LLM SEO: how AI answers choose their sources

LLM SEO means making your pages easy for large language models to find, understand and cite when they answer a question. When an AI assistant cites a source, it is almost always because it searched the web while writing the answer: it ran searches against an index, fetched pages, picked passages and wrote from them. That retrieval step is where a site owner has influence. What the model absorbed in training, months earlier, you can't change after the fact.

Two ways a model "knows" about you

A language model can mention your business for two different reasons:

Training data. The model read text about you before it was released. This is fixed until the next model, it can be out of date, and it rarely comes with a link.
Retrieval at answer time. The assistant searches, reads current pages and writes an answer based on them, usually with citations. Google calls this "grounding": giving the model content from the search index when the question is asked (Google).

For a business, the second one matters. It's current, it links to you, and it happens on the questions people ask before they buy: "best physio near Kitsilano", "how much does a will cost in Ontario", "tutoring for 11+ in Richmond". Everything below is about retrieval.

How AI answers pick sources

The steps differ by product, and none of them publish their ranking. But what Google, Microsoft and OpenAI do publish, plus practitioner data, gives a consistent picture.

1. The question becomes several searches

Google says AI Overviews and AI Mode may use "query fan-out", "issuing multiple related searches across subtopics and data sources" (Google Search Central). ChatGPT does the same. Grow and Convert ran 100 buying-intent prompts through ChatGPT; 83 triggered a web search, and ChatGPT typically split each prompt into two to four narrower searches (Grow and Convert, 2026).

So your page isn't competing only for the words the person typed. It competes for the follow-up questions: price, location, duration, "for beginners", "compared with".

2. The searches run against an index

The assistant needs a search index to search. Google's AI features use Google's index. Copilot uses Bing's. ChatGPT search uses OpenAI's own crawler and outside search services; an OpenAI engineering lead has said "Bing is an important one" (Search Engine Roundtable). Perplexity runs its own crawler.

This is why being indexed in both Google and Bing matters more than it used to. A site that only checks Google can be missing from half the assistants.

3. Pages are read in pieces

Microsoft's guidance for Bing and Copilot says AI assistants "don't read a page top to bottom like a person would"; they break pages into smaller, usable pieces and rank those (Microsoft, October 2025). Headings mark where one idea ends and the next begins. Content hidden in tabs or expandable menus may not be read at all.

4. Sources are chosen across all the searches

Because the answer draws on several searches, the cited pages often aren't the top result for the original prompt. Ahrefs compared citations with search rankings over 15,000 prompts: on average about 12% of links cited by ChatGPT, Gemini, Copilot and Perplexity were in Google's top 10 for the same prompt, and about 8% for ChatGPT (Ahrefs, August 2025). Grow and Convert matched against the fan-out searches instead and found more overlap: 27% of ChatGPT's cited pages ranked on Google for those searches, and about half when counting any page from the same domain.

Google's AI Overviews are the exception: the same Ahrefs article cites separate research in which 76% of AI Overview citations came from top-10 pages. Rankings still count; they just aren't the whole story.

5. The model writes, quoting what's easy to use

The 2023 research paper that named generative engine optimisation tested which page changes made a source more visible in AI answers. Adding quotations, statistics and cited sources helped most; keyword stuffing made things slightly worse; and lower-ranked pages gained more than the top result (Aggarwal et al.). Those tests mostly ran on a simulated engine, so read them as direction, not a formula. The direction matches Microsoft's advice: direct questions with clear answers, lists and tables, schema, examples and data.

What you control and what you don't

You control	You don't control
Whether crawlers can reach the site (`robots.txt`, firewall, CDN settings)	Which searches the assistant runs
Whether pages are indexed in Google and Bing	How each product ranks sources
How each page is structured and how direct its answers are	Whether an answer shows citations at all
Facts, prices and specifics on the page	What older models learned in training
Structured data and page speed	Personalisation (Ahrefs notes the same prompt can cite different pages for different people)
How consistently your business is described elsewhere	What other sites say about you, beyond asking them to correct it
How fast engines hear about changes (sitemaps, IndexNow)	When they recrawl

The llms.txt question

llms.txt is a proposal by Jeremy Howard, first published in September 2024: a Markdown file at /llms.txt that gives AI tools a short summary of a site and a list of links to its most useful pages, ideally as plain Markdown. The proposal's site says thousands of sites publish one, the AI labs publish them for their developer docs, and Chrome's Lighthouse checks for one in its agentic browsing audit.

