How AI Assistants Choose What to Cite

Citation is a pipeline, not a judgement

Teams tend to treat citation as a single verdict: the model liked us, or it did not. That framing makes the problem unfixable, because there is nothing specific to act on.

What actually happens is a sequence of filters. Understanding the sequence turns one vague problem into four concrete ones, each with its own fix.

Stage 1: Query fan-out

Your buyer's question is not the query that hits the retrieval index. The model rewrites it, usually into several queries at once.

Someone asking "what should I use to track whether ChatGPT mentions my company" might generate retrieval queries for AI brand monitoring tools, ChatGPT visibility tracking, answer engine optimisation platforms, and share of voice measurement in LLMs.

Two consequences follow.

First, you are competing for queries nobody typed. Optimising for the literal phrasing of a buyer question is less useful than covering the concept space around it, because the concept space is what gets searched.

Second, the vocabulary the model uses in its rewrites matters. If your category has an insider term and a plain-language term, and your pages only use the insider term, you will miss the rewrites generated from plain-language questions. Covering both registers is cheap and effective.

You are failing here if: your pages rank for your preferred terminology but the model's rewritten queries use different words, so you never enter the candidate pool at all.

Stage 2: Retrieval

The rewritten queries hit an index, and a set of candidate documents comes back. This is the stage that resembles conventional search most closely, and where classic SEO signals do most of their work.

It is also where the most common and most boring failure lives: you are not in the index in a usable form.

• robots.txt disallows the engine's crawler.
• The page requires JavaScript the crawler does not execute, so the fetched HTML is an empty shell.
• Bot mitigation serves a challenge page instead of content.
• The content is inside a PDF, an image, or a component that renders only after interaction.

None of this is exotic. In our reading of published audits and vendor case studies, retrieval failures of this kind are the single largest category of AEO problem, and the cheapest to fix.

You are failing here if: fetching your URL with an AI user agent returns something other than your content. Test this directly rather than assuming.

Stage 3: Re-ranking

Now the interesting part. The engine has more candidates than it can use, and it scores passages, not pages.

This is the detail that changes how you write. A 4,000-word guide is not evaluated as one unit. It is chunked, and individual chunks compete. A brilliant document with a diffuse structure can lose to a mediocre document containing one crisp, self-contained paragraph that directly answers the query.

What re-rankers reward, as far as can be inferred from observed behaviour:

• Direct answers. The claim stated plainly, near a heading that matches the question.
• Self-containment. A passage that makes sense lifted out of context. Chunks beginning "As mentioned above, this approach..." are structurally handicapped, because the antecedent is gone.
• Specificity. Concrete numbers, named entities and dates outrank hedged generalities. "Pricing starts at $99 per month" beats "affordable pricing options are available".
• Verifiability. Claims attached to a source, a date, or a method are preferred over bare assertions.
• Recency, conditionally. For questions with a temporal dimension — pricing, comparisons, anything with "2026" in it — fresh content is strongly favoured. For stable conceptual questions, recency matters much less.

You are failing here if: you are retrieved but not quoted. Your content is in the candidate pool and losing the passage-level competition. This is a structural and editorial problem, not a technical one.

Stage 4: Composition and the citation threshold

The model writes an answer from the surviving passages, then decides what to attribute. Crucially, using a passage and citing it are separate decisions, and engines differ enormously in how freely they attribute.

This produces a frustrating and real phenomenon: your content shaping an answer without your name appearing. A model can absorb your framing, adopt your numbers, and cite nobody, or cite an aggregator that copied you.

It also means citation counts understate influence. If you are measuring only explicit citations you are measuring the visible portion of your contribution.

At composition, models also apply consistency checks. A claim contradicted by other retrieved sources tends to get dropped, hedged, or attributed to the consensus rather than to you. Being the lone dissenting source is a weak position even when you are correct — which is why getting independent corroboration published elsewhere is legitimate AEO work.

You are failing here if: your framing appears in answers but your brand does not. Usually a corroboration and authority problem rather than a content problem.

Diagnosing your own stage

A rough sequence that works without specialist tooling:

1. Fetch your key page as an AI crawler. Set the user agent to GPTBot and request the URL. If you do not get your content, stop here and fix it. Nothing downstream matters.
2. Search the model's likely rewrites in a conventional engine. If you are absent from those results, your problem is Stage 2 and it is ordinary SEO work.
3. Ask the assistant the buyer question directly, several times. If competitors appear consistently and you never do while ranking well conventionally, your problem is Stage 3: structure and extractability.
4. Read the answers for your own framing. If your ideas are present without your name, your problem is Stage 4: corroboration and authority.

Doing this by hand is fine for one page and one question. It does not scale to a category, which is where instrumentation earns its cost. Tools that attribute a citation loss to a named factor — CiteCue publishes a thirteen-factor model, for instance — are doing this diagnosis systematically rather than anecdotally, which is the difference between a programme and a hunch.

What this means for how you write

Three practical rules follow from the pipeline:

Write for the chunk. Every section should survive being read alone. Answer first, elaborate second, and avoid opening a section with a reference to something above it.

Be specific enough to be checkable. Numbers, dates and named things are what re-rankers reward and what composition steps can verify. Vagueness is penalised twice.

Consolidate rather than proliferate. One page that owns a question decisively beats four that each cover it partially. The re-ranker is choosing between passages, and splitting your best material across several URLs guarantees that none of them is your strongest possible candidate.

Further reading

Structuring Content So AI Can Quote It covers Stage 3 in depth. Technical AEO for AI Crawlers covers Stage 2. Why Your Brand Is Not Being Cited works through the diagnostic in a more structured form.