How AI decides what to cite

Ask an assistant "what's the best tool for transcribing interviews?" and you'll get an answer with two or three named products and, usually, a short list of sources. Those named sources won the query. Everyone else — no matter how good their page is — was not in the room. Understanding how that shortlist gets built is the whole game, so here is the pipeline as it actually runs, without the marketing gloss.

The answer is not a search result

A classic search engine returns a ranked list of ten links and lets you choose. An answer engine does something different: it writes a single response and decides, mid-composition, which sources to lean on and name. The user reads the answer, not the list. If your domain isn't among the handful the model grounds its answer in, you don't get a lower ranking — you get zero visibility, with no signal that it happened.

That changes the job. You are no longer competing for position five on a page. You are competing to be one of the few sources a model considers trustworthy and quotable enough to build an answer on.

The pipeline, step by step

Whether it's ChatGPT with browsing, Perplexity, Google's AI Overviews, or Claude with search, the shape is broadly the same four stages:

1. Retrieval — get a candidate set

The engine turns the user's question into one or more search queries and pulls a set of candidate pages, usually from a conventional search index (Bing, Google, or its own crawl). If your page is not in that underlying index, or is too slow or blocked to fetch, the pipeline never sees it. Retrieval is a hard gate: nothing downstream can rescue a page that wasn't retrieved.

2. Ranking — pick the few worth reading

From dozens of candidates, the engine keeps a small number — often three to six — to actually read. This is where source reputation, topical relevance, freshness, and how directly the page answers the query do their work. A page that clearly and specifically answers the exact question beats a broad page that mentions the topic in passing.

3. Grounding — write the answer from those pages

The model reads the shortlisted pages and composes an answer grounded in them. Here, the writing on your page matters enormously. Self-contained, factual sentences that stand on their own are easy to lift. Claims buried in a sales narrative, split across a slideshow, or dependent on a video are not. The model quotes what it can cleanly extract.

4. Attribution — decide whom to name

Finally the engine attaches citations to the answer. It tends to credit the sources that most directly supported specific claims — and it tends to prefer sources that agree with each other. A fact corroborated across several reputable pages is "safe" to state and attribute; a lone, uncorroborated claim is riskier for the model to repeat.

What actually makes a page citable

Translating those four gates into things you can change:

• It's indexed and fast to fetch. Verify the page is actually in the search index — not just submitted — and that it returns quickly without login walls, aggressive bot-blocking, or content that only appears after heavy JavaScript.
• It answers one real question directly. Pages that state the answer in the first paragraph, then support it, are far easier to ground on than pages that build up to a conclusion or hide it behind a form.
• Its claims are self-contained. Write sentences a model can lift verbatim and have them still be true out of context. Concrete numbers, dates, and named specifics are more quotable than adjectives.
• It's structured cleanly. Real headings, lists, and tables — plus a clean machine-readable rendering — reduce parsing ambiguity. This is exactly why InfoRelay serves structured markdown to agents.
• It's dated and maintained. Freshness is a ranking signal for many queries. A visible, honest "last updated" beats a stale page competing on the same question.
• It agrees with the record. If reputable sources describe your product one way and your page claims something contradictory, the model will usually trust the corroborated version — sometimes describing you incorrectly as a result.

What does not reliably work

The GEO space is already full of tactics that sound plausible and don't survive contact with the pipeline above:

• Keyword stuffing. Answer engines rank on relevance and clarity, not density. Stuffing makes a page less quotable, not more.
• Thin, AI-spun content at volume. Publishing a hundred shallow pages gives the model a hundred weak, uncorroborated candidates. It does not manufacture trust.
• Schema markup alone. Structured data helps a machine parse a page; it does not make the underlying claims true, relevant, or corroborated. It's a multiplier on good content, not a substitute.
• Buying "mentions." Paying to appear on low-trust pages adds noise the model discounts, and can pull your described facts in the wrong direction.

How to find out where you stand

You cannot optimize what you haven't measured, and the number that matters is not traffic — it's your Share of Citation: across the buyer questions in your category, how often do the engines name your domain as a source? We wrote a full field report on measuring exactly that, across the sites we own, in We measured our own AI citation share for 90 days. The method there is the same one we'd point any site at — baseline the real questions, score who gets cited today, then close the specific gaps.

None of this is a trick. The engines are trying to give a good answer from trustworthy, quotable sources. "Getting cited by AI" is mostly the discipline of being that source, on the exact questions your buyers ask, and then proving it moved with a repeatable measurement.