AI Search Glossary

Occurrences of a brand name in the text an AI assistant generates, and why a mention is a different measurement from a citation, a rank, or a visit.

How an AI system orders candidate sources before writing an answer, why it is a selection problem rather than a results page, and where each stage of the ordering fails.

Editing a page so that the specific sentences an assistant lifts into an answer are correct, self-contained and attributable — the passage-level half of AI search work.

The practice of measuring how content fares inside AI assistants' answers — what can actually be observed, what has to be sampled, and what cannot be attributed at all.

A composite number some tools compute to summarise how often a brand appears in AI answers — what it is made of, why the components are not comparable, and why Attensira does not compute one.

Planning a content estate on the assumption that its main reader is a machine that extracts one fact and leaves — what that changes about structure, access policy and maintenance.

Making a page eligible to be fetched, retrieved and cited by ChatGPT's search features, and the reasons a page that ranks well on Google is still never quoted.

The credibility a specific piece of content and its named creator carry on a specific topic — how raters are told to assess it, and why it is a page-level property rather than a site-wide score.

A group of pages covering one subject at different depths, linked to a central page, so that a system can see coverage of a topic rather than one page about it.

The practice of getting a source reproduced inside an AI-generated answer, introduced as a named paradigm in a 2023 research paper.

The practice of making a page usable by a system that reads sources and writes an answer from them, rather than by a system that returns a ranked list of links.

Writing and publishing so that a language model can fetch a page, retrieve a passage from it, and repeat the passage correctly — the three separate stages, and the failure that happens at each.

The practice of treating every surface where people ask questions — engines, assistants, marketplaces, video and community sites — as a retrieval target, rather than optimising one engine.

How an AI answer attributes what it says to the pages it read, and why a citation is a distinct outcome from a click or a mention.

Being known as the go-to source for one subject rather than a thin source for many: what Google's rater guidelines call Authoritativeness, and what it means for retrieval.

Publishing with the intent of shaping what a future model learns about you — why it is the slowest lever available, and which crawler decisions actually control it.

A constrained HTML subset built to make mobile pages load fast, and a case study in what happens when a format's distribution advantage is withdrawn.

Inbound links from other sites: what search engines actually do with them, what the rel attributes change, and why a link's value to an AI assistant is mostly about being reachable and quotable.

The share of sessions that failed an engagement test — a definition that changed materially between Universal Analytics and GA4, and that measures less every year as answers move off your page.

The one URL you nominate as the indexable version of a page, how the rel=canonical hint is actually weighted, and what a split canonical costs you when an agent needs a link to cite.

Clicks divided by impressions, as Google defines it in Search Console — and why the ratio stops describing your reach once answers are generated instead of listed.

Google's three published field metrics for loading, responsiveness and visual stability, with the exact thresholds and the percentile they are judged at.

The two separate stages that decide whether a page can be retrieved at all, and the reason a serving rule on a blocked page is never read.

A third-party predictive score for a whole domain, not a signal any search engine publishes — and what actually carries across a site when a model decides whom to cite.

The interval between a searcher clicking a result and returning to the results page. It is an SEO-industry term, not a metric Google publishes, documents, or exposes in any tool.

The superseded three-letter form of what Google's rater guidelines now call E-E-A-T, after Experience was added in December 2022, and which Google states is not a ranking factor.

Publishing so that a machine can resolve which real-world thing your page is about, and reconcile it with the same thing described elsewhere.

Google's reversed search result, where a passage lifted from a page is shown above the link — selected by Google's systems, requestable by nobody, and refusable with one meta rule.

Google's free property-verified reporting tool for how a site appears in Google Search — and the reason most teams have no equivalent record of how they appear in AI answers.

Making images cheap to load and legible as text — the two independent jobs, why only the second one survives into an AI answer, and where each fails.

Links from one page of a site to another: how crawlers discover pages, what makes a link machine-readable at all, and why the anchor text is the part that carries meaning.

A store of entities and the typed relationships between them — Google's since 2012, and the reason a machine can tell your company from a word that happens to spell it.

The automatically generated information box Google shows for an entity in its Knowledge Graph, and why it is a read-out of the web's consensus about you rather than a page you write.

Making a business eligible and prominent in place-qualified results — how Google says local ranking is decided, what the structured data requires, and what changes when an assistant answers instead.

Longer, more specific queries that each occur rarely but together make up most of what people ask — why the tail exists mechanically, and why it is the part of demand an assistant answers directly.

Google indexes and ranks the mobile version of a page, crawled with the smartphone agent — which means the mobile rendering, not the desktop one, is the document everything downstream reads.

Google's collective term for the signals describing how it feels to use a page — explicitly not a single ranking signal, and the reason a fast page still loses to a more relevant slow one.

How fast a page loads and responds, the thresholds Google publishes for it, and why latency decides whether an AI agent ever reads you at all.

