Guide · 7 min read · updated 2026-08-22

How answer engines pick sources: documented vs guessed

No answer engine publishes a citation ranking function; here is exactly what each one does document, and the short checklist that survives the evidence test.


No answer engine publishes a ranking algorithm for citation selection. Google, Microsoft, OpenAI, Anthropic and Perplexity all document eligibility and controls. None documents ordering. Every "GEO ranking factor" beyond the eligibility statements below is inference, and most of it is inference from small correlation samples against systems that are documented to be non-deterministic.

That is not a reason to do nothing. It is a reason to know which of your decisions rests on a vendor sentence and which rests on a guess. This page separates them, engine by engine, with the source for each claim.

Google: the best-documented, and the most surprising

Primary source: Google Search Central, AI features and your website, last updated 2025-12-10.

"To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements."

Read the last sentence twice. Google is not describing a new surface with new prerequisites. It is saying the prerequisite set is unchanged. If you are indexed and snippet-eligible, you are eligible, and there is no separate AI checklist to satisfy.

Three further mechanisms are documented:

  • Query fan-out. Both surfaces "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response." Google adds that this lets it "display a wider and more diverse set of helpful links associated with the response than with a classic web search." The practical implication is that the entry query is not the only query you are competing on, and breadth of subtopic coverage plausibly matters more than it does in classic search. Note that the second half of that sentence is our inference; Google states the mechanism, not the optimisation.
  • AI Overviews are conditional. They are "only shown when our systems determine that it is additive to classic Search, and as such, often don't trigger." A query that shows no AI Overview today may not be evidence of anything about your site.
  • The two surfaces differ. AI Mode and AI Overviews "may use different models and techniques, so the set of responses and links they show will vary." Testing one tells you little about the other.

On controls, robots.txt for Googlebot is the access control; noindex, nosnippet, data-nosnippet and max-snippet limit what can be shown. Google-Extended covers Gemini training and grounding, and Google states it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search." Measurement rolls into Search Console's Performance report under the Web search type.

Google's entire structured-data instruction for AI features is one line: "Making sure your structured data matches the visible text on the page." We cover what that does and does not license in structured data for AI answers.

One vendor assertion to treat as such. Google writes: "We've seen that when people click from search results pages with AI Overviews, these clicks are higher quality (meaning, users are more likely to spend more time on the site)." No figures accompany it. It is a claim by an interested party, not published data, and should not be cited as evidence of anything.

Microsoft: the only first-party citation data

Primary source: Introducing AI Performance in Bing Webmaster Tools, 2026-02-10.

Microsoft is the only vendor shipping a publisher-facing citation dashboard. The public preview reports total citations, average cited pages per day, a sample of the "grounding queries" the AI used when retrieving cited content, page-level citation activity, and visibility trends over time. Coverage spans Microsoft Copilot, AI-generated summaries in Bing, and "select partner integrations".

And then Microsoft undercuts the obvious misreading, in its own words: citation counts "reflect how often pages are cited, not page importance, ranking, or placement", and average cited pages "does not indicate ranking, authority, or the role of any page within an individual answer." The one available citation metric is deliberately not a ranking signal.

Microsoft's content guidance, framed as suggestions rather than requirements, is the closest thing to published advice anyone offers: strengthen depth and expertise; improve structure and clarity via "clear headings, tables, and FAQ sections"; support claims with evidence ("examples, data, and cited sources"); keep content fresh and accurate; and reduce ambiguity across formats, aligning "text, images, and video so they consistently represent the same entities". IndexNow is recommended as the freshness signal. Microsoft also states that "Bing respects all content owner preferences expressed through robots.txt and other supported control mechanisms."

Anthropic: an architecture, not a ranking

Primary source: Anthropic's web search tool documentation.

The documented flow has three steps: Claude decides when to search based on the prompt; the API runs the searches and provides the results, a process that "can repeat multiple times throughout a single request"; Claude then answers with cited sources. Anthropic also documents when Claude searches at all — recent events, current prices or statistics, information about organisations, people or products that might have changed, and explicit requests — and when it does not: established facts, mathematics, coding concepts, creative work, and analysis of content already in the conversation.

The structurally important fact sits in the tool versioning. From web_search_20260209 onward, Claude writes and runs code that filters search results before they reach the context window, which Anthropic calls dynamic filtering. Relevance selection is therefore model-authored program logic executed per request, not a static scoring function applied to a candidate set.

There is no stable ranking surface to optimise against, and that is a design property rather than an omission. Two levers survive: be present in the underlying search index, and be cleanly machine-parseable so that a filtering program written on the fly can find and extract the relevant span. The domain controls Anthropic documents (allowed_domains and blocked_domains) are consumer-side settings for the org or agent, not publisher-side controls.

