertising cost. Expressed as a percentage, it is conversion value divided by ad spend, multiplied by 100.
A practical example
An illustrative value of 500 against spend of 100 gives a ROAS of 500%, or 5 times spend.
Limits
Keep period, currency, attribution and value definitions compatible. Revenue-based ROAS does not subtract every business cost. For zero ad spend, the ratio is undefined; do not report it as an ordinary finite ROAS.
Reader FAQ
Is target ROAS the achieved ROAS?
No. A bidding target is an objective. Observed ROAS is a measured result and can differ from that objective.
Further reading: return on ad spend.
rich result
A rich result is a Google Search result with supported features beyond an ordinary text result, such as additional visual or informational details. Eligible structured data can help Google understand content for some of these features.
A practical example
An eligible recipe page could provide recipe information that Google may display in a richer result.
Limits
The page must meet the relevant feature requirements and policies. Correct markup and a successful validation result do not guarantee that Google displays the feature.
Reader FAQ
Does passing the Rich Results Test guarantee a rich result?
No. It checks relevant markup conditions, while actual appearance also depends on Google’s selection and the applicable guidelines.
Further reading: rich result — 1; rich result — 2.
robots.txt
robots.txt is a file that tells compliant crawlers which paths they may request on a site. Its rules control crawling for specified user agents; they are not an access control system.
A practical example
A site could disallow a crawler from requesting a low-value utility path while allowing public articles.
Limits
A disallowed URL can still appear in search results if discovered elsewhere. Private content needs actual access protection.
Reader FAQ
Can robots.txt reliably remove a page from Google’s index?
No. Blocking crawling can prevent Google from seeing a noindex instruction. Crawl restrictions and index exclusion serve different purposes.
Further reading: robots.txt — 1; robots.txt — 2.
Routing
Routing classifies an input and directs it to an appropriate task, tool or specialist path.
Hypothetical example
A request about campaign figures goes to an analysis path.
Limits
Misclassification can send work to an unsuitable path; fallback handling matters.
Reader FAQ
Does routing require an LLM?
No. Rules can also select a route.
Sources: ANTH-PATTERNS. Checked 2026-10-17.
Sandbox
A sandbox is an isolated execution environment that limits how a program or agent interacts with surrounding resources.
Hypothetical example
An agent processes test files within a restricted workspace.
Limits
Isolation strength depends on its design and configured filesystem, network and permission boundaries.
Reader FAQ
Does a sandbox guarantee complete security?
No. Its protections must be assessed for the actual environment.
Sources: DOCKER-SANDBOX. Checked 2026-10-17.
Schema.org
Schema.org is a shared vocabulary of types and properties for describing things in structured data. A publisher can use its vocabulary with formats such as JSON-LD, Microdata or RDFa to express what a page discusses.
A practical example
A page about an organization could describe that organization and its name using the appropriate type and properties.
Limits
Markup should describe the actual content. A valid Schema.org type is not automatically supported as a Google rich-result feature.
Reader FAQ
Is Schema.org the same thing as JSON-LD?
No. Schema.org supplies vocabulary; JSON-LD is one format that can carry that vocabulary.
Further reading: Schema.org — 1; Schema.org — 2.
search engine optimization
Search engine optimization, or SEO, is work that helps search engines understand a site’s content and helps people find useful pages through search. It can involve content clarity, technical accessibility and how pages are presented, rather than only inserting keywords.
A practical example
A publisher could clarify a guide’s structure and make its important pages accessible to crawlers while keeping the explanation useful for readers.
Limits
SEO does not guarantee indexing, a first-place ranking or a fixed time to results. Improvements need to fit the site and be evaluated against observed search performance.
Reader FAQ
Does repeating more keywords guarantee a higher ranking?
No. Useful content and appropriate technical practices matter; Google does not promise a ranking position from keyword repetition or any single change.
Further reading: search engine optimization.
Search engine results page (SERP)
A SERP is the page of results a search engine presents in response to a query.
Hypothetical example
A query returns organic listings alongside labeled advertising placements.
Limits
Results and layouts vary by query, location, time and search product.
Reader FAQ
Is everything on a SERP an organic result?
No. A results page can also contain paid placements.
