LLM Search Engine Optimization: How to Get Your Pages Cited

An LLM search engine does three things.

It searches an index. It reads the pages that come back. It writes one answer, with sources.

ChatGPT search does it. Perplexity does it. Google AI Overviews and Copilot do it.

So "LLM search engine optimization" means: get your pages retrieved and cited. Get your brand named.

One rule comes before all others. A page the crawler can't fetch can't be cited.

Is this just SEO? Mostly. Google's guide to generative AI features says that, from its side, optimizing for generative AI search is still SEO.

What changes:

• Several indexes and crawlers. Not one.
• Rewritten queries.
• A result you count as a citation or a mention. Not a rank.

Where an engine hasn't published how something works, I'll say so.

How an LLM Search Engine Finds and Cites a Page

A language model alone is not a search engine.

It writes from its training data. It has no index.

An LLM search engine adds retrieval. Two techniques do the work.

1. Retrieval-augmented generation. Also called grounding. Google's guide names it: Google's core Search ranking systems fetch relevant, current pages from the Search index. The model writes the response from those pages.

2. Query fan-out. The model generates related queries and runs them at the same time.

Google's example. "How to fix a lawn that's full of weeds" may fan out to:

• "best herbicides for lawns"
• "remove weeds without chemicals"
• "how to prevent weeds in lawn"

ChatGPT does it too. OpenAI's help page on web search says ChatGPT typically rewrites your question into one or more targeted queries. And can send more specific ones after it reads the first results.

(For the other assistants, the evidence is Ahrefs' analysis of query fan-out. Not a vendor page.)

So your page competes on the searches inside the question. Not on your sentence.

The prize: a place among the few sources under one answer.

The sequence is broadly the same in each engine:

1. Decide whether to search. OpenAI says ChatGPT may search when a question needs current information.
2. Rewrite the question into several short queries.
3. Look up pages in an index that a crawler filled earlier.
4. Select some of the retrieved pages.
5. Write the answer with links to the pages it used.

A site wins or loses at steps 3 and 4.

Step 3 is documented for each engine. Step 4 is not.

What Each Engine Needs From Your Site

Sources: OpenAI's help page above, its crawler documentation, Perplexity crawler documentation, Google's AI features documentation, the Bing Webmaster blog, and Microsoft's guidance for AI search answers.

Four things that are easy to get wrong:

• GPTBot is not the search crawler. OpenAI lists GPTBot for model training and OAI-SearchBot for search. Each setting is independent. Block GPTBot and you stay in ChatGPT search. Block OAI-SearchBot and OpenAI says the site isn't shown in the answers. (It can still appear as a navigational link.)
• A correct robots.txt is not enough. OpenAI also asks your host or CDN to allow its published search crawler IP addresses. A firewall rule can remove you with no change to robots.txt.
• Google-Extended does not control AI Overviews. To limit what Search shows, Google points to nosnippet, data-nosnippet, max-snippet and noindex.
• robots.txt does not control a fetch that a user triggers. OpenAI says its rules may not apply to ChatGPT-User. Perplexity says Perplexity-User generally ignores them.

What Decides Which Pages Get Picked

Nobody outside knows for sure.

OpenAI says only that ChatGPT ranks results on multiple factors. And that placement is not guaranteed.

So the evidence on step 4 is data from the output.

One question and five engines

Our public leaderboard scans questions monthly on five engines. It lists the pages each one cited.

Here's "best SEO tool", dated 24 September 2026:

The engines cited 27 pages between them. 24 of those are different pages.

Three were cited by two engines. None by three.

Ahrefs found the same shape in much larger data. Take the 50 most-mentioned websites in each of AI Overviews, ChatGPT and Perplexity (June 2025). Only 7 were in all three lists.

The scan record shows one more thing.

• Copilot's recorded search was "best SEO tools 2025". For the question "best SEO tool". All five pages it cited have "2025" in the title.
• Perplexity's recorded search was "best SEO tool". 7 of its 10 cited titles say "2026".

The engine rewrote the question. Then cited pages whose titles matched the rewrite.

(One question, one date, one run per engine. An example. Not a rate.)

What larger studies show

Cited pages have titles that match the sub-question.

Ahrefs looked at 1.4 million ChatGPT prompts (April 2026).

• ChatGPT cited about half of the URLs it retrieved.
• Cited titles were closer to the prompt than non-cited ones. Cosine similarity 0.602 against 0.484.
• Against the best-matching fan-out query, they scored 0.656. (A best match. Not comparable with those two figures.)

Ahrefs used open-source embedding models as a stand-in for ChatGPT's own scoring.

A top-10 Google rank for the prompt is not the test.

