How do AI engines decide which sources to cite?

AI engines decide which sources to cite by testing a candidate page on five things in order: whether the page can be fetched at all, whether a self-contained answer can be lifted out of it, whether the claims in it can be checked, whether its wording matches the question as the person actually asked it, and whether anything else on the web agrees with it. The test runs on passages rather than on whole sites, which is why a well-known company can be missing from an answer its own documentation could have written. The table below reads the five as a diagnosis, because the fix for a page that was never fetched has nothing in common with the fix for a page that was fetched and then passed over.

The five factors behind a citation decision, how a page fails each one, and the fix

Factor	What the engine is testing	How a page fails it	What the fix is
Retrievability	Whether this page can be fetched, and whether it sits in the index this particular engine reads	A blocked search agent, a body that only exists after JavaScript runs, or a page strong in Google and absent from Bing	Allow the search agents, serve the answer in HTML, and get the page into the index the engine actually uses
Extractability	Whether a passage answers the question on its own, without the rest of the page around it	The answer arrives in the fourth paragraph, or it depends on a sentence two sections earlier	Put the answer in the first two sentences under a heading that asks the question
Trust	Whether the claims on the page can be checked by something outside the page	Superlatives with nothing behind them, figures with no source, no author and no date	Name the standard, the number, the source and the date, in the paragraph that makes the claim
Query-language match	Whether the page uses the words the buyer used, including the sub-questions the prompt breaks into	Internal product vocabulary where the buyer has a plain-language problem	Write the heading in the buyer's phrasing, and cover the adjacent questions on the same page
Corroboration	Whether the claim, or the company making it, is confirmed anywhere else	One firm trading under two names, or a specification that appears nowhere but your own site	Keep one name and one description everywhere, and make the facts findable off your own domain

Two things follow from reading the decision this way. The factors are not weights that average out, because a failure at retrievability ends the matter before any of the others is consulted, which makes it a gate rather than a score. And every row describes a property of your own page or your own identity on the web, which is why an absent citation is usually something a company can act on rather than something the engine did to it.

The citation pipeline, in plain terms

When a buyer asks a question, the engine doesn't pick a winner from a ranked list. It runs a sequence: it gathers candidate sources it can reach, reads them for an answer it can lift, judges which it trusts, assembles a response, and names some sources. A page can be excellent and still drop out at any one of those steps. Thinking in stages is useful because the fix depends entirely on where a page falls out. (For the technical detail of how retrieval-augmented generation works under the hood, see the RAG pipeline explained; this page is about what determines selection.)

One detail of the sequence decides how much of it you can influence. The naming of sources is not a judgement the model makes at the end from memory, it is attached to the passages that were put in front of the model during generation, which means a page that was never retrieved cannot be rescued by being well written and a page that was retrieved is competing on how usable its text is. Our walkthrough of the stage of the pipeline where citation decisions are actually taken follows that handover step by step, and it is the reason the five factors below are ordered the way they are rather than ranked by importance.

Does being retrieved mean being cited?

No: being retrieved does not mean being cited, and most pages an engine fetches are never named in the answer. AirOps logged 548,534 pages that ChatGPT fetched while answering 15,000 prompts, which fanned out into 43,233 queries in total, and found 15% of those fetched pages cited in the final answer, leaving 85% read and unused (AirOps, The Influence of Retrieval, Fan-out, and Google SERPs on ChatGPT Citations, as reported by Search Engine Land; figures confirmed from two independent accounts rather than from the report itself, 5 October 2026). The same analysis put the citation rate at 18.3% for product discovery prompts, 16.9% for how-to prompts and 11.3% for validation prompts, so the kind of question being asked changes how generous the engine is with its sources.

That 85% is the working space of the four factors below retrievability. Every page in it was reachable, was fetched, and was then passed over for a reason that lives in the content rather than in the plumbing. The practical consequence is worth stating plainly, because teams get it backwards: a server log full of search-agent requests is evidence that the page was read, not evidence that it was cited. Retrieval is measurable on your own side and citation is not, so the log tells you the gate is open and nothing more. Our guide to checking whether AI can read your website sets out that log check and what it does and does not prove.

Factor 1 — Retrievability: can the engine reach the page at all?