Sources: MS-SERP. Checked 2026-10-17.
search impression
A search impression records a link being seen or potentially seen under the rules of the search feature and report. In Search Console, counting can differ by result type and aggregation.
A practical example
A result in the current set of search results can count without a click; some features have additional visibility rules.
Limits
This sense is distinct from an advertising impression or a site visit. Compare reports with compatible grouping.
Reader FAQ
Must the user scroll to the link for it to count?
Not always. The counting rule depends on the feature; use the documented rule for the specific report.
Further reading: search impression.
search intent
Search intent is the purpose behind a person’s search, inferred from their query and context. They may want an explanation, a particular destination, a comparison or help taking an action.
A practical example
The query ‘what is an NPU’ likely seeks an explanation; ‘compare NPU laptops’ likely calls for comparison evidence.
Limits
Intent categories are editorial interpretations, not certain knowledge about an individual. Similar wording can have different purposes depending on context.
Reader FAQ
Can intent always be identified from one keyword?
No. Wording provides clues, but context and the information needed to satisfy the request matter too.
Further reading: search intent.
search query
A search query is the word or phrase a person submits to a search engine to seek information or a result. It is the actual request, which may differ from the terms a publisher chose when planning content.
A practical example
A reader could search for ‘what counts as a conversion’ rather than the publisher’s preferred label ‘conversion action’.
Limits
A query does not reveal every detail of the searcher’s intent. Search Console reports available query data for Google Search performance, not an exhaustive record of every query.
Reader FAQ
Is a search query the same as a planned keyword?
No. The query is submitted by the searcher; a planned keyword is a term the publisher or advertiser chooses to work with.
Further reading: search query — 1; search query — 2; search query — 3.
Search ranking
Search ranking is the ordering of results according to a search engine’s systems and signals for a particular query.
Hypothetical example
Different relevant pages appear in a different order for different searches.
Limits
A position is contextual and can change; no single optimization guarantees first place.
Reader FAQ
Do search engines use only exact keywords?
No. Ranking systems consider several signals, including relevance and usability.
Sources: GOOGLE-RANK. Checked 2026-10-17.
Semantic search
Semantic search seeks results by meaning or conceptual similarity rather than relying only on matching literal words.
Hypothetical example
A search about reducing delays finds a passage about latency.
Limits
Vector similarity can support semantic retrieval, but similar meaning does not establish factual correctness.
Reader FAQ
Is semantic search identical to vector search?
They describe different aspects: semantic search is the objective; vector search is a retrieval approach.
Sources: AZ-SEARCH. Checked 2026-10-17.
Single source of truth (SSOT)
A single source of truth is an agreed, reconciled reference that stakeholders and applications use for a defined body of information.
Hypothetical example
A team references one maintained dataset instead of separate conflicting copies.
Limits
Its scope, ownership and update rules need to be defined; designation alone does not remove errors.
Reader FAQ
Does SSOT mean every kind of information belongs in one file?
No. The aim is a consistent authoritative reference, not necessarily one physical file.
Sources: IBM-SSOT. Checked 2026-10-17.
Smart Bidding
Smart Bidding is Google’s family of bidding strategies that uses AI to optimize for conversions or conversion value at auction time. It includes strategies such as target CPA and target ROAS.
A practical example
A campaign could use a conversion-focused strategy to set auction bids from the selected goals and available signals.
Limits
Its optimization objective depends on the conversions and values supplied. Review measurement and actual results; the strategy name does not guarantee profitability.
Reader FAQ
Is every automated bidding strategy Smart Bidding?
No. Smart Bidding specifically focuses on conversions or conversion value; automated bidding is a broader category.
Further reading: Smart Bidding.
structured data
Structured data is a standardized way to describe a page’s information and classify its content. Google uses supported markup to understand content and assess eligibility for particular search features.
A practical example
A recipe record describes ingredients and preparation steps in a supported vocabulary.
Limits
Use accurate content and the feature’s required properties. Valid markup alone does not guarantee a visible rich result.
Reader FAQ
Is every Schema.org type a Google rich-result feature?
No. Google Search documentation defines its supported features and eligibility; the vocabulary covers broader uses.
Further reading: structured data.
System instruction
A system instruction provides application-level guidance about a model’s role, style or constraints alongside the task input.
Hypothetical example
An application instructs an assistant to distinguish facts from assumptions.
