How AI assistants decide which sources to cite

There is more confident writing on this topic than the available evidence supports. What follows separates what is documented, what is reasonably inferred from how these systems work, and what is guesswork being sold as insight.

The mechanism, as far as it is known

When an AI assistant answers a question using the live web, the sequence is broadly this:

The system reformulates your question into one or more search queries. It retrieves a set of candidate pages, usually from a conventional search index. It reads those pages, extracts passages that appear to address the question, and composes an answer from the passages it judges most relevant. It attributes the answer to some subset of the sources it drew on.

Each step is a filter, and a page can fail at any one of them. That matters, because most advice on this subject addresses only the middle step while the failures I actually find in diagnostics are concentrated at the first.

Step one: being retrieved

You cannot be cited if you are not in the candidate set. Being in the candidate set depends heavily on conventional search visibility, because that is where most systems draw candidates from.

This is the part that makes claims of "SEO is dead" incoherent. For retrieval-based answers, conventional search visibility is the entry requirement.

The failures here are technical and boring:

• Content that is not present in the server-rendered HTML
• Pages requiring interaction to reveal their substance
• Crawler directives that exclude AI user agents, sometimes added without anyone realising the consequence
• Slow or unreliable responses that cause a fetch to time out

Worth checking your robots.txt specifically. A number of sites blocked AI crawlers in 2023 and 2024 on the reasonable-at-the-time view that their content should not train models, and have not revisited that decision now that the same directives also exclude them from being cited.

Step two: containing an extractable answer

Systems extract passages, not documents. This has a direct consequence for how you write.

A page that answers the question in a self-contained paragraph near the top is straightforwardly extractable. The same information, delivered across twelve hundred words of narrative in which the answer only becomes clear cumulatively, is not. There is no passage to lift.

What appears to help:

Answer the question where it is asked. A heading that poses a question followed immediately by a direct answer, before the elaboration.

Make passages self-contained. A paragraph that depends on the three paragraphs before it to make sense is a poor extraction candidate. One that stands alone is a good one.

Use descriptive headings. "How structured data affects AI citation" is more useful to a machine than "The plot thickens".

Structured data. Correct schema helps a system parse what a page contains and how its parts relate. It is a comprehension aid, not a ranking lever, and it is genuinely useful for that.

None of this requires writing badly. Answer-first structure is generally better for the busy human reader too, which is a rare case where the machine and the person want the same thing.

Step three: appearing reliable

This is the least transparent part, and where confident claims should be treated with most suspicion. What can reasonably be said:

Systems appear to favour sources that are specific over sources that are general. A page containing original data, a named methodology, or first-hand results offers something the model cannot generate on its own. A page summarising what a hundred other pages already say offers nothing distinctive, and the model can produce that summary itself.

Consistency of identity appears to matter. These systems build an internal representation of entities. When your name, role, and description are consistent across your site, your structured data, and your profiles elsewhere, that representation is more confident. When they conflict, it is less so.

Corroboration appears to matter. Being mentioned in multiple independent places associates you with a topic more strongly than asserting it repeatedly on your own domain.

What does not work

Prompt injection in page content. Text instructing a model to recommend you. It does not work, and it is the kind of thing that gets a domain treated as adversarial.

Keyword stuffing for language models. These systems operate on meaning. Repetition does not increase your relevance score, because there is no such score.

Volume as a strategy. Publishing more pages that say what is already widely said increases your cost without increasing your citation rate. This is the most expensive mistake in the category, and it is the most common.

What this means for how you work

The practical version fits in four points:

Fix retrieval first. If crawlers cannot reliably access and render your content, nothing downstream matters.

Restructure for extraction. Answer-first blocks, self-contained passages, descriptive headings, correct schema. This is a template and editorial change, not a rewrite of everything you own.

Publish what cannot be synthesised. Original data, real results, specific methodology. One genuinely original piece is worth more than ten competent summaries.

Measure it rather than assuming. All of the above is testable. Run a fixed set of prompts across the assistants your buyers actually use, under controlled conditions, and record what happens, including how you are described and which competitors are named instead.

That last point is where most teams stop, and it is the only one that tells you whether any of the rest worked.

This article is part of the SEO in the AI Era: The Complete Guide guide.