Nothing downstream matters if the engine can't fetch the page. This means the AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended) are allowed, the content is in server-rendered or static HTML rather than JavaScript that the parser may never execute, and — critically — the page is present in the index each engine reads. ChatGPT's web search leans on the Bing index, so a page strong in Google but thin in Bing can be unreachable to ChatGPT specifically. Retrievability is binary and comes first.

Retrievability is also the only one of the five that leaves evidence on your side of the wire, which makes it the cheapest to rule out first. A month of logs with no request from OAI-SearchBot, Claude-SearchBot or PerplexityBot against a page is a retrieval failure, and a month of requests answered with 403 or 429 is the same failure wearing a different number. The other four factors leave no trace at all: the engine arrived, read the page, and chose something else, and nothing in your analytics will say so.

Factor 2 — Extractability: can the engine lift a clean answer?

Engines favor content they can quote as a self-contained unit. A section that opens with a direct answer in its first two sentences, under a clear heading, is far easier to cite than the same fact buried mid-paragraph. Structure measurably helps: 68.7% of pages cited by AI engines use logical H1→H2→H3 hierarchy (ConvertMate, 2026), and controlled testing found direct-answer formatting and added statistics raised citation rates substantially (Princeton GEO, KDD 2024). Extractability is largely a formatting and clarity property — among the cheapest factors to fix. The reason it works this way is mechanical: retrieval scores passages rather than whole documents, which is set out in how a page is split into chunks before retrieval.

What makes a passage extractable is whether it still means what it meant once it is lifted out of the page, because that is literally what happens to it. "It supports both" is a sentence that dies on removal. "The skid supports clean-in-place and sterilize-in-place cycles on the same frame" survives, carries its own subject, and can be quoted without the paragraph above it. Pronouns, "as noted above" and a figure whose units sit in an earlier sentence are the three things that most often make an otherwise good answer unquotable.

Factor 3 — Trust: are the claims verifiable?

Engines weight the sources whose claims can be checked from outside the page, because a claim nobody can check gives an engine nothing to stand behind. Specific, sourced, verifiable claims beat vague superlatives every time — "certified to ASME BPE with 21 CFR Part 11 records" is citable; "industry-leading quality" is not, because nothing in it can be checked. Named authors, dates, and references all raise trust. This is why precise, evidenced pages get cited over polished but unsubstantiated ones. Trust here means checkable, not disinterested: nothing in the pipeline asks who benefits from a claim, which is why a studio's own ranking of itself as the best AI glasses developer gets quoted back as though a third party wrote it.

Checkable is a lower bar than true and a higher bar than confident. A reader cannot verify "market-leading accuracy" at all, so an engine has nothing to weigh; it can verify "certified to ASME BPE 2022 with 21 CFR Part 11 electronic records" against a published standard, so the claim carries weight whether or not the engine does the checking. Dates and named authors work the same way, by making a claim locatable in time and attributable to someone. This is also the dimension our own content checker scores hardest, because a figure with no source in its own paragraph is the most common defect we find on otherwise strong pages.

Factor 4 — Query-language match: does the page speak the buyer's words?

Query-language match is the strongest single lever. The match between a page's vocabulary and the buyer's actual search language is the citation signal that survives statistical controls for domain authority. Discovered Labs analysed 2 million AI citations across 10,000 cited pages and reports prompt-content alignment at a standardised effect of +0.37, against +0.07 for the strongest on-page signal it tested, which was FAQ blocks (Discovered Labs, What actually drives AI citations, opened and read at source 19 October 2026). Discovered Labs sells AEO services, so the choice of what to measure is not disinterested, and that is worth knowing while reading the figure rather than instead of it. A buyer's prompt is also decomposed into sub-questions the engine answers piece by piece, so a page that covers the adjacent questions — cost, requirements, comparisons — is eligible for more of them. Speaking the buyer's language, not internal product language, is what lets a smaller site get cited over a larger one.

The reason this factor rewards breadth as well as wording is that a prompt rarely stays one query. In the AirOps data above, 15,000 prompts became 43,233 queries once the engine had broken them apart, close to three retrieval passes for every question a person typed. Each of those passes is matched against passages separately, so a page that answers the cost question, the requirements question and the comparison alongside the main one is eligible for several of them rather than one. The mechanism is set out in what query fan-out is and how it changes which pages get read.

Factor 5 — Corroboration: is the source confirmed elsewhere?

