E-E-A-T Signals in AI Search

A rating framework, not a mechanism

E-E-A-T stands for Experience, Expertise, Authoritativeness and Trust. It comes from Google's Search Quality Rater Guidelines, the manual issued to the external human contractors Google pays to rate sample search results. Raters apply it when judging Page Quality. It was E-A-T from 2014; Google added the leading Experience — first-hand or life experience with the subject — on 15 December 2022.

What it is not is a score. From the guidelines themselves, in the version dated 11 September 2025:

No single rating can directly impact how a particular webpage, website, or result appears in Google Search, nor can it cause specific webpages, websites, or results to move up or down on the search results page.

Google's publisher-facing documentation says the same thing from the other direction: while E-E-A-T itself is not a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful. And on the internal hierarchy, both documents agree — the guidelines call Trust the most important member at the center of the E-E-A-T family, and the publisher docs repeat that trust is most important while the others contribute to it.

So: raters apply E-E-A-T, ratings do not move rankings, and Google builds systems it hopes correlate with what raters reward. Everything downstream of that is inference, and it was inference for eight years before AI search existed. The question this page answers is narrower and harder. Given that AI answer systems select sources, is there any evidence they select on anything resembling E-E-A-T, and is there anything a publisher can do about it?

The eligibility bar Google actually published

Google's stated rule for appearing in its AI features contains no quality language at all. From AI features and your website, stamped 10 December 2025:

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.

That is a technical gate: indexed, snippet-eligible. The quality judgment happens upstream, in ordinary ranking, because the grounding corpus for AI Overviews and AI Mode is the ordinary Google Search index. There is no separate AI quality assessment documented anywhere by anyone.

Which yields the strongest defensible pro-E-E-A-T argument available, and it is structural rather than direct. Whatever E-E-A-T-adjacent signals influence ordinary ranking are inherited by AI features, because AI features draw from the same ranked index. That follows from documented architecture. It is not a Google statement about AI and E-E-A-T, because no such statement exists, and the difference between those two things is exactly what separates an honest page on this subject from a confident one.

Google does state one differential, and it is worth having precisely. Its systems give even more weight to content that aligns with strong E-E-A-T for topics that could significantly impact the health, financial stability or safety of people. That is the YMYL carve-out, it is documented, and it is about Search rather than about AI features.

The document that should have said it, and did not

In May 2026 Google published its first official guide to optimizing for generative AI features. It is the natural place to write E-E-A-T matters more than ever for AI. It contains a whole section on myths, listing things publishers are being told to do that Google says are unnecessary. E-E-A-T does not appear in it. Not as advice, not as a myth, not once.

The nearest thing anywhere in Google's AI documentation is the AI features page's standing recommendation to create helpful, reliable, people-first content — which is Google's ordinary phrase for ordinary Search, and has been since 2022.

Every non-Google operator is more silent still. OpenAI's only ranking statement is that ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information, and that placement is not guaranteed; its publisher guidance is entirely about crawler access — allow OAI-SearchBot, permit the IP ranges — with no mention of authorship, credentials, bylines or expertise. Anthropic publishes crawler documentation and nothing about source quality. Perplexity publishes crawler documentation and nothing about source quality. Microsoft rewrote the Bing Webmaster Guidelines in February 2026 to cover Copilot, and the guidance is Bing's ordinary quality guidance.

Four operators, one open question, zero documentation. No one documents whether an author entity is identified or scored, whether credentials or professional bodies are recognized, whether an About page or editorial policy changes retrieval, whether site-level reputation carries across topics, or how any of this differs between AI features and ordinary ranking.

Every E-E-A-T-for-AI checklist in circulation is filling that silence with the pre-AI E-E-A-T checklist and relabeling it. The pre-AI checklist — author bios with credentials, an author archive page, an About page, an editorial policy, outbound citations to authoritative sources, Person and Organization markup — was itself always an inference about what might correlate with what raters reward. Carried into a context where nobody can check it, and sold as an AI-visibility deliverable, it becomes an inference about an inference with a price attached.

What the citation data actually shows

This is the uncomfortable part. If AI systems selected on expertise and authoritativeness the way practitioners describe, the most-cited domains would look like credentialed publishers. They partly do and partly very much do not.

The Pew Research Center published the only large independent non-vendor study of the question on 22 July 2025, tracking 900 US adults across 68,879 unique Google searches in March 2025, of which 12,593 produced an AI summary. Wikipedia, YouTube and Reddit together supplied 15% of AI summary sources, against a 17% share of standard results. News sites were 5% in both. AI summaries were not, on that measure, markedly more credentialed than blue links.

The same study found the one clean piece of evidence pointing the other way: government sites were 6% of AI summary sources against 2% of standard results, a threefold over-representation. That is a real, measured signal that something trust-shaped operates in AI source selection, and it is routinely left out of both the pro- and anti-E-E-A-T cases.

