Does llms.txt increase AI citations? The 2026 data, who should publish one, and what to fix first.
llms.txt for AEO promises a shortcut: place one markdown file in your site root and AI engines cite you more. The 2026 evidence does not support that promise for most B2B sites. SE Ranking found no correlation between the file and AI citations across nearly 300,000 domains, and Google says its Search systems ignore the file. This guide shows what the data covers, who should still publish one, how to write it in an hour, and which fixes pay back faster.
llms.txt is a plain markdown file at yourdomain.com/llms.txt that lists your most useful pages with one-line descriptions for AI agents. Jeremy Howard of Answer.AI proposed it on 3 September 2024, and a second version of the proposal appeared in August 2026, according to INITE's 2026 review. No standards body has adopted it.
Unlike robots.txt, llms.txt controls nothing. It does not allow or block any crawler. It works as a curated index: you pick the pages, you write the summaries, and an agent that chooses to fetch the file gets a map of your site without parsing navigation menus and scripts. A companion file, llms-full.txt, packs full page content into one markdown document.
For AEO the question is narrow. Do answer engines use that map when they decide which sources to cite? The data below answers it.
No published study shows a citation lift from llms.txt. SE Ranking analyzed nearly 300,000 domains and published its results on 7 November 2025. It found the file on 10.13% of domains and no correlation between the file and AI citations. When SE Ranking removed the llms.txt variable from its XGBoost model, the model predicted citations slightly better, which suggests the file added noise.
Trakkr compared 37,894 AI-cited domains. Sites with llms.txt averaged 6.8 citations and sites without it averaged 6.7, with a p-value of 0.81. A gap that small is noise. Trakkr concluded that citations track domain authority, content depth, and training data exposure.
Both studies measure correlation across many sites. Neither rules out an effect on one narrow class, such as documentation sites that coding agents read.
Google does not use llms.txt. Gary Illyes said at a Google Search Central event in Bangkok in July 2025 that Google does not support the file and has no plans to. John Mueller compared it to the keywords meta tag and pointed out that bots already fetch the real pages.
Google's AI optimization guide, last updated 10 July 2026, states that publishing the file will neither harm nor help your visibility in Search. AI Overviews and AI Mode are Google Search features, so treat Google as a closed question.
Wix AI Search Lab reported 30,000 to 60,000 llms.txt files in Google's index in October 2025. Indexing a file is different from using it to write answers.
No major AI vendor documents llms.txt as a citation input. INITE reviewed the crawler documentation from OpenAI, Anthropic, and Perplexity on 7 October 2026 and found no statement that OAI-SearchBot, GPTBot, ClaudeBot, PerplexityBot, or their user-triggered counterparts request or prefer the file. Those crawler names are robots.txt controls.
Agents that already sit on your site fetch the file more often. Coding assistants reading developer docs are the clearest case. The OpenAI and Perplexity documentation pages carry a hidden note that points agents to each company's own /llms.txt, which INITE reads as a doorway for agents already on the site, not a citation signal.
Other operators report small upticks in requests for the file in their server logs. Nobody has published an authoritative log study, so check your own logs before you trust anyone's anecdote.
Developer-tool and AI companies lead adoption, and their use does not prove a citation lift. Anthropic, Stripe, Cloudflare, Mintlify, Vercel, and Perplexity publish the file. INITE's sweep from 30 September to 7 October 2026 found that 5,473 of 10,066 live AI-tool websites (54%) served an /llms.txt file.
Broad samples look different. SE Ranking found the file on 10.13% of its domains, and larger sites adopted less: 8.27% of sites with more than 100,000 visits had it, against 9.88% of sites with up to 100 visits.
File sizes show the intent. On 7 October 2026 INITE measured Anthropic's llms-full.txt at 41,007,961 bytes and Cloudflare's at 67,357,207 bytes, while OpenAI's llms.txt measured 6,201 bytes. A 41 MB file serves a developer's agent. It does not serve a buyer who asks ChatGPT for a vendor shortlist.
Publish it if you maintain technical documentation or can generate it from a canonical source in about an hour. Skip it if it would replace content work. The cost is low and the downside is near zero while the file stays accurate. A stale file costs you credibility, because an agent that reads it gets dead links and old pricing.
Publish if:
Skip or postpone if:
Follow the proposal's structure: an H1 with your name, a blockquote summary, then H2 sections of markdown links with short descriptions. The steps take about an hour for a mid-size SaaS site.