When an engine can cross-check a claim or an entity across multiple independent sources, it cites with more confidence. Being referenced consistently across the web — under one coherent company name and description — strengthens this. Fragmented identity (the same firm appearing under two names) splits the signal and weakens it. Corroboration builds slowly, but consistency of identity is something a company can stop undermining immediately.

Corroboration is the slowest factor to build and the easiest to undermine by accident. A firm that trades as one name on its website, another in its directory listings and a third in its press coverage has split its own evidence three ways, and each fragment is weaker than the whole would have been. Nothing about this requires new coverage: picking one legal name, one trading name and one sentence of description, then using them everywhere, recovers the signal already earned. Where the fragments have already produced wrong answers, why AI gets your business information wrong covers the repair.

Does how recently a page was updated change whether it gets cited?

Recency shifts a citation decision at the margin rather than at the gate, and it acts on the date a page was last updated rather than the date it first went up. The largest comparison published so far is Ahrefs' analysis of roughly 17 million cited URLs, which put the average age of a page cited by an AI assistant at 1,064 days against 1,432 days for the pages in Google's organic results (Ahrefs, New Study: AI Assistants Prefer to Cite "Fresher" Content, July 2025; figures confirmed from two independent accounts rather than from the report itself, 14 October 2026). A cited page roughly a year younger than a ranking one is a mild preference, not a filter, and an average of 1,064 days at the time of that measurement puts a page of average age in the set before ChatGPT was released. Within the same study ChatGPT leaned furthest towards newer pages, while Google's AI Overviews and organic results were the most likely to cite old ones.

The gap between the two dates is where refresh advice usually goes wrong. Seer Interactive's 2026 study of content recency separated the date a cited page was last updated from the date it was first published, and the cited set looked far fresher by update date than by publication date: 42% of the pages cited had first been published within the previous year (Seer Interactive, Study: Content Recency's Impact on AI Visibility in 2026; per-platform percentages differ between the secondary accounts of this study, so only the figure they agree on is quoted here, checked 14 October 2026). Read alongside the Ahrefs ages, that describes a cited web made mostly of old pages somebody kept current.

So the obvious shortcut is the wrong one. Moving a date without changing the page gives an engine nothing new to lift, because the passage it reads is the passage it read last time, and Google treats it as a quality problem in its own right: its 2019 guidance on dates says that artificially freshening a story without significant new information goes against its guidelines, and John Mueller has repeated that the date should move when the content substantially changes and stay put when it does not (as reported by Search Engine Journal, confirmed from two independent accounts, 14 October 2026). Recency is best read as a tiebreaker that runs after the five factors above, strongest on questions where the answer itself has a date attached. The change that earns it is a section answering something the page did not answer last quarter, a figure brought up to date with its source named in the same paragraph, or a claim withdrawn once it stopped being true. Telling a page that is genuinely sinking from an answer that is only varying between runs is a measurement problem, and how to measure AI citation decay sets out the method; the lag between making a real change and an engine noticing it is covered in how long until ChatGPT indexes a new page.

Do the four engines choose the same sources?

No: the four engines do not choose the same sources, and the overlap between them is far smaller than most teams assume. GetMentions ran 2,398 questions from 56 brand accounts on ChatGPT with web search, Google AI Mode, Perplexity and Gemini once a day for seven days in June 2026, logging 530,875 citations across 181,225 distinct URLs, and found 84% of the sources cited for a question used by only one of the four engines (GetMentions, AI Citation Volatility: A 530,875-Citation Study; figures confirmed across two independent search passes, 5 October 2026). On that evidence the sources an engine draws on for a question are mostly its own rather than shared with the other three, so the five factors are satisfied separately by each engine rather than once for all of them.

Our own measurement shows the same split from the supplier's side. In the SIGNALS four-engine study of pharmaceutical process skid fabricators, Rodem was named in 100% of ChatGPT runs and 20% of Perplexity runs for the same bioreactor query, while GEA was named in 100% of Google runs and none of the Perplexity ones (SIGNALS, 2026; the lane-by-lane figures are in AI visibility for process skid fabricators). A buyer reading one engine and a buyer reading another would describe two different markets, and a supplier reading one would draw the wrong conclusion about which factor was failing.

For a page owner the consequence is procedural rather than theoretical. There is no single answer to "are we cited", there are four, and a fix aimed at the engine that was already citing you buys nothing. Which one to work on first is a question with a real answer, and which AI engine to optimise for first works through it.

