Does llms.txt increase AI citations?

Not on its own, and the honest evidence is thin. llms.txt is a discovery aid: it tells a client which pages exist and what each one answers. Published experiments that measured citation share after adding the file found no reliable lift, and large-scale crawler audits show most llms.txt files are never requested. Ship it because it is cheap and because some agents do read it — then put your effort into crawl eligibility, indexation in the engine's own index, and passages a model can quote.

Why is my site never cited even though I rank on Google?

Because the engine may not be searching Google. ChatGPT's retrieval layer is fed by Bing, so a page that Bing has not indexed is effectively invisible to it; Google AI Overviews, by contrast, draws on the Google index. Check the engine-specific index first: verify the domain in Bing Webmaster Tools, submit the sitemap, and confirm the URL is indexed there. Ranking well on Google while being absent from Bing is the single most common cause of missing ChatGPT citations.

How do I measure whether AI engines cite me?

Three signals, none of them perfect. Run a fixed list of twenty buyer-intent prompts monthly across ChatGPT, Perplexity and Google AI Overviews, and record whether your domain appears in the answer and which URL is cited. Then watch referral traffic: filter GA4 or your server logs for referrers such as chatgpt.com, perplexity.ai and copilot.microsoft.com. Finally, check your log files for OAI-SearchBot, PerplexityBot and Google-Extended fetches to confirm the retrieval layer can even reach your pages.

GEO · September 2026

How to Get Cited by AI Search Engines

A citation is what happens when a model decides one page is the most useful thing it read this second. That decision is made upstream of your prose: in crawl eligibility, in the index the engine searches, and in whether a single paragraph of yours can be lifted without edits.

Eligibility Comes Before Content

Every AI answer pipeline has a door, and most sites are accidentally holding it shut. OpenAI publishes three user agents that do three different jobs, and confusing them is the most common technical error in this space:

OAI-SearchBot — builds the index behind ChatGPT's search and citation layer. Block it and you have removed yourself from consideration.
GPTBot — collects training data. Allowing it shapes whether your brand is part of the model's baseline knowledge, not whether you get cited today.
ChatGPT-User — fetches a specific URL live when a user's request triggers a visit. Block it and the model cannot open the page it was about to quote.

Perplexity uses PerplexityBot and Perplexity-User in the same two roles, and Microsoft's bingbot plus IndexNow govern Copilot. The full allow/block decision table is in robots.txt for AI crawlers.

Which Index Does Each Engine Actually Read?

"I rank on Google" is not a retrieval strategy. Each assistant queries a different corpus, and a page missing from that corpus is invisible regardless of quality.

Engine	Retrieval layer	Practical consequence
ChatGPT	Bing-fed index plus OAI-SearchBot	Bing indexation is a prerequisite; verify in Bing Webmaster Tools
Google AI Overviews	Google's own index	Classic technical SEO still applies; top-10 pages dominate citations
Perplexity	Its own real-time crawl	Freshness and forum/news-adjacent sources carry unusual weight
Claude	Third-party search backend	Being present in general web indexes matters more than any single file
Microsoft Copilot	Bing plus IndexNow	Push updated URLs through IndexNow instead of waiting for a recrawl

A Seer Interactive analysis of SearchGPT citations found roughly 87% matched Bing's top ten results, and a June 2025 Ahrefs study reported only around 14% overlap between the top sources cited by Google AI Overviews, ChatGPT and Perplexity. Treat every engine as a separate channel with its own index, not as one audience. The engine-by-engine fetcher audit in which AI engines read llms.txt makes the same point from the crawler side.

Write Passages a Model Can Lift Verbatim

Answers are assembled from short spans, so the unit of optimisation is the paragraph, not the page. Four habits separate quotable pages from pages that get read but never mentioned:

Answer in the first 40–60 words. State the conclusion before the context. A model scanning for evidence needs a self-contained span, not a build-up.
Keep one idea per paragraph, with the entity named in full. "It reduces latency" is unciteable outside its page; "Cloudflare's content negotiation reduces latency" survives extraction.
Make tables and lists carry data. Comparisons, limits, thresholds and prices are the spans most often reproduced, because they are hard to paraphrase without loss.
Ship numbers inside the sentence, not only inside an image or a chart. If the figure is not in the text, it is not in the corpus.

One striking data point cuts against the intuition that higher rankings always win: a June 2025 Semrush study found that nearly 90% of pages ChatGPT cited were ranking 21 or lower in classic search. Retrievability and quotability beat position, which is why passage formatting is worth more than another round of keyword tuning.

Corroboration: The Part You Cannot Write Yourself

Citation selection is biased toward sources the model has seen agree with each other. Aggregated audits consistently find that the cited mix for ChatGPT is dominated by brand pages, news and media, community threads and encyclopedia entries — a large share of which are not owned by the brand being discussed. The practical implication is unglamorous: digital PR, a coherent entity story, and an honest presence in the communities where your topic is debated do more for citation share than a rewrite of your own homepage.

Entity consistency across the web — the same name, description and category everywhere — is what lets an engine decide your page and that third-party mention are about the same thing. That is the same alignment problem the llms.txt and Schema.org combination solves at the file level.

Where llms.txt Fits — and Where It Does Not

Be sceptical of anyone selling llms.txt as a citation lever. Controlled experiments that added the file and measured citation share reported no reliable lift, and crawler audits keep finding that the large majority of published llms.txt files are never requested. What the file genuinely does is two smaller jobs: it gives agent-style clients an accurate map of what your site contains, and it shortens the path from a fetched index to the right page.

That makes it a hygiene item, not a strategy. Publish a short, accurate file, keep the descriptions specific enough to be quoted, and spend the remaining effort on the three things that move citations: crawl eligibility, presence in the engine's own index, and extractable paragraphs. The wording rules for the summary and link annotations are covered in how to write llms.txt descriptions, and the measurement loop — logs, referrals and prompt tracking — is in measuring whether AI engines read your file.

A 30-Day Citation Check

Run this in order; each step takes minutes and the fourth is the only one that changes week to week.

# 1. Can the citation crawlers reach you at all?
curl -s https://example.com/robots.txt | grep -i -E "oai-searchbot|gptbot|perplexitybot"

# 2. Is the page in the index ChatGPT actually searches?
curl -sI https://example.com/pricing | grep -i -E "x-robots-tag|last-modified"

# 3. Does your llms.txt describe the pages you want cited?
curl -s https://example.com/llms.txt | head -20

# 4. Baseline twenty buyer prompts monthly and record whether your domain appears

Steps one to three fix eligibility and discovery; step four tells you whether any of it worked. Start from a validated file rather than a hand-typed one — the llms.txt generator produces a spec-compliant skeleton, and the llms.txt checker catches the broken links and delivery errors that quietly remove a file from consideration.