How do AI search engines decide which websites to cite in their answers?
How AI search engines choose which websites to cite: retrieval, fan-out queries, ranking and grounding, what is known and unknown, and how to measure it for your site.
No AI search engine publishes its full rulebook for choosing sources. What we have is a mix of official documentation, observed behavior and a lot of speculation. This article separates the parts that are well established from the parts that are guesses, then shows how to measure what actually happens for your own site.
The general pipeline
Most AI search products that cite sources work in a similar shape, though details differ by engine and change often.
- Interpret the question. The model works out what is being asked, often rewriting it.
- Fan out. Instead of one search, the system issues several sub-queries. Ask for the best project management tool for a ten-person agency and it may look separately for pricing, reviews, integrations and comparisons. These are called query fan-outs.
- Retrieve. Those queries hit a search index or live fetch. Pages must be crawlable, indexable and, for live fetches, reachable by the engine's bots.
- Rank and select. Candidate pages are filtered and scored. Exact signals are not public, but relevance to the sub-query, the quality and clarity of the passage, and source credibility are widely believed to matter.
- Ground and write. The model composes an answer from the selected passages and attaches citations to some of them.
- Present. The interface decides how many citations to show and how prominent they are.
A citation is therefore the end of a chain, and a page can fail at any link: blocked from the crawler, not indexed, not retrieved for any sub-query, retrieved but not selected, or selected but not shown.
What is documented, and what is not
Google's guidance on AI features in Search says the same foundations apply as for classic search: pages must be indexable and eligible to appear with a snippet, and helpful content is what to aim for. OpenAI documents separate crawlers for different purposes, including OAI-SearchBot for search results and GPTBot for training, and Anthropic and Perplexity document their own. Bing and Microsoft Copilot have their own webmaster guidance.
What is not documented is any exact weighting of signals. Claims that a particular schema type, word count or link count "makes AI cite you" are usually extrapolation from small samples. Treat them with the skepticism you would give any unverified ranking-factor claim.
What tends to matter in practice
These are observations and working hypotheses, not guarantees.
- Access: if bots cannot fetch the page, it cannot be used by a live-fetch engine.
- Direct answers: passages that state the answer plainly are easier to quote.
- Specificity: named entities, numbers and dates give a model something concrete to cite.
- Topical coverage: engines issue several fan-out queries, and a page that answers only one of them has fewer chances.
- Third-party presence: answers often cite sources other than the brand's own site, so mentions on forums, video and trade sites can decide the answer.
- Freshness: for topics that change, recently updated pages may be favored. How strongly varies.
How to measure it for your site
The useful move is to stop theorizing and look at your own data.
Fan-outs. Promptwatch lists query fan-outs per model, so you can see which sub-queries each engine ran. Its free GEO Inspector Chrome extension shows ChatGPT fan-outs and the difference between fetched and cited sources, and runs locally.
Citations. Citation analytics shows page- and domain-level citations, URL positions, a Page Tracker for specific URLs, offsite mentions, and Reddit, YouTube and LinkedIn reports. Citation Trends classifies each citation by content type and source type, with first cited, ramp-up days, peak rate and decay half-life per page.
Crawling. Agent Analytics shows IP-verified bot hits and a crawl-to-citation path with a citation rate per page, so you can see how many fetched pages end up cited. Crawler logs begin on the Professional plan ($245 a month).
A published example. Promptwatch's case study with TreeHouse, which runs Pararius and Huurwoningen.nl, reports that AI search became about 25% of organic conversions, 99% of it from ChatGPT, after fixing crawler access and automating weekly reporting. Attribute that to TreeHouse, not as a general expectation.
Other tools and what they show
| Tool | Starting price | What it shows about citation |
|---|---|---|
| Promptwatch | Free; $95/mo | Fan-outs, page-level citations, crawl-to-citation path |
| AthenaHQ | $295/mo | Visibility across 11 models, Action Center |
| Scrunch AI | $250/mo | Visibility on four LLMs, bot traffic reporting |
| Ahrefs Brand Radar | Included with paid Ahrefs plans | Mentions, Reddit and YouTube sources |
| Profound | Trial; custom | Citation share, answer-level analytics |
What to do with this
Check bot access first. Then list the sub-queries engines run for your top prompts and see whether you have a page for each. Look at who is cited instead of you and ask why. Fix what you control and accept that some of the choice is outside it. Nobody can guarantee a citation, and anyone who claims to know the exact formula is guessing.