Do different AI engines decide which sources to cite in the same way?

Different engines largely do decide the same way once they are looking at the same pages, which is a separate question from whether they look at the same pages at all. The first benchmark built to hold retrieval still handed ten answer engines the same query and the same five candidate sources, and found they agreed with one another about which of the five deserved the citation at a within-group Spearman correlation of 0.704 (Elisha Bajemon and Andre-Louis Rochet, Predicting What Answer Engines Cite, arXiv 2610.38090, 28 October 2026, read at source 6 November 2026).

Two details of that figure matter before anybody leans on it. The paper reports the agreement as "the engines agree with one another at 0.704 [0.669, 0.737]" across 1,141 groups, and as 0.689 with an interval of 0.652 to 0.723 once engines from the same vendor are left out of the mean, so it is not an artefact of four of the ten engines sharing a maker. And the authors disclose a commercial interest in the answer: the paper states that "The authors are employed by TW3 Partners, which offers generative engine visibility services under the Citead name", and the same two authors wrote the earlier paper arguing that a page can be scored without an engine at all (arXiv 2609.07559, 7 September 2026). Treat it as a measurement worth having from a team adding to its own side of an argument, rather than as independent confirmation.

Held next to the 84% single-engine figure above, the two results give a usable division of the work rather than a contradiction. Engines read largely different slices of the web, and they rank what they have read in much the same way. So the part of a citation decision that turns on the page, whether a passage answers the question on its own and whether its claims can be checked, is settled in roughly the same terms by every engine and is work you do once. The part that decides whether your page is in front of the engine to be judged is each engine's own retrieval, and that is work you do once per engine.

Three limits travel with the 0.704, and the first is the largest. No engine in that benchmark searched for anything: the paper states that "Every engine receives the query and the five sources in its prompt, and no engine retrieves documents", and says of its own design that "It does not establish that the ranking transfers to engines that search the web". It therefore measures agreement inside the stage Chen and Koudas size at 2.8% of citation variation, and is silent about the 32% that sits in retrieval below it. Second, every scorer the paper tested fell short of the engines' own agreement, the best of them reaching 0.575 to 0.578 against the engines' 0.704, which is the figure to put to any vendor selling a page score as a prediction of what engines will do, and what to judge an AEO tool on covers the rest of that conversation. Third, the benchmark is English only with five candidates per query, and ranking those five by position alone scores 0.131, so the ordering it is agreeing about is not simply the order they arrived in.

Does that agreement survive when the engines retrieve pages themselves?

The agreement does not survive at anything like 0.704 once engines retrieve for themselves, and the first benchmark to test it put the figure at roughly half. LIVE-GEO published rewritten product pages as live web pages and let three engines find and rank them, rather than pasting candidate texts into a prompt, and it measured within-brand agreement between engine pairs at a mean Spearman correlation of 0.32 to 0.48 for ChatGPT against Grok, ChatGPT against Gemini and Grok against Gemini (Runze Kong, Haibo Jin, Xucheng Yu, Xiaoqi Wang, Huimin Zeng and Haohan Wang, Do GEO Gains Survive the Live Web?, arXiv 2611.01126, submitted 31 October 2026, read at source 9 November 2026). The same study found the three engines selecting the same top product for only 48.8% to 66.3% of brands.

Two further numbers from it are worth carrying around. Presentation order moved the outcome about as much as the writing did: the leading rewriting strategy took a top-one rate of 73.8% on ChatGPT under one fixed order of candidate URLs and 93.8% under the reverse, with the randomised order the benchmark uses landing at 82.5%. And a repeat run under the same presentation agreed with itself at 0.85 on ChatGPT, which is higher than any cross-engine pair in the study, so the disagreement is between engines rather than inside them.

Read next to the controlled result above, this gives a narrower claim than either paper alone supports. Cross-engine agreement about which source deserves a citation looks like a property of the test as much as of the engines: high when every engine is handed the same five candidates, roughly half that once retrieval and ordering are in play. Neither measurement cancels the other. Taken together they say that the page-level half of a citation decision is judged in broadly similar terms, and that nothing about that similarity rescues a page the engines never retrieved.

