Being cited by an AI: the four signals ChatGPT, Perplexity and AI Overviews actually look at

An answer generated by ChatGPT, Perplexity or Google's AI Overviews usually cites between three and eight sources. Dozens of pages covered the same subject. The question is therefore no longer "am I indexed?" but "why that page and not mine?". The selection models are not known in detail, but four reliable references let you reason about them: Google's official documentation on its AI features, OpenAI's and Perplexity's on their robots, and the academic work that measured what makes a page's visibility vary inside a generated answer. Those sources converge on four signals.

Signal 1: the page is reachable by the right robot, not only by Googlebot

This is the precondition many sites fail without knowing. Generative engines do not all use the same robot as the classic search engine, and one operator sends several. OpenAI documents three distinct agents: GPTBot for training, OAI-SearchBot for the ChatGPT search index, and ChatGPT-User for fetching a page on a user's request. Each is controlled separately in robots.txt, as the OpenAI bots documentation explains. Anthropic does the same with ClaudeBot, Claude-SearchBot and Claude-User, and Perplexity with PerplexityBot and Perplexity-User.

In practice, a site that blocked "all AI robots" in 2024 as a precaution often blocked the ones that answer users too. It can no longer be cited, whatever the quality of its content. Reading server logs is the only way to know who actually passes through; we devote a separate article to identifying AI robots.

Signal 2: the engine selects passages, and every section has to be one

Google is explicit on this in its documentation on AI features in search: AI Overviews and AI mode rest on the same systems as classic search, and there is "nothing special to do" beyond the usual recommendations. But the way those systems read a page has changed. Google describes a query fan-out technique: the user's question is broken into several sub-queries, each run in parallel, and the answer is assembled from the most relevant passages for each sub-question, as the AI Mode announcement explains.

The unit of selection is therefore no longer the page but the passage. A paragraph that answers a precise sub-question cleanly stands a better chance of being picked up than a long page where the answer is diluted. That argues for writing where every section carries an identifiable question, opens with the answer, then develops the qualification. It is the subject of our article "Answer first, qualify second".

Signal 3: the content carries quotable elements

Research has measured this signal better than the others. In 2023 a team from Princeton, Georgia Tech and the Allen Institute published GEO: Generative Engine Optimization, a study testing nine ways of rewriting the same page and measuring their effect on that page's visibility inside generated answers. Three methods stand out clearly: adding source citations, adding statistics, and adding attributed direct quotations. The authors measure up to 40% more visibility, depending on the rewriting method applied. Keyword stuffing, the historic SEO technique, brings nothing and sometimes degrades the result.

The reason is easy to see: a model producing a reliable answer prefers to lean on a passage stating something precise, dated, quantified and attributed, rather than on a generality. A figure with its source is a fragment the AI can reproduce as is while saying where it came from. A sentence like "the market is growing strongly" is not.

Signal 4: the source is identifiable and consistent

Google's systems have for years assessed what the search quality rater guidelines call experience, expertise, authoritativeness and trustworthiness, grouped under the acronym E-E-A-T. It is not a ranking factor in the strict sense, but it is the frame within which Google trains and checks its systems. And those criteria bear on the source as much as on the page: who writes, with what standing, and whether the site holds a consistent line from one page to the next.

For a generative engine, consistency counts twice over. A site giving three different figures for the same data across three pages is a risky source. A site where every article is signed by an identified person, with an author page and a stable field of expertise, is a source the system can attribute. Google's Article structured data documentation sets out the fields that make that attribution machine-readable: author, publication date, modification date.

What these four signals change in production

Put end to end, these signals sketch a method with little left in common with volume SEO:

• Check access before writing: robots.txt is read at the domain root, and each AI robot has its own token. Blocking training does not oblige you to block citation.
• Write in passages: one question per section, the answer up front, the qualification after.
• Source every claim that matters: a figure, a date, an origin. It is the strongest signal research has measured, and it is also what protects a publisher against error.
• Sign and hold a line: an identified author, stable vocabulary, data that does not contradict itself from page to page.

None of this is a technical trick. These are editorial requirements, and they only hold at scale when they are built into the production tool rather than checked by hand afterwards. In practice: content organized into coherent clusters, one source per claim, one signature per article.

What we do not know, and should say

These signals are documented, but they do not add up to a formula. Generative engines change fast, their indexes do not overlap, and the same question asked twice can produce two different lists of sources. The only honest way to steer your presence is to measure it over time: put the same questions to the same engines regularly, note who gets cited, and separate pages merely retrieved from those genuinely attributed. We describe that measurement method in an article on share of voice inside AI answers.