Platform-level data agrees on the direction. Profound, analyzing 680 million citations from August 2024 to June 2025, found Perplexity's single most-cited domain was Reddit at 6.6%. ChatGPT's was Wikipedia at 7.8%, with Reddit at 1.8%. In AI Overviews, Reddit, YouTube and Quora took 2.2%, 1.9% and 1.5% respectively.

Reddit is anonymous user-generated content with no bylines, no credentials and no About page. On the framework's own terms it scores near zero on Expertise and Authoritativeness. It is nonetheless among the most-cited sources on every platform anyone has measured.

Reddit refutes the checklist, not the framework

Both of the loud readings of that finding are wrong, and it is worth being precise about why.

The reading that says Reddit proves quality is irrelevant to AI systems ignores the government-site figure and the fact that a framework with four components is not falsified by a source scoring low on two of them. The reconciliation available inside Google's own framework is that Experience — first-hand accounts from people who actually did the thing — is precisely what a forum supplies, and that Trust in the guidelines is about reliability rather than about credentials. Google added Experience to the framework in December 2022, before any of this was measurable, and a forum thread is the purest form of it.

That reading is coherent. It is also retrofitted, and honesty requires saying so. Nobody predicted Reddit's prominence from the framework; the framework was stretched to accommodate it afterward.

What Reddit's prominence does refute is the checklist version of E-E-A-T — bylines, credential displays, author schema, an About page — because the most-cited sources on several platforms have none of it. The claim that author bios and named-expert bylines increase AI citations is unsupported: no operator documents author-level signals, and no published controlled test isolates authorship and measures citation change. Google's own guidance is conditional, encouraging accurate authorship information such as bylines where readers might expect it — a reader-trust argument, not a machine-parsing one. Google has never said every page needs a byline, and a byline on a product category page signals nothing to anyone.

The Person and Organization schema version of the argument fails twice over. Google says there is no special schema.org structured data you need to add for AI features, and the one controlled test of adding JSON-LD — 1,885 pages against roughly 4,000 matched controls, published 11 May 2026 — found citation changes of −4.6% on AI Overviews, +2.4% on AI Mode and +2.2% on ChatGPT. Separately, and more fundamentally, markup does not create expertise. It asserts it.

Source preferences move by a factor of five in a fortnight

There is one further finding that should be attached to any strategy built on observed source mixes, and it is rarely mentioned.

Semrush, analyzing more than 230,000 prompts and over 100 million citations between 14 July and 12 October 2025, recorded ChatGPT's Reddit citation rate falling from around 60% of responses to around 10% in mid-September 2025. Wikipedia fell from about 55% to under 20% over the same window. AI Mode and Perplexity were unchanged.

No announcement accompanied it. A single platform's source preferences moved by roughly a factor of five inside a fortnight, and the only reason anyone knows is that a vendor happened to be sampling continuously through the period.

Anything built on a snapshot of current platform source preferences has an unknown and possibly short half-life. That applies to the Reddit findings above as much as to any tactic derived from them. The Pew and Profound figures are accurate as measurements of the periods they cover and should be dated every time they are used — Pew for March 2025, Profound for August 2024 to June 2025. They are not standing descriptions of how these systems behave in August 2026.

What survives the evidence

Very little of the E-E-A-T-for-AI advice market survives, and what does survive is unglamorous.

• The inheritance argument holds. AI Overviews and AI Mode require a page to be indexed and snippet-eligible in Search, and they are grounded on the Search index. Whatever quality signals move ordinary ranking carry through. That is an argument from architecture and it is the honest floor of the case for doing quality work.
• E-E-A-T is not a score and cannot be audited as one. Google states it is not a ranking factor and that no single rating moves a result. Third-party E-E-A-T scores, authority scores and trust scores are vendor constructs. They may correlate with something useful; they are not measurements of anything Google computes.
• Author attribution is a reader decision, not a machine one. Google's guidance is explicitly conditional on reader expectation. Add bylines where a reader would look for one and skip them where nobody would.
• The YMYL differential is documented and worth respecting. Google says its systems weight E-E-A-T-aligned content more heavily for health, financial stability and safety topics. That is the one place a differential is stated in writing.
• First-hand experience is retrievable in a way credential displays are not. That is the reading of the Reddit data that actually generalizes. An account of what happened when someone did the thing produces text a retrieval system can match against a question. A credential in a sidebar does not.

An E-E-A-T audit for AI search that is the 2019 E-A-T checklist with the word AI added is selling a list of items none of which is documented against any AI system — and the single item that has been tested, structured data, came back null. The most useful thing anyone can say about E-E-A-T and AI search in August 2026 is what Google did not say in May 2026: it wrote the definitive document on optimizing for generative AI features, included a section debunking myths, and never used the term. That absence is information, and it points the same direction as everything else on this page.

Frequently asked questions