# Acme Analytics
> Acme Analytics is a product analytics platform for B2B SaaS teams. This file lists the pages most useful for answering questions about Acme.
## Product
- [Platform overview](https://acme.com/platform): Features, supported data sources, and architecture.
- [Pricing](https://acme.com/pricing): Current plans, limits, and billing terms.
## Docs
- [API reference](https://acme.com/docs/api): Endpoints, authentication, and rate limits.
## Optional
- [Changelog](https://acme.com/changelog): Release notes by date.Monitor your progress with Nobori: track whether the pages you list in llms.txt gain citations over the following weeks, engine by engine, at nobori.ai.
llms-full.txt makes sense only as a self-contained documentation export for agents. Tobira's 2026 review of the evidence recommends it for that use alone. Anthropic's and Cloudflare's files run to tens of megabytes because they package entire documentation sets.
Marketing pages add nothing to such a file. Put API references, setup guides, and changelogs in it, and generate it from source on every deploy so it cannot drift from the live pages. If your site has no documentation, skip llms-full.txt and keep the short index.
Do not wrap the file in a plugin that sells itself as AI optimization. A plain, hand-checked file is easier to keep accurate than a generated file nobody reads.
Request the file with curl and confirm an HTTP 200 and a text content type. Run curl -I https://yourdomain.com/llms.txt and read the status line. A page that returns 200 with an HTML "not found" body is a soft 404, and INITE counted soft 404s as missing files in its sweep.
Then check every link inside the file. Add a CI step that requests each listed URL and fails the build on any non-200 response or redirect. Compare each URL to its canonical tag so the file points to the version you want cited.
Google's Lighthouse audit flags llms.txt only on a server error, and it marks a 404 as Not Applicable because the file is optional. A clean Lighthouse report does not confirm that your file works.
Monitor your progress with Nobori: see which of your listed pages AI engines cite and which they skip at nobori.ai.
A correct llms.txt does not hurt AI visibility, and Google states it will neither harm nor help Search. The risk comes from errors you introduce yourself. Three mistakes show up most often.
The first is staleness: the file lists a pricing page you retired or a plan you renamed. An agent that trusts the file repeats the old claim, and our research on AI brand sentiment shows pricing as the claim category where engines err most.
The second is blocking: a CDN rule or robots.txt line that returns 403 to the file's requesters, so the file exists but nobody can read it. The third is sprawl: a list of 400 links with no priority tells an agent nothing. Keep the index short and send everything else to the Optional section.
Fix the factors that citation data does support before you spend a day on llms.txt. Trakkr's data points to domain authority, content depth, and training data exposure. Four fixes map to those signals.
First, make your content readable by crawlers that skip JavaScript. Our guide to JavaScript rendering for AI crawlers covers the curl test and the rendering fixes.
Second, confirm robots.txt and your CDN let answer-engine crawlers in. The crawl-to-refer ratio study includes a three-layer access audit.
Third, restructure pages into self-contained sections with the chunk-first framework.
Fourth, build the brand signals models use to recognize you. Start with entity optimization for AI search and then earn third-party mentions.
Run a before-and-after test with a fixed prompt set and a control group of pages. Build 30 prompts that match your buyers' questions. Record citation rate per engine for four weeks, publish the file, and record for another four to eight weeks.
AI citations drift from week to week, so a raw before-and-after comparison misleads. Compare pages listed in llms.txt against similar pages you left out, and count a change only if the listed pages move while the control pages stay flat. Our citation half-life data shows how much baseline churn to expect.
The head-to-head comparison of citation monitoring tools helps you pick a tracker that can segment results by URL.
Monitor your progress with Nobori: compare cited-page share across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews at nobori.ai.
Nobori tracks your brand across ChatGPT, Google AI Overviews, Perplexity, Gemini, and Claude with a fixed prompt set, so you can run the llms.txt test described above without a spreadsheet. You see which URLs each engine cites for your prompts, and you can tag the pages listed in your file to compare them against the rest of your site.
The same view shows where competitors earn citations that you miss. That tells you whether your next hour belongs in llms.txt or in the rendering, structure, and entity work that citation data supports.
Daily tracking also catches drift. If a listed page loses citations after a deploy, you see it that week, not at the next quarterly review.
See if AI engines are citing you → nobori.ai