Blog · Foundations

An AI search glossary that defines the terms properly

Published 2 October 2026 · Updated 12 October 2026 · 9 min read

An AI search glossary that defines the terms properly
Photograph: Annie Spratt / unsplash

The short answer: A glossary earns its place only when every entry names something you can go and check. Below are the terms that matter, grouped by where the work breaks: retrieval and grounding, the crawler agents that decide whether you get fetched, the entity vocabulary that decides whether an engine knows who you are, and the measurement words that get misused most often.

Vocabulary here moves faster than practice. Plenty of terms in circulation were coined to describe one product feature, and a few describe nothing. Each cluster below maps to a failure you can reproduce.

Terms for the machinery that writes the answer

Grounding, also called retrieval-augmented generation, is the step where a model fetches live documents and writes from them rather than from memory. Google defines it plainly in its optimisation guide, updated 10 July 2026, as a technique that uses the core ranking systems to retrieve current pages before the answer is composed. Keep that definition handy. It means your page is competing to be retrieved, not to be remembered.

Query fan-out is Google's term for the related queries a model issues on its own behalf after reading your question. Ask about a weedy lawn and the system may also ask about herbicide choice. One user question becomes five or ten machine questions, which is why a prompt list built from a keyword report tends to miss. The buyer's words are the first hop.

Passage or chunk is the unit that actually gets retrieved. Google describes its systems as reviewing the specific information from retrieved pages rather than whole documents, so assume the thing being judged is a few hundred words from the middle of your page, read without the navigation, the hero line or the paragraph that gave it context. None of the five engines we test publishes its chunk boundaries, so the size is an assumption on our part and not a measured figure.

Reranking is the scoring pass that orders those passages before any of them reach the model. In published retrieval architectures it is a separate model from the one writing the prose. Rerank scores are not exposed by the engines we test, so from outside you cannot tell whether a page was never fetched or fetched and then ranked below the cut. That is why we read server logs next to the answers rather than inferring the stage from the output.

Retrieval knowledge versus training knowledge is the distinction that explains why two engines contradict each other about your pricing. One answered from a document fetched a minute ago. The other answered from parameters frozen months back.

Terms for access, where the word "blocked" hides four different things

Treat "we allow AI crawlers" as a single setting and you will get this wrong. OpenAI alone documents four distinct agents, each with its own job:

  • GPTBot collects content for model training.
  • OAI-SearchBot indexes pages so they can appear in ChatGPT's search results.
  • ChatGPT-User fetches a page because a person, right now, asked something that needs

    • it. OpenAI's documentation adds two cautions: because the fetch is user initiated, robots.txt rules may not apply, and this agent is not used to decide whether you appear in Search.

  • OAI-AdsBot checks pages submitted as advertisements.

Here is the failure mode that costs real visibility. A policy decision to stay out of model training gets implemented as a blanket rule, and OAI-SearchBot goes with it. OpenAI's documentation is explicit about the consequence: a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though it can still appear as a navigational link. Nobody intended that, and nobody joins the dots, because the robots.txt edit was filed as a training question. This is documented for OpenAI's agents. Other engines publish their own agent lists and you have to read each one on its own terms. Write the two rules separately.

Content-Usage is the robots.txt directive, and matching HTTP header, coming out of the IETF AI Preferences working group. The attachment draft dated 14 September 2026 defines where the statement goes. Its companion vocabulary draft of the same date defines three categories, AI training, AI use and search, each set to allowed, disallowed or unknown, serialised as something like train-ai=y, search=n. Two things are worth knowing. It separates training from search by design, which is exactly the distinction most sites currently cannot express. And the draft says outright that whether a preference is honoured sits outside its scope. It is a stated wish, not a lock.

llms.txt is a proposed file listing your content for language models. Google is unusually direct about its own surface: you do not need to create AI text files like llms.txt to appear in Google Search, because Search itself does not use them. That is a statement about one engine. For the others there is no equivalent published statement we can point to, and from outside a site you cannot see whether the file was fetched or what was done with it, so we cannot tell you either way. Publish one if you like the tidiness, and do not book it as work done until something shows it changed a retrieval.

Terms for identity, which decide whether you are a thing at all

Entity means a resolvable thing, a company, product or person, that a system can hold as one object rather than as a string of characters. Two companies share your name, your schema says one legal name, your footer says another, and the engine hedges. A hedge reads to a buyer as absence. Organization schema done properly is the unglamorous fix.

sameAs is the schema property that points from your markup to your other profiles. It costs an afternoon, and across the audits we have run, a few dozen to October 2026, it is the identity field we most often find empty.