What isn't established is that any AI search product uses it to choose or cite sources. In June 2025 Google's John Mueller wrote "no AI system currently uses llms.txt" (Search Engine Roundtable), and Google's guidance for AI Overviews and AI Mode says you don't need new AI-specific files to appear in them. No major AI search provider has published documentation saying it reads the file for ranking or citation. Treat it as cheap and harmless: useful to coding agents and tools that look for it, not a lever for AI visibility.

A practical LLM SEO checklist

Access

robots.txt allows Googlebot, Bingbot and the AI search crawlers you want to appear in (for OpenAI that's OAI-SearchBot; see ChatGPT SEO).
Your host, CDN or firewall doesn't block those crawlers. OpenAI publishes IP lists for its crawlers so they can be allowed (OpenAI).
Decide separately whether you allow AI training (GPTBot, Google-Extended). Google says blocking Google-Extended doesn't affect Search inclusion or ranking.

Indexing

Site verified in Google Search Console and Bing Webmaster Tools, sitemap submitted to both.
Changes reach the engines quickly. Bing recommends IndexNow so AI systems reference the current version of a page (Bing Webmaster Blog).

Pages

One page per service, location and common question.
Each section opens with the answer and names its subject, so it makes sense out of context.
Descriptive headings, not "Overview" or "More information".
Nothing important hidden in tabs, accordions or images of text.
Prices, hours, durations, qualifications and numbers written out, with sources for outside claims.
A short FAQ with the follow-up questions customers actually ask.

Structured data

Business details, each location's address and opening hours, services, people, articles and FAQs, all matching the visible text.

Reputation

Same name, address and phone on your site, Google Business Profile, Bing Places and directories. Microsoft specifically suggests Bing Places for local details.
Reviews and mentions on sites the assistants trust in your field.

Measurement

Ask the assistants your customers' questions every month or so and note who's cited.
Bing Webmaster Tools' AI Performance report for citations in Copilot and Bing, including the "grounding queries" it used to find your pages.
Search Console's Performance report (Google counts AI Overviews and AI Mode under "Web").
Referral visits from chatgpt.com, perplexity.ai and copilot in your analytics.

Where Firetend fits

Several items on that list are built into every Firetend site: fast static pages, structured data generated from your content, a sitemap, a robots.txt that allows search engines and AI answers (training stays off unless you turn it on), an /llms.txt file, a plain Markdown copy of each page for AI tools, and IndexNow on every publish. When the content agent writes or refreshes a page, it checks the page for search and AI-readiness and links to the sources it used. See How your site shows up in search and AI answers for the full list.

The parts no platform can do for you are the facts and the reputation. If your current site fails the access and pages sections, an AI website redesign is the quicker fix. For the difference between this and classic SEO, read GEO vs SEO.

FAQ

Is LLM SEO different from GEO?

No, they're two names for the same work: being found and cited by AI assistants. "GEO" comes from a 2023 research paper; "LLM SEO" is the plainer description. GEO vs SEO compares the practice with classic SEO.

Can I get my business into an AI model's training data?

Not in any way you can plan around. Training happens on the model maker's schedule and you can't see what was included. Retrieval is the part you can influence, and it's the part that comes with links.

Does ranking on Google still matter for AI answers?

Yes. Google's AI Overviews mostly cite pages that already rank in the top 10. Other assistants cite more widely, but they still find pages through search indexes, so being indexed and ranking for the related questions raises your chances.

Should I add an llms.txt file?

It does no harm and takes minutes, but don't expect it to change what AI assistants cite. It's a proposal, and no major AI search provider has said it uses the file to choose sources. Every Firetend site has one by default.

How long does it take for changes to show up in AI answers?

It depends on when the engines recrawl. Google says recrawling can take days to months. OpenAI says robots.txt changes for its search crawler take about 24 hours to apply. IndexNow tells Bing and other participating engines about changes when you publish.