Ahrefs studied 15,000 prompts (August 2025).

• On average, 12% of the URLs that AI assistants cited ranked in Google's top 10 for the same prompt.
• 28.6% for Perplexity.
• 6% to 9% each for ChatGPT, Gemini and Copilot.

Why so low? Ahrefs points to query fan-out.

The engine searches the index for its fan-out queries. Not for your prompt. So a page can rank nowhere for the head term and still be a good result for one sub-question.

$49 report

Does AI name you?

See if AI mentions your brand. Then get the pages to fix.

Your website

Get my $49 report

Create a free account first, then pay $49 once.

Check one question freeHow it works

What's the best AI SEO tracker for a startup?

Most answers name Semrush. You aren't cited.

Ask ChatGPT

How to Get Your Pages Cited

1. Let the crawlers in and test them

Check robots.txt for OAI-SearchBot, PerplexityBot, Googlebot and Bingbot.

Then check your firewall or CDN.

• OpenAI and Perplexity ask you to allow their published IP ranges. (OpenAI, Perplexity.)
• Cloudflare has separate Search, Agent and Training settings for AI bots.
• OpenAI says a robots.txt change can take about 24 hours to reach its search systems.

Two commands test what a crawler gets.

The first sends OAI-SearchBot's user agent and prints the final status code. (OpenAI says the version number may change.)

The second sends PerplexityBot's. It counts the raw HTML lines that contain a sentence you want cited.

curl -sL -o /dev/null -D - -A "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36; compatible; OAI-SearchBot/1.4; +https://openai.com/searchbot" https://example.com/your-page | grep HTTP | tail -1

curl -sL -A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot)" https://example.com/your-page | grep -c "a sentence from your answer"

Three limits:

• This tests user agent rules only.
• A firewall that filters by IP address needs the published IP lists and your server logs.
• A 200 doesn't mean the engine has indexed the page.

The sentence is missing from the raw HTML? Then a crawler that doesn't run JavaScript can't read it.

In December 2024, Vercel and MERJ found that the crawlers of OpenAI, Anthropic, Meta, ByteDance and Perplexity did not render JavaScript. That included OAI-SearchBot and PerplexityBot.

Gemini uses Googlebot's infrastructure. That one renders JavaScript.

The data is almost two years old. So test your pages. The free Page Inspector shows what AI can read on a page.

2. Cover the sub-questions with titles that say so

The engine retrieves for its fan-out queries.

Your page gets retrieved when it's a good result for one of them.

That favors a page with one clear subject. And a title that names it. The Copilot example above shows it.

Can you see the fan-out queries?

• ChatGPT: I know of no first-party report that shows them for your prompts.
• Microsoft: yes. The grounding queries in Bing Webmaster Tools are the phrases the AI used when it retrieved pages that were then cited. Microsoft says the list is a sample.

Don't write a separate page for every phrasing. Google's guide says that doing so primarily to manipulate rankings or AI responses violates its scaled content abuse policy.

3. Put the answer first with its evidence

Give the direct answer at the top of each section. Then the support.

Microsoft's guidance recommends clear headings, question-and-answer formats, lists and tables.

The GEO paper (KDD 2024) built a benchmark of 10,000 queries. It tested content changes on the 1,000-query test split.

Three changes gave a source 30 to 40% more visibility on its main metric:

• Added quotations
• Added statistics
• Cited sources

(The metric counts a source's words in the answer and weights them by position.)

Keyword stuffing scored below the unchanged page.

The limits: results vary by domain. And the engines were those of 2023 and 2024.

I read it as direction. Sourced, specific statements are easier for a model to use than general ones.

The statistics must be real.

4. Get onto pages the engines already cite

Ahrefs studied 75,000 brands (May 2025).

• Branded web mentions: the strongest correlation with brand visibility in AI Overviews. 0.664.
• Backlinks: 0.218.

(Correlation. Not cause. The author says so.)

The source lists point the same way.

• Pew: Wikipedia, YouTube and Reddit were 15% of the sources in AI summaries. (Pew data.)
• Our leaderboard: listicles are 77% of the cited pages across 343 questions (read 2 October 2026).

(Those are consumer and software "best X" questions. Our own classifier sets the page types. The numbers change monthly.)

Two companies report the same pattern. Self-reported, with no isolated cause.

• Vercel said in mid-2025 that ChatGPT referred about 10% of its new signups. Its practices include earning community mentions.
• Tally credits years of answering questions in communities and forums.

So: find the third-party pages that the engines cite for your prompts. Get onto them.

The outreach: getting into AI visibility listicles. The wider list: how to get cited in AI answers.

5. Keep the pages that matter current

Ahrefs analyzed 17 million citations (July 2025).

