LLM SEO: How Language Models Choose Which Sources to Cite
No AI provider publishes how sources are selected, so anyone describing a ranking algorithm is speculating. What is observable: models favour content they can retrieve easily, parse unambiguously, corroborate across independent sources, and quote as a self-contained passage. Those four are what you can actually act on.
This article is deliberately cautious. There is a great deal of confident writing about “how to rank in ChatGPT” built on inference from small samples, and much of it will age badly. What follows separates what is documented, what is reasonably inferable, and what is guesswork.
How retrieval actually works, roughly
Most AI answers involving current information are produced by retrieval rather than from memorised training data. The system runs a search, pulls back a set of candidate documents, and generates an answer grounded in them. That matters because it means two separable things decide whether you appear: whether you are retrieved, and whether you are used once retrieved.
- Retrieval depends on being indexed by whatever search layer the engine uses, being crawlable, and being relevant to the query as the system rewrites it.
- Use depends on whether your passage answers the question cleanly, and whether what you say agrees with other retrieved sources.
The second point is why contradicting yourself across your own web presence is so damaging. A model reconciling several sources will favour the claim it can corroborate.
What is documented, inferable, and guesswork
| Claim | Status |
|---|---|
| Engines run searches to answer current questions | Documented — providers describe browsing and retrieval |
| Crawler access is controlled via robots.txt user-agents | Documented — the user-agent names are published |
| Clear, self-contained passages are easier to quote | Reasonable inference — consistent with how extraction works |
| Corroboration across sources increases confidence | Reasonable inference |
| “Mentioning your brand N times increases citation” | Guesswork |
| “Posting on Reddit guarantees being cited” | Guesswork — forums are retrieved sometimes, which is not the same thing |
What to do about the observable parts
- Be retrievable. Allow the AI crawlers explicitly. Make sure content is in the HTML rather than assembled by JavaScript. Keep the site indexed by conventional search, since that is often the retrieval layer.
- Be unambiguous. One set of facts about your business everywhere. Structured data that matches the page.
- Be quotable. Self-contained answers, specific numbers, comparison tables. A passage that needs three paragraphs of context is hard to lift.
- Be corroborated. Directories, review sites, third-party lists. A claim that appears only on your own site is the weakest kind.
Frequently asked questions
Can you submit a site to ChatGPT?
There is no submission process equivalent to Search Console. What you can do is ensure the crawlers are permitted, the site is indexed by conventional search, and the content is clear enough to be used when retrieved.
Does being in training data matter?
Less than people assume for current questions, because those are answered by retrieval rather than memory. It matters more for general background claims — but you cannot influence training data directly, and a strategy built on it is not actionable.
Why do answers change between sessions?
Generation is probabilistic, retrieval results vary, and models update. This is why a single screenshot proves very little and why measurement needs a fixed prompt set run repeatedly.
Is Bing Webmaster Tools worth setting up?
Yes, and it is under-used. Bing’s index feeds several AI products, so being indexed there has a more direct line to some answer engines than Google alone. It takes minutes and most UK businesses have never touched it.
We work on this as AI SEO and generative engine optimisation, measured against a fixed prompt set rather than anecdotes.