The study also contains the clearest published argument for abstaining rather than inventing a figure. Its statistics-adding strategy was flagged by a verifier on all 80 of its rewritten descriptions, and after that adjustment it scored a top-one rate of zero in every condition tested, against a verifier that matched reference labels on 90.7% of 194 numerical cases. Adding numbers works when the numbers are real and collapses completely when they are not, which is why a page here abstains when there is nothing to ground a claim on. Treat the whole paper as early evidence: it is a preprint a week old, it ran on 80 invented brands in a product-choice task rather than on real pages answering a research question, and nobody has replicated it.

When several sources support the same claim, how does an engine choose between them?

When several retrieved sources support the same claim, the engine cites some and passes over the others, and the best causal evidence on how it chooses says that presentation redistributes the credit rather than winning admission. Sriram Selvam and Anneswa Ghosh, in CITECHOICE: A Causal Audit of How Document Presentation Redistributes Citation Credit in Agentic Search, arXiv 2609.15164, 14 September 2026, took 129 multi-turn agentic-search transcripts of everyday questions, selected 113 pairs of documents that were retrieved in the same call and supported the same pre-specified fact, had 103 of those pairs confirmed by blinded human review, and then replayed each one with the target document rendered either as structured content or as prose while the rest of the transcript stayed fixed.

Two figures out of that study matter to a page owner and they point in different directions. Structured rendering raised the target's citation count by 0.50 citations per answer, with a 95% confidence interval of +0.20 to +0.84 and a Holm-adjusted p of 0.033, and it did so without raising the answer's total citations or cutting the competitor's credit, which makes it a reallocation among sources already in the answer. Whether the structured version was cited at all moved by 4.5 percentage points, which the authors themselves describe as inconclusive on a 95% interval running from -1.4 to +10.4 at a p of 0.168. Both read at source on 19 October 2026.

The useful consequence is an ordering of work rather than a formatting tip. Extractability operates on pages an engine has already retrieved: it governs how much of the visible credit you take once you are in the candidate set, not whether you get into it. So a page nobody fetches does not become fetched by being better structured, and an audit that recommends extraction fixes for a page that never appears in the logs has named the right work at the wrong stage. Ruling that out first is what checking whether AI can read your website is for.

The authors add warnings of their own on two points, position and precision. On position, they found that the observational gap between being listed first and being listed fifth was 42.3 percentage points, while controlled reordering of the same material moved citation by only 7.9 points in the main replay and 0.0 points in the held-out set, so most of the apparent advantage of being first is not caused by being first, and in any case it is not a lever a publisher sets. On precision, the study found that re-decoding 30 frozen cases changed 15% of the individual cite-or-not decisions, with decoding accounting for an estimated 45% of the variance in a single generation's effect. It is one small-team preprint on 113 pairs, and the authors are explicit that it does not establish reliable source admission or a general rank advantage.

Does the page or the engine's own retrieval decide more?

The engine's own retrieval decides more than the page does, and the largest study built to separate the two says so. Mahe Chen and Nick Koudas of the University of Toronto, in What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), arXiv 2610.15033, 10 October 2026, took two 90-day collections of answers to commercial questions in contrasting markets, 158,099 answers carrying 468,037 citations, and replayed every question through the engines' own APIs to record the sub-queries each engine wrote, the pages its search retrieved, and which of those it went on to cite.

What makes it worth reading over the rest is that it compares pages on the same website, which is the only design that tells "this page is well built" apart from "this site is trusted". Within a single site, the study found that the pages cited more often are the ones whose titles and paragraphs match the questions people actually ask, while FAQ blocks, structured data, summary boxes and page speed appear effective only across websites, travelling with the kind of site that has them rather than predicting which page on a site gets cited. Its abstract puts the retrieval step at 32% of the variation against 2.8% for page properties, and reports that every page ChatGPT cited was one its own search had retrieved. Read at source on 19 October 2026.

The finding is narrower than it first looks. It does not say page work is wasted, because the one page-level property that separated pages on the same site is the vocabulary match that Factor 4 above is about, so the lever this page calls the strongest is the lever the study keeps. What it withdraws is the idea that a checklist of on-page features predicts which of your pages gets cited. It also explains the engine-by-engine split further up: citations track each engine's own retrieval ranking and rarely coincide with what Google or Bing return for the same question.

