Getting Cited by AI: How ChatGPT, Perplexity and AI Overviews Choose Sources

AI search visibility is a different problem from ranking, and AI assistants changed its shape. In classic search, the goal was a position on a results page. In an AI answer, there are no positions — a page is either retrieved and quoted, or it is not mentioned at all. That makes the mechanics familiar but the failure modes different.

Most of what determines whether you get cited is ordinary technical hygiene. A few parts are specific to how retrieval-augmented systems work, and one of them is a mistake that removes sites from AI search results without any visible symptom.

Ranking versus retrieval

When someone asks an assistant a question, the system usually does three things: it converts the question into a search, it retrieves a small set of candidate pages, and then it reads those pages and writes an answer with citations. Your page has to survive all three steps.

This is why a page can rank well on Google and still never be quoted. Ranking gets you into the candidate set. Being quotable is what gets you into the answer. They are related properties, but they are not the same one, and the second is the one most sites neglect.

The robots.txt layer, and the mistake in it

AI systems reach pages through named crawlers, and the names matter because they do different jobs:

The mistake is treating these as one group. A site that writes a single wildcard rule blocking all of them has decided not only “do not train on our content” — a defensible position — but also “do not show our content in AI answers”, which for most B2B companies is a straightforward commercial loss. If you want to opt out of training but still be cited, block GPTBot, ClaudeBot, CCBot and Applebot-Extended, and leave OAI-SearchBot, PerplexityBot, ChatGPT-User and Bingbot alone.

A related trap: Google-Extended does not control Google Search. Blocking it removes your content from Gemini and Vertex grounding while leaving your ordinary Google rankings untouched. Some publishers block it believing they have opted out of Google entirely, and are surprised to still be ranked.

Note on llms.txt: it is a proposed convention, not an adopted standard, and there is no evidence major assistants use it as a selection signal today. Publishing one is cheap and harmless. Treating it as a strategy is not.

What makes a page quotable

Once a page is in the candidate set, an extraction pipeline reads it and decides what to use. Pages that get quoted share a set of properties:

• A stated fact early in the document. A number, a date, a definition, a named entity — in the first few hundred words.
• Attribution. “Approximately 55% of measured search traffic in China, according to StatCounter-derived estimates” is quotable. “Most of the market” is not, because nothing in it can be carried into an answer.
• Headings that read like questions or topics. H2 and H3 headings that paraphrase how people actually ask things make the relevant passage easy to isolate.
• Tables and lists. Structured comparison is the single most extractable content format, and it is the format that most often survives into an AI answer.
• Server-side rendering. If the text only exists after JavaScript runs, many extraction pipelines see an empty document. This is the most common reason a modern site is invisible to AI systems despite looking fine in a browser.
• Stable URLs and visible dates. Machines weight recency, and they need a URL they can cite and a date they can attach to it.

Entity clarity

Before quoting a page, a model has to know who is behind it. Ambiguity here is expensive, and it is where small companies lose to large ones for reasons that have nothing to do with content quality.

The fixes are unglamorous:

• Use one exact company name everywhere — on the site, in the footer, in the page title, in press releases, in schema.
• Publish an about page that states legal name, registration details, founding date, address and contact. Machines resolve entities from exactly the kind of facts a legal page contains.
• Add Organization schema with sameAs links to profiles you control, so the model can connect the name to a consistent identity.
• Keep a physical address and a working telephone number on the contact page. This sounds like a formality; it is one of the strongest trust signals available to a small site.

What to do about freshness

Assistants weight recent material, and retrieval systems prefer pages that are updated rather than duplicated. In practice this means updating an existing page with a new date and new figures beats publishing a near-identical second page — the second page competes with the first and neither accumulates authority.

For a company publishing regularly, the compounding asset is not the number of pages. It is the number of pages that state a specific, attributed fact which no one else has published.

Measuring AI search visibility when there are no rankings

There is no equivalent of a rank tracking tool, and the workaround is direct: write down the twenty questions your buyers actually ask, run them monthly across ChatGPT, Perplexity, Copilot and Google's AI results, and record whether you appear and which sources do.

The list of sources is the useful output. It tells you which domains the model trusts on your topic — and presence on those domains is usually a faster route to being cited than optimising your own pages further.