Limits
Instruction handling is model- and platform-specific; a prompt is not a substitute for access controls.
Reader FAQ
Can system instructions define response style?
Yes. They can guide style and constraints for the application.
Sources: GC-PROMPT. Checked 2026-10-17.
target CPA
Target CPA is a Google Ads bidding strategy with a desired average cost per conversion. It sets bids to seek conversions while trying to achieve that average target.
A practical example
A team could set a target for qualified lead conversions, then compare the achieved average with the effective target over a suitable period.
Limits
Individual conversions can cost more or less than the target. A very low target can restrict opportunities; the goal and actual measurement remain essential.
Reader FAQ
Does a target CPA cap every conversion’s cost?
No. It is an average objective, not a fixed price for every conversion.
Further reading: target CPA.
target ROAS
Target ROAS is a Google Ads bidding strategy that seeks conversion value while trying to achieve a target average return on ad spend. The target expresses conversion value relative to ad cost, often as a percentage.
A practical example
An illustrative 400% target represents a goal of four value units for each cost unit, using the defined conversion-value basis.
Limits
Revenue, profit and assigned lead value are different bases. A target is not an achieved result, and an overly restrictive target may reduce traffic.
Reader FAQ
Does target ROAS always mean profit?
No. It depends on what the recorded conversion value represents; revenue-based ROAS does not subtract all business costs.
Further reading: target ROAS.
Technical SEO
Technical SEO addresses the technical conditions that help search engines access, process and understand content.
Hypothetical example
A review checks crawl access, indexability and canonical signals.
Limits
Meeting technical requirements does not guarantee crawling, indexing or ranking.
Reader FAQ
Does technical SEO replace useful content?
No. Technical eligibility and content quality address different needs.
Sources: GOOGLE-ESSENTIALS. Checked 2026-10-17.
Telemetry
Telemetry is data emitted about a system and its behavior, including metrics, logs and traces.
Hypothetical example
A workflow records its duration and errors.
Limits
Collection must be scoped to avoid unnecessary sensitive data; telemetry alone does not explain every cause.
Reader FAQ
How does telemetry relate to observability?
Telemetry supplies evidence used to understand the system.
Sources: OTEL. Checked 2026-10-17.
Throughput
Throughput is the amount of work or data processed per unit of time.
Hypothetical example
A pipeline reports the number of records processed each minute.
Limits
Specify the work unit and conditions; higher throughput need not mean lower latency.
Reader FAQ
How is throughput different from latency?
Throughput measures a processing rate; latency measures delay for a request or item.
Sources: IBM-PIPELINE. Checked 2026-10-17.
title link
A title link is the linked title displayed for a search result. Google generates it using several sources, including page titles, headings and other references.
A practical example
A search result may show a title derived from the page’s main heading rather than copying its title element exactly.
Limits
The HTML title is an input to the displayed title, not a promise of exact reproduction.
Reader FAQ
Can I set the displayed title directly?
You can provide clear, consistent title information; Google’s title-link generation remains automated.
Further reading: title link.
token
In a language model, a token is a unit of text the model processes, such as a word, part of a word or a character. Tokenization is the conversion from text to those units.
A practical example
A tokenizer might divide an unfamiliar word into several parts, so a document’s word count alone would not identify its token count.
Limits
Use the tokenizer for the actual model when checking a limit. Do not assume one word always equals one token.
Reader FAQ
Why can two texts with the same word count use different token counts?
Their characters, language and division into token units can differ. Word count is not the model’s unit of measurement.
Further reading: token.
tool calling
Tool calling lets a model request functionality exposed by an application. The application controls whether the requested tool is executed and can return its output to the model.
A practical example
A system requests a permitted document lookup and uses the returned evidence in a response.
Limits
A model request and successful execution are separate events. Available tools and application permissions determine the action boundary.
Reader FAQ
Does a tool call give a model unrestricted access?
No. It can request exposed functionality; the application must control execution and permissions.
Further reading: tool calling.
TOPS
TOPS means trillions of operations per second. In AI hardware specifications it commonly describes potential peak arithmetic throughput under stated assumptions. Qualcomm’s technical explainer relates the figure to arithmetic units, operating frequency and precision.
A practical example
Two devices with similar advertised TOPS may still differ in the time an application takes to answer. Memory, compatible software and workload execution affect that result.