Its other half is about memory rather than retrieval, and it is the part that explains answers naming a company with nothing of that company's cited underneath. The study found that ChatGPT wrote brands it already knew into its own sub-queries, preferred websites it knew, and named brands with no retrieved page behind them, and that in controlled experiments the brand name reordered its ranking of products whose facts were identical. The paper's own summary of the split is that pages are cited when they are searched and brands are also named when they are remembered. The second half of that is why third-party mentions matter for AI visibility even when they send no traffic at all.

If an engine does cite your page, does it go on to name your company?

Often, and the step from a cited page to a named company has now been measured instead of assumed. Benjamin Tannenbaum, in From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search, arXiv 2609.23162, 19 September 2026, took 34,960 unbranded prompt-engine observations from 75 anonymised projects on the Aiso platform, covering 2,854 prompts and repeated GPT and Gemini runs from June to September 2026, logging for every run whether the brand's own domain showed up in the retrieval path the engine exposed. Where neither the brand nor its own domain appeared there, the brand was named in 2.8% of GPT answers and 3.8% of Gemini ones. Where its own domain was cited, the same brands were named in 49.0% and 58.4%. Read at source on 28 October 2026.

A second condition in that study sits outside the page entirely. The engines rewrite a question into several sub-queries before searching, and sometimes one of those carries the brand's name, which the paper calls a branded fan-out. Where the brand's own domain was cited and a branded fan-out also occurred, mention rates reached 91.4% on GPT and 100% on Gemini. The 100% is one cell of a crosstab rather than a law, and the mechanism it points at is query fan-out: what the engine decides to go looking for, which no page controls.

Brand mention rate by what the engine's own retrieval path contained, from 34,960 unbranded prompt-engine observations (Tannenbaum, arXiv 2609.23162, read at source 28 October 2026)

What the retrieval path showed	Brand named, GPT	Brand named, Gemini
Neither the brand nor its own domain	2.8%	3.8%
Own domain cited, no branded sub-query	49.0%	58.4%
Own domain cited and a branded sub-query	91.4%	100%

The comparison that carries the most weight is made inside a single prompt. Among prompt cells that varied in whether the brand's own domain was cited while no branded fan-out occurred, the study found a mean mention-rate increase of 40.2 percentage points on GPT and 49.0 points on Gemini, which holds the project, the prompt and the engine fixed and so is not a comparison between well-known and obscure companies (same source and date).

The study's other finding is less comfortable and worth reading next to it. Prior visibility persisted on its own: a brand that went unmentioned last run and had no own-domain citation this run was named in 1.6% of GPT answers and 1.9% of Gemini ones, while a brand mentioned last run that was also cited this run was named in 80.5% and 83.7% (same source and date). Whether that is memory, a stable retrieval pattern or both, the practical reading is that the first citation is harder to win than the tenth, and that a brand with no presence anywhere is starting from under 2%. That is the case for third-party mentions stated in numbers rather than in principle.

Three limits belong with those figures. The paper describes its own model as predictive and observational rather than a causal description of proprietary engine internals, so none of this is the selection rule. It is one author, two engines, and one platform's monitored projects, a sample of companies already paying somebody to watch their AI visibility. And being cited and being named are separate outcomes, which is the distinction mention rate against citation rate works through.

What do the engines themselves document about how sources are chosen?

All four engines publish documentation about reaching a page and being eligible to appear, and none of them publishes how the final citation is chosen. That division is worth knowing before you read anybody's advice on this subject, including ours: the first of the five factors above can be settled by reading the vendors' own pages, and the other four have to be measured, which is why they are evidenced here with studies rather than with documentation. The table below is what each vendor actually states, and what it stops short of.

What each engine's own documentation covers, and what it leaves unsaid about citation