Web apps built on a manifest and a service worker so they install and work offline — and the one architectural habit that makes them invisible to the agents writing AI answers.

The standardised file that tells crawlers what they may fetch, now the main place where AI training and AI retrieval access are decided separately.

The Schema.org vocabulary and the JSON-LD you embed with it, what it is genuinely good for, and the limits of what it can make happen.

What a person is actually trying to accomplish with a query, in Google's own four categories, and why an answer engine resolves it before it ever reaches your page.

The practice of making a page understandable and retrievable by search systems — defined by Google as helping engines understand your content and helping users decide whether to visit.

The page a search engine returns for a query — organic links, ads, and a growing set of documented result types that answer without a click.

The non-blue-link elements on a search results page — what Google documents you can and cannot influence, and why each one is a chance to be read without being visited.

Everything about how a person perceives and operates a site — speed, stability, responsiveness, clarity — and the part of it that also decides whether a machine can read the page at all.

Making a video findable by describing it in text a machine can read: structured data, a sitemap entry, a stable file URL and a watch page. The pixels themselves are not what gets indexed.

Search initiated by speech and usually answered by speech, which collapses a page of results to one spoken answer and makes the retrieved passage — not the ranking — the thing worth optimising.

Writing for a channel that returns one spoken answer instead of a list — and the reason its main technique, speakable structured data, has always been a narrow beta.

A machine-readable list of the URLs you want discovered, bounded by the protocol at 50,000 URLs and 50MB per file, and a hint rather than an instruction.

Google's rater-guideline classification for topics that could significantly impact health, financial stability, safety, or the welfare of society, raising the accuracy bar.

Techniques that try to decide whether a text or image was machine-generated, why post-hoc classifiers are unreliable, and why provenance signing answers a different question.

When a language model states something false with the same fluency it states something true — why the training pipeline produces it, and what it means for a brand being described by a model.

Search where a model composes the answer and fetches pages through named crawlers, rather than returning a ranked list of links for you to read.

The corpus a model learns from, how it differs from what an assistant retrieves at query time, and how publishers have actually responded.

The AI safety company that builds the Claude models, and the three separately-named web crawlers a publisher has to answer if they want Claude to read their pages.

The bidirectional transformer Google uses to read a query as a whole sentence rather than a bag of keywords, and why that ended keyword-shaped writing.

OpenAI's conversational assistant, and — because it fetches pages while answering — one of the surfaces where a page you publish today can be quoted today.

Splitting a corpus into retrievable pieces before embedding it, why the split destroys context, and what contextual retrieval measured about fixing it.

Anthropic's family of large language models and the assistant built on them, and the three separately-documented agents that decide whether your page can reach it.

The token budget a model can reference in one request, what counts against it, and why a bigger window does not remove the need to retrieve selectively.

Vectors of floating point numbers whose distance measures relatedness, the representation underneath semantic search and retrieval.

Google's family of multimodal models, and the grounding layer that lets one of them answer from live web sources and return citations with the text.

Connecting a model's answer to retrieved real-world sources so it can cite them, and why a grounded answer is still not a verified one.

The date beyond which a model has no trained knowledge, why vendors publish a reliable cutoff, and what it does and does not stop a model saying.

A neural network with many parameters, trained on large amounts of text to predict the next token — and the reason an answer about your product can be fluent, confident and wrong.

Software that derives a rule from data instead of being given one, and the reason an AI answer about your brand can be confidently wrong in a way no bug fix reaches.

The open standard that lets an AI application connect to external data and tools over JSON-RPC, and what its three roles and three server features actually do.

The field concerned with getting machines to process human language, and the lineage — transformers, then bidirectional pre-training — behind the systems that pick which passage gets quoted.

The company behind ChatGPT and the GPT model family, and — for a publisher — the operator of four separately named crawlers whose access decides whether your page can be quoted.

An answer engine that retrieves live web documents and writes cited answers from them, via two documented agents — one bound by robots.txt and one that, per its own docs, generally is not.

Structuring the text given to a language model so its output meets a stated success criterion — and, for anyone measuring AI answers, the variable that decides what the measurement is of.

The vulnerability where input alters a model's behaviour in unintended ways, why the indirect kind is a publishing problem, and what OWASP says can be done.

The architecture that retrieves documents at query time and has a model write from them, and the reason your page can be quoted without ever being trained on.

Google's 2015 deep learning ranking system for relating words to concepts — the first of its kind in Search, and the point at which exact keyword matching stopped being the job.

Fetching current documents at answer time instead of relying on a model's fixed training data, and the reason a page you publish today can be quoted today.

The second-pass scoring step that reorders retrieved passages before a model reads them, what it costs, and why it changes which page gets quoted.

The units a language model actually reads and writes, why they are not words, and why every limit and price you meet is denominated in them.

Retrieval by geometric proximity between embedding vectors rather than by word overlap, and the reason a passage about your product can be found without containing the words that were searched.