Does llms.txt work?

The short answer

It depends entirely on what you mean by "work."

If you want llms.txt to improve your Google rankings: no, it does not work for that. Google has publicly stated it does not use llms.txt as a ranking signal.

If you want automatic citations from ChatGPT, Claude, or Perplexity: that outcome is not established. No major provider has documented automatic inference-time use of the file.

The testable value is narrower: you can publish a curated index, observe requests in your own logs, and configure a retrieval workflow to consume it. Each result must be measured separately.

What "working" depends on your goal

Before looking at evidence, it helps to be precise. People ask "does llms.txt work?" with four different goals in mind:

1. Goal A, LLM citation: When someone asks an AI assistant about my product, does it give accurate, up-to-date answers and cite my pages?
2. Goal B, AI crawler coverage: Do GPTBot, ClaudeBot, PerplexityBot fetch my llms.txt and use it to prioritize what they crawl?
3. Goal C, Google SEO: Does publishing llms.txt improve my ranking on Google?
4. Goal D, controlled retrieval: Does a client or RAG pipeline that I configure use llms.txt as its reading list?

Each goal has a different answer. We cover them in order below.

Does it help LLM citations? (Goal A)

The honest caveat first: there is no Google Search Console equivalent for LLM citations. You cannot open a dashboard and see "Perplexity fetched your llms.txt and cited you 47 times this week." The measurement problem is real.

llms.txt can provide a curated reading list to a compatible client. If you control that client, you can verify that it loads the file, follows selected links, and uses the retrieved content. That does not predict the behavior of an unrelated chatbot.

A citation test also needs a control group and repeated prompts. A better answer after publication is only an observation until the retrieval trace shows that the file or one of its linked pages was actually used.

The weakest part of Goal A is the major inference-time LLM providers (OpenAI, Anthropic, Google). None has publicly confirmed that ChatGPT, Claude, or Gemini fetches /llms.txt at inference time when a user asks a question. The file likely benefits retrieval-augmented pipelines more than it benefits base models responding from their training weights.

Verdict on Goal A: unconfirmed for consumer assistants. Measurable only in a retrieval workflow whose inputs and traces you control.

Do AI crawlers actually fetch it? (Goal B)

A crawler may request the file, but publication alone does not guarantee any request. Provider documentation distinguishes crawler purposes and robots.txt controls; it does not establish special llms.txt handling.

Check server or CDN logs for the exact /llms.txt request, timestamp, status, and user-agent. A user-agent string is evidence of a request, not proof of identity, parsing, indexing, prioritization, or use in an answer.

Use a verified reverse-DNS or provider-recommended validation method when attribution matters. Keep "no request observed" separate from zero use when log coverage is incomplete.

Verdict on Goal B: measurable per site from complete logs. Structured or priority use remains unconfirmed unless a provider documents it.

Does it help Google SEO? (Goal C)

No. This is the clearest answer of the four.

Google's current guidance states that Search does not use llms.txt or other special AI text or Markdown files, including for generative AI features. Publishing one brings no special treatment in Google Search.

This is consistent with how Google treats robots.txt (access control, not ranking) and sitemap.xml (crawl discovery, not ranking). Neither file improves your position in search results on its own; they affect whether and how Google can access your content. llms.txt is not in the same category as those files from Google's perspective, it simply is not part of its pipeline at all.

Verdict on Goal C: No effect on Google rankings. This is confirmed, not speculative.

Who has published llms.txt? (Goal D context)

Publication is measurable, but it must not be confused with consumption. In the fixed llmtxt.info panel, 119 of 217 reachable hosts served a valid root file on 2026-08-31, a rate of 54.8%.

This panel measures HTTP availability and file validity on a defined set of hosts. It does not measure traffic, provider support, citation lift, or ranking impact. See the adoption tracker for the panel and methodology.

For Goal D, run a controlled ingestion test: record the fetched file, the pages followed, the retrieved passages, and the answer. Compare it with the same workflow without llms.txt.

Verdict on Goal D: testable and potentially useful in a configured workflow, not proven by publication counts alone.

How to measure whether it is working

No single dashboard answers this today, but you can triangulate:

• Server logs. Filter for requests to /llms.txt and /llms-full.txt. Look for user-agents: GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Applebot-Extended. Frequency and recency of these hits tells you whether AI crawlers are actively interested.
• Referrer traffic. Watch for referrers from chat.openai.com, claude.ai, perplexity.ai, and similar domains. An increase after publishing llms.txt is not proof of causation, but it is worth tracking.
• Retrieval traces. In a workflow you control, store the fetched URLs and passages. This is the strongest evidence that llms.txt influenced an answer.
• Manual spot-checks. Ask Claude, ChatGPT, and Perplexity a question that your site should answer authoritatively. Note whether the answer is accurate, whether your pages are cited, and whether the quality improves after you publish or improve your llms.txt.
• Developer feedback. If your site serves developers, ask them directly: "When you ask an AI tool about [your product], do you get accurate answers?" This qualitative signal is often the most actionable.

Verdict

llms.txt reliably works as a publishable, machine-readable index. Anything beyond that depends on the receiving client and requires separate evidence. It is most testable for:

• Developer documentation sites
• API and SaaS products with technical buyers
• Any site whose users regularly ask AI assistants questions about the product

It does not work as a Google SEO lever. If that is your goal, you are in the wrong place, focus on structured data, Core Web Vitals, and editorial backlinks.

Publishing also creates maintenance work and a risk of stale or contradictory guidance. The rational decision is to publish only when you can keep the file current and name a compatible workflow or measurement you intend to test.

Next steps

• How to create an llms.txt file, templates and deployment guide for every common stack.
• Generator, fill a form and download a spec-compliant file in minutes.
• Validator, check your existing file against the spec.
• Best practices, what separates a useful llms.txt from noise.