OpenAI and Perplexity: eligibility only

OpenAI's bots documentation is precise about which token governs appearance: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links." The stated route in is to allow OAI-SearchBot in robots.txt and allow requests from the published IP ranges, with roughly 24 hours for a robots.txt change to take effect. ChatGPT-User "is not used to determine whether content may appear in Search". Crawls may be shared between tokens to avoid duplicative fetching. On ranking or citation selection, OpenAI publishes nothing at all.

Perplexity's crawler documentation is the same shape: PerplexityBot "is designed to surface and link websites in search results on Perplexity. It is not used to crawl content for AI foundation models." Allowing it plus the published IP ranges is the stated route to appearing, with changes taking up to 24 hours. Nothing on ranking, ordering or selection criteria.

What is not published

No vendor publishes any of the following. Any tool that claims to measure them is inferring, and should say so.

  • Citation-ranking weights or ordering functions, for any engine.
  • Whether specific schema.org types influence citation selection. The sole exception is Apple's isAccessibleForFree, which is documented as consumed, and as a suppression signal rather than a promotion signal.
  • Whether llms.txt affects citation likelihood in any production answer engine.
  • Whether content freshness, author markup, or outbound citations measurably change citation rate.
  • Any domain-authority equivalent for answer engines.

The checklist that survives the evidence test

Strip out everything unsourced and this is what remains. Each row names the vendor that documents it, so you can audit the claim rather than trust the table.

Do thisDocumented byWhat it buys
Allow the right token per engine, retrieval separately from trainingOpenAI, Anthropic, Google, Apple, Perplexity, Amazon, Mistral, DuckDuckGoAccess. The precondition for everything else. See the crawler registry.
Stay indexed and snippet-eligibleGoogleStated to be sufficient for AI Overviews and AI Mode eligibility, with no additional requirements.
Keep important content in server-rendered textGoogleSnippet eligibility depends on extractable text; most AI fetchers do not execute JavaScript.
Use clear headings, tables and visible FAQ sectionsMicrosoftThe only vendor-endorsed content-structure guidance in existence.
Keep structured data consistent with visible textGoogleGoogle's entire structured-data instruction for AI features. Mismatch is a policy violation.
Keep content fresh, and push updates via IndexNow where supportedMicrosoftFaster reindexing, which is a prerequisite for being retrieved on time-sensitive queries.
Verify your edge is not blocking the fetchers you allowedCloudflare (mechanism), operator IP lists (verification)Prevents the failure where robots.txt says yes and the WAF says 403. See AI crawler blocking at the edge.

How to read the next GEO article you find

Three questions separate evidence from folklore, and they are quick to apply. First, does the claim quote a vendor sentence, or a correlation study? Second, if it quotes a vendor, does the quote say what the article says it says — Google's fan-out statement, for example, describes a retrieval mechanism and is routinely rewritten as advice about content length. Third, would the claim be falsifiable given that AI Overviews "often don't trigger" and Claude's filtering is written per request? A great deal of published GEO advice fails the third test before you reach the first.

Scan your site to check the parts of this that are actually checkable: whether every retrieval token is allowed, whether your pages are indexable and snippet-eligible, whether your body text survives without JavaScript, and whether your structured data agrees with what a reader sees. The census shows how the same checks land across sites already measured.

Common questions

Do I need to do anything special to appear in Google's AI Overviews?

Google says no. Its AI features documentation states that to be eligible as a supporting link in AI Overviews or AI Mode, "a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements." That last sentence is the strongest statement any vendor has published, and it contradicts most of what is sold as AI search optimisation.

Is there a ranking factor list for ChatGPT citations?

No. OpenAI documents eligibility only: allow OAI-SearchBot in robots.txt and allow requests from its published IP ranges. It publishes nothing about how candidates are ordered or selected. Any list of ChatGPT ranking factors you encounter is inference from correlation studies, not vendor documentation.

Does Bing's citation count tell me how I rank in AI answers?

Microsoft says explicitly that it does not. The AI Performance dashboard states that citation counts "reflect how often pages are cited, not page importance, ranking, or placement", and that average cited pages "does not indicate ranking, authority, or the role of any page within an individual answer." It is the only first-party citation data any engine gives publishers, and the vendor has pre-emptively denied it is a ranking signal.

Can I optimise for how Claude selects sources?

Not in the sense of a ranking function. From the web_search_20260209 tool version onward, Anthropic documents that Claude writes and runs code which filters search results before they reach the context window. Relevance selection is model-authored program logic executed per request, so there is no stable surface to optimise against. The two levers that remain are being present in the underlying search index and being cleanly machine-parseable so a filtering program can extract the relevant span.

Where can I actually measure whether I am being cited?

Two places, both free and both first-party. Google Search Console rolls AI Overviews and AI Mode appearances into the Performance report under the Web search type. Bing Webmaster Tools AI Performance, in public preview since 2026-02-10, reports total citations, average cited pages, sampled grounding queries and page-level citation activity across Copilot, Bing AI summaries and select partner integrations. Both require verified site ownership.

Keep reading