Citation versus mention is a distinction worth enforcing in your own reporting. A citation is a linked source under the answer. A mention is your brand named inside the prose, often with no link anywhere. A mention can move a buyer while leaving nothing in your analytics, because these surfaces mostly send no usable referrer, and that gap is a large part of why this work can look like it is doing nothing.

Corroboration is whether a source the engine already trusts repeats your claim. You do not own it, it takes the longest, and in the engagements we have run it is the layer that most often separates a brand that gets named from one that merely gets indexed.

Terms for measurement, where the abuse concentrates

Prompt set is a fixed list of buyer questions, re-run on a schedule. Fixed is the operative word. Change the prompts between runs and you have a new baseline, not a trend.

Retrieval on and retrieval off describes running the same prompt with live browsing enabled and disabled. The gap between the two answers tells you something no single number will: what the model believes about you from training alone, against what it finds when it looks. A brand that scores well only with retrieval on has a corroboration problem.

Volatility is the spread across identical runs. In the audit runs we have logged, a few dozen prompt sets through October 2026, two runs of one prompt minutes apart often come back with different sources. That is a sample drawn from our own client work, not a measurement of the engines at large, and we publish no rate for it. Take it as a reason to run each prompt several times and record the spread rather than the best result.

Visibility score names whatever weighted blend a given tool computes from the figures above. Fine as a dashboard header. Useless as a diagnosis, because a score carrying no evidence beneath it reports that a number moved and never what to change. The longer version of that argument sits in AI search is not a ranking game: there is no position eleven to climb from.

One term earns its number. GEO, generative engine optimisation, comes from the paper "GEO: Generative Engine Optimization", published at KDD 2024, which tested content tactics on a 10,000 query benchmark the authors built themselves and reported visibility gains of up to 40 percent for the strongest ones, with quotations, statistics and explicit citations at the top. The same paper reports up to 37 percent when those methods were tried on one live engine, Perplexity. Treat that as a direction of travel rather than a promise, because it was measured on a benchmark and not on your site, and because the measurements date from 2024 and the engines have moved since. The underlying point survives: concrete, attributable sentences get lifted more often than smooth ones. That is the same argument as what answer engine optimisation is, arrived at from the research side.

The check that takes ten minutes

Open your robots.txt. For each of the four OpenAI agents, and the equivalents from Google, Anthropic, Perplexity and Microsoft, write down whether it is allowed, and whether anyone intended that. Across the audits we have run, a few dozen to October 2026, the most common access finding is at least one agent blocked by a rule written for a different purpose, which is our client sample and not a survey of the web. After that, run five real buyer questions three times each on two engines and record whether you were named, whether the claim was true, and what was cited. Checking what engines say about you, properly sets out the full method, and the self-check is the ten-item version.

Honest limits

This glossary does not settle four things. The definitions of grounding, fan-out and llms.txt above are documented by Google for Google's own surfaces, and we have quoted them as such rather than as rules holding across the field, because the other engines publish less and we cannot see inside any of them. The figures attributed to our own work, the empty sameAs fields, the agent blocked by accident, the spread between identical runs, come from a few dozen client audits through October 2026 and are a sample of the sites that hire us, which skews to software and to teams that already suspect a problem, so no rate should be read off them. The GEO paper's numbers were measured on a benchmark its authors built in 2024, not on your site and not this year. And the Content-Usage drafts are drafts, so the syntax may shift, while the vocabulary draft says in terms that whether anyone honours a stated preference sits outside its scope.

FAQ

Are AEO, GEO and LLMO different disciplines? No. They are three names for the same work, from different origins: the interface, the academic literature and the technology. Pick one internally and stop arguing. What matters is whether somebody owns access, identity and corroboration.

Does blocking GPTBot remove me from ChatGPT? Not by itself. GPTBot concerns training. Appearing in ChatGPT's search answers depends on OAI-SearchBot, which is a separate user agent and a separate robots.txt rule.

Do I need schema markup to appear in AI answers? Google's own guidance says structured data is not required for generative AI search and there is no special markup for it. Schema still earns rich results and still helps an engine resolve your identity, so it remains worth doing for reasons other than the AI answer.

What is a reasonable number of prompts to track? Enough to cover the questions a buyer asks before purchase, re-run often enough that you can see a spread rather than a point. Our audits use a fixed set of about 45 prompts across five engines, run with retrieval on and off, which is a starting shape rather than a rule.

Sources

Related reading

What is answer engine optimisationAI search is not a ranking game, and what replaces itHow to check what engines say about you, properlyOrganization schema done properlyEight claims about AI search that do not survive a testDoes SEO still matter for AI search