• Cited by ChatGPT, Perplexity, Copilot and Gemini: 1,064 days since publication, on average.
• Google's organic results: 1,432 days.
• Google's AI Overviews: the exception. Pages 16 days older than the organic results.

The Ahrefs author suspects that most brands will see better results from new quality content than from extremely frequent updates.

And Google's John Mueller has warned against changing dates without changing the page.

Update the substance of a page that matters. Not only its date.

What you can skip

Google's guide says:

• Search doesn't use llms.txt files.
• You don't need to break content into small chunks for AI.
• Structured data isn't required for generative AI search.

Microsoft's guidance does recommend schema markup.

My reading: keep the structured data your CMS already produces. Don't start a schema project to chase citations.

AI search optimization goes through Google's full list.

How to Tell If It Worked

Clicks alone undercount this work.

Pew Research Center tracked 900 US adults in March 2025.

• 8% of visits to a page with an AI summary included a click on a search result.
• 15% when it had none.
• 1% of visits to a page with an AI summary included a click on a link inside the summary.

Ahrefs looked at 300,000 keywords. An AI Overview correlated with a 34.5% lower average click-through rate for the top-ranking page in March 2025.

When it repeated the study on December 2025 data: 58%. The baselines differ. The first run compared March 2025 with March 2024. The second compared December 2025 with December 2023.

So a citation doesn't replace the click. It matters because the answer is what the person reads.

Track two outcomes separately:

• Citation: one of your URLs is in the source list.
• Mention: your brand is named in the text of the answer. With or without a link.

(The difference: AI citations. The metrics: how to measure AI brand visibility.)

Then use what each engine gives you:

• ChatGPT visits. OpenAI says ChatGPT adds utm_source=chatgpt.com to its referral URLs. So publishers who allow OAI-SearchBot can track visits in analytics. In the scan above, every URL ChatGPT cited ended in it. Gemini's carried utm_source=gemini. (An observation from the scan. Not a Google document.)
• Google Search Console. In June 2026 Google launched separate generative AI performance reports, with impressions by page and country. All sites had them by August 31, 2026.
• Bing Webmaster Tools. The AI Performance report shows citations, cited pages and a sample of grounding queries.
• OpenAI and Perplexity. I know of no equivalent report from either. For those engines, you run the prompts and read the answers.

Your Next Steps

1. Check access. Run the two curl tests from tactic 1 on your key pages. Confirm that Google and Bing index them.
2. Write a prompt set. 20 to 50 questions a buyer asks before they choose a product like yours. Full sentences. The free keyword research tool finds prompts from buyer questions. Then freeze the set.
3. Record a baseline. For each prompt and engine: are you mentioned? Cited? Which URL? Who else is cited? Run each prompt more than once. One answer is a sample, not a rate.
4. Fix the pages the engines already reach. Start where a competitor's page is cited for a question you also answer. Move the answer to the top. Source each number. Make the title name the sub-question.
5. Work the source list. The third-party pages in your baseline are the places to appear.
6. Scan the same set again. Compare mentions and citations for each engine. Next to clicks.

Steps 2 and 3 are what the AI SEO Tracker report does.

$49, once. You enter your website. It writes up to 50 prompts from your site for you to confirm. It scans ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews. Daily for 7 days.

You get mentions, citations, position in the answer, competitor gaps, the sources each engine cited and a prioritized action plan. The daily scans give you repeat runs for step 6. Weekly tracking is optional, from $79/mo.

A spreadsheet does the same job on a small set. If you have the time.

See also: how Perplexity picks sources, how to rank in ChatGPT, and LLM SEO for the brand side.

FAQ

What is LLM search engine optimization?

The work of getting your pages retrieved and cited, and your brand named, in LLM search answers. (Optimisation, in UK spelling.)

Three parts:

1. Crawler access.
2. Pages that answer the sub-questions an engine searches for.
3. Mentions of your brand on the sites the engines already cite.

How do I do keyword research for LLM search?

Research prompts. Not keywords.

Collect the full questions buyers ask in sales calls, support tickets and forums.

Break each prompt into the sub-questions an engine would search for. Check that your site answers each one. See tactic 2.

Is LLM SEO the same as GEO and AEO?

Mostly. Google calls AEO (answer engine optimization) and GEO (generative engine optimization) terms for work on visibility in AI search. And says that from its side this is still SEO.

AI SEO vs LLM SEO vs GEO vs LEO compares the names.

Do I need an llms.txt file?

Not for Google. Google says Google Search itself doesn't use llms.txt files. Creating one will neither harm nor help your visibility in Google Search.

That statement covers Google Search only.