Counting matters
Qualcomm’s example counts a multiply and an addition as two operations. Its formula therefore doubles the multiply-accumulate unit count before multiplying by frequency. That is a documented example, not a universal convention for every accelerator. Compare ratings only after checking precision, counting rules, dense or sparse assumptions, and whether they describe a whole platform or a particular processor.
Reader FAQ
Is TOPS the same as tokens per second?
No. TOPS counts arithmetic operations; tokens per second describes generated output throughput.
Does a higher number guarantee a faster application?
No. Measure the actual workload under comparable conditions.
Transformer
A Transformer is a neural-network architecture that uses attention to work with relationships within sequences. The original research introduced an attention-based encoder-decoder for sequence tasks without recurrent or convolutional layers.
A practical example
A translation model could use attention to relate a word to other relevant parts of the sentence.
Limits
Transformer identifies an architecture, not an accuracy or reasoning guarantee. Later systems may use different variants; do not assume every model has the original encoder-decoder structure.
Reader FAQ
Does Transformer mean chatbot?
No. The original work studied machine translation; a chatbot is an application, and its architecture requires separate identification.
Further reading: Transformer.
Trigger
A trigger is the event or condition that starts an automated workflow, such as a schedule or a specific incoming event.
Hypothetical example
A new approved record starts a reporting workflow.
Limits
A trigger starts processing; it does not independently authorize every later action.
Reader FAQ
Can a schedule be a trigger?
Yes. Workflows can start on a schedule as well as from events.
Sources: GH-ACTIONS. Checked 2026-10-17.
vector search
Vector search retrieves items by comparing numerical representations, often embeddings, with a query representation. In semantic search, the embedding space is used to find items with similar meaning.
A practical example
A document service could retrieve passages about account closure even when they use wording different from the question.
Limits
Similarity is not factual support. Inspect retrieved passages, and evaluate the search configuration on relevant and misleading candidates.
Reader FAQ
Does vector search require exact keyword matches?
Not necessarily. Semantic similarity search compares representations; a hybrid system can also include lexical matching.
Further reading: vector search.
Webhook
A webhook delivers an event notification to a configured external receiver when a subscribed event occurs.
Hypothetical example
A service notifies a workflow that a record changed.
Limits
Delivery behavior and authentication must be checked for the particular provider.
Reader FAQ
How does a webhook differ from polling?
The sender notifies the receiver instead of the receiver repeatedly checking for changes.
Sources: GH-WEBHOOK. Checked 2026-10-17.
Workflow
A workflow is an organized set of steps and decisions that takes work from an initial input to an outcome.
Hypothetical example
A report workflow collects figures, checks them, drafts commentary and requests review.
Limits
A workflow may include people, software or both; it need not use AI.
Reader FAQ
Does a workflow have to be fully automated?
No. Human review can be one of its steps.
Sources: AWS-WF. Checked 2026-10-17.
Workflow step
A workflow step is one defined stage of a process, which may perform work or control what happens next.
Hypothetical example
A validation step checks a file before analysis.
Limits
Step types and input/output rules depend on the workflow system.
Reader FAQ
Is every step an external tool call?
No. A step can also make a decision or control execution.
Sources: AWS-STATE. Checked 2026-10-17.
Workspace
In VS Code, a workspace is the collection of one or more folders opened in a window, with configuration scoped to that workspace.
Hypothetical example
An editor opens a project’s folders and local settings as its workspace.
Limits
Products define workspace boundaries differently; being visible does not grant permission to publish or change every resource.
Reader FAQ
Must a workspace contain only one folder?
No. Some tools support multiple folders within one workspace.
Sources: MS-WORKSPACE. Checked 2026-10-17.
XML sitemap
An XML sitemap is a machine-readable file that lists site URLs and can provide information about those resources. Search engines can use it to discover pages and understand which URLs the site presents for crawling.
A practical example
A publisher could list its approved article and glossary URLs in a sitemap.
Limits
A sitemap is a discovery aid, not a promise that listed URLs will be crawled, indexed or ranked. Its URLs and update information need to reflect the site accurately.
Reader FAQ
Does adding a URL to a sitemap guarantee indexing?
No. Google decides whether to crawl and index a resource separately from receiving a sitemap.
Further reading: XML sitemap.