Engine	What the vendor documents	What it does not document
Google AI Overviews and AI Mode	That no special markup, schema type or llms.txt file is needed; that a page must be indexed and eligible to be shown with a snippet to be eligible as a supporting link; that both features may use a query fan-out technique across subtopics and data sources; and that nosnippet, data-nosnippet, max-snippet and noindex all apply (Google Search Central, AI features and your website)	Which eligible page becomes a supporting link for a given question. Google states plainly that meeting every requirement does not mean a page will be crawled, indexed or served
ChatGPT	Four agents with separate jobs: GPTBot for training, OAI-SearchBot to gather pages so search features can list sites, ChatGPT-User for fetches a person's question or a GPT Action triggers, and OAI-AdsBot, which "only visits pages submitted as ads". Of those, user-initiated fetches sit outside robots.txt, which OpenAI states as "Because these actions are initiated by a user, robots.txt rules may not apply", and a robots.txt change takes about 24 hours to register (OpenAI, crawler documentation)	Which of the pages it fetched gets named in the answer, and how the search layer ranks what it retrieves
Claude	ClaudeBot for training, Claude-SearchBot for indexing for search features, and Claude-User for user-directed fetches; that blocking Claude-SearchBot may reduce a site's visibility and accuracy in search results; and that Crawl-delay is honoured for ClaudeBot (Anthropic support, crawlers and how to block them). Separately, Anthropic states that a noindex tag tells its search partners not to send a page in response to a web search query (Anthropic support, report, block and remove content from Claude)	How a retrieved passage becomes a citation, and which partner index served it
Perplexity	PerplexityBot, which surfaces and links sites in results and is not used to crawl for foundation models, and Perplexity-User, which fetches a page because a person asked something and generally does not apply robots.txt. Allowing PerplexityBot and its published IP ranges is the documented way to be eligible (Perplexity, crawler documentation)	The ranking or selection method behind a cited answer, which Perplexity's crawler pages do not address

Documentation read 9 October 2026, and the ChatGPT row rechecked at source on 2 November 2026, when OpenAI documented a fourth agent and the robots.txt exemption above. Vendors revise these pages without notice, so treat the dates as the point the claim was checked rather than as settled fact.

Which part of a citation decision is documented, and which is inferred?

The documented part of a citation decision is the gate: crawler access, indexing, and eligibility to be shown with a snippet. The choice itself is documented by nobody, which sets the honest division of labour on this page, where retrievability is a reading task and the other four factors are a measurement task. It also explains why so much writing about AI citation sounds confident and disagrees with itself. Once past the documented gate, everybody is inferring from samples, ours included. The studies cited above are sampling behaviour at a point in time rather than reporting a published rule, and a vendor can change the unpublished half of this without telling anybody.

Two of the documented lines are more useful than they look. Anthropic's instruction about noindex is an admission of architecture: Claude's web search reaches publishers partly through search partners, so a page's standing with Claude depends on an index its owner never deals with directly. And Google's eligibility rule cuts both ways, because a page that carries nosnippet to keep its text out of search previews has also removed itself from the supporting links under an AI Overview. Both are documented levers rather than inferred ones, and both are things a site can get wrong by accident while tidying something else. Which index sits behind each assistant is a separate question with a moving answer, and whether ChatGPT searches with Google or Bing works through the current position. Which agents to allow, and what each one costs you, is the subject of whether you should block AI crawlers.

Why this maps to a fixable diagnosis

Every absent citation has a locatable cause, which is what makes this a fixable diagnosis rather than a mystery: each cited page passed all five factors and each absent one failed at least a single named factor, so "why aren't we cited?" has a specific answer. The SIGNALS framework scores pages across the dimensions behind these factors — Structure, Grounding, Alignment, Substantiation and the rest — to isolate the failing stage rather than guessing. The fix for a retrievability failure (an index gap) is nothing like the fix for an alignment failure (wrong vocabulary), so naming the stage is what makes the work efficient.

The order to work them in follows from the gate. Retrievability comes first because nothing downstream is even consulted until it clears, and because it is the one failure your own server logs will confirm in an afternoon. Extractability comes second because it is the cheapest to change: the facts are already on the page and the work is moving the answer to the top of a section and giving the section a heading that asks the question. Query-language match is third, and it is the one that usually needs a decision rather than an edit, because matching the buyer's words means giving up some of the internal vocabulary a company is attached to. Trust and corroboration come last, not because they matter least, but because they accumulate: a sourced figure added today is working a year from now, and a consistent company name pays back over the same horizon rather than this week.

What else do people ask about how AI engines choose sources?

What else do people ask about retrieval, recency and engine overlap?

Related guides

• The RAG pipeline explained — the technical mechanism behind these factors, including at what stage of the pipeline citation decisions are actually taken.
• How to get cited by ChatGPT and how to appear in Google AI Overviews — applying the factors per engine.
• AI citation checklist: 25 signals that matter.
• The 7 SIGNALS dimensions — the scored version of these factors.
• Does AI-generated content get cited by AI search engines?