LLM SEO: how large language models choose what to cite.

A large language model learns patterns from a large body of text during training, which shapes how it writes and what it knows in general. The links in an AI answer come from a different step, retrieval: the tool searches, reads a set of pages and cites the ones it used. The three platforms people ask about most document the two routes separately:

Two things follow. First, blocking a training crawler does not by itself keep you out of AI search. OpenAI treats GPTBot and OAI-SearchBot separately, and says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Second, robots.txt is not a complete off switch: Perplexity says Perplexity-User, which fetches pages a person asked about, generally ignores robots.txt rules, and OpenAI says robots.txt rules may not apply to ChatGPT-User. Because retrieval runs on search, the pages a tool can cite are the ones it can find and read now.

None of these platforms publishes the full criteria it uses to choose citations, and anyone outside these companies who claims to know the formula is guessing. What they do document is where the process starts. Google says its generative features rely on its core Search ranking systems to retrieve pages from the Search index, and may use query fan-out, running several related searches to gather more results. To be eligible as a supporting link, a page must be indexed and eligible to show with a snippet in Google Search. OpenAI says any public website can appear in ChatGPT search, and that content can be included in its summaries and snippets as long as the site is not blocking OAI-SearchBot.

Research adds direction, with caveats. In the generative engine optimization (GEO) paper presented at KDD 2024, adding citations, quotations from relevant sources and statistics made pages more visible in generated answers, by over 40% in the authors' tests on a benchmark of 10,000 queries. That is a research result, not a platform rule, so treat it as an experiment to run on your own pages.

A tool that answers by searching reads what others say about you, not only what you say about yourself. In Ahrefs' study of 75,000 brands across ChatGPT, Google's AI Mode and AI Overviews (Louise Linehan, December 12, 2025), mentions on YouTube had the strongest correlation with how often AI answers mentioned a brand, about 0.737, and branded web mentions followed at 0.66 to 0.71. Link metrics, including the number of backlinks, showed very weak correlations. The authors add the usual caution: correlation isn't causation.

Read it as direction, not formula. Coverage, reviews, videos, comparisons and discussion in places your customers trust are part of how an AI tool meets your business, and much of that is the work of good digital PR.

Semrush uses the term LLM seeding for publishing and distributing content so that the AI systems behind tools like ChatGPT, Perplexity and Google's AI Overviews can find, understand and reference your brand. In practice that means clear information on your own site, plus earned presence elsewhere: reviews, partner pages, videos and community discussion.

The line to hold is that seeding has to be earned and true. In the US, the FTC's 2024 rule on fake reviews and testimonials bars businesses from creating or selling fake reviews, and from buying them when they knew or should have known they were fake. It also covers certain reviews by company insiders that do not clearly disclose their connection to the business, and it lets the FTC seek civil penalties against knowing violators.

• Write a fixed list of questions your customers actually ask, including ones that name your category but not your business.
• Ask each tool the same questions on a set schedule, such as monthly, and record the answer, the pages cited and the conditions: date, tool, signed in or not, location.
• Check the facts it gives about you, not just whether you appear. A wrong description is a problem to fix at the source.
• Read the traffic you can see. Google says traffic from its AI features is counted in Search Console's Performance report, under the Web search type. OpenAI says ChatGPT adds utm_source=chatgpt.com to its referral links, so those visits can be separated in your analytics.

An LLM audit can do two useful things: check the crawler access, indexing and pages that decide whether AI tools can use your site, and record what the tools say about you today. It cannot reveal a platform's hidden ranking rules, so be wary of any audit or agency that promises placements. Our AI search optimization work starts the same way: crawler access first, then a baseline of where the business appears, then the source material and the evidence around it.