llms.txt AI Citation: Does It Actually Work?.

A lot of B2B engineering teams spent an afternoon in the past year shipping an llms.txt file because someone on the marketing team asked for it. The llms.txt AI citation promise sounds reasonable on the surface: a plain-text manifest at the root of your domain, written for language models instead of search crawlers, that tells ChatGPT, Perplexity, Gemini, and Copilot exactly which pages matter. The evidence says that promise has not materialised. No major AI provider has confirmed reading the file in production, and the largest study of real-world adoption found that the overwhelming majority of published files are never requested at all.

This is not an argument against technical AEO work. It is an argument for spending that engineering time on the handful of things that verifiably move an llms.txt AI citation outcome, instead of a file most AI systems never open.

What is llms.txt and why did B2B teams rush to add one?

llms.txt is a proposed standard, published on a root domain alongside robots.txt and sitemap.xml, that lists a site's most important pages in a short markdown format intended to be easy for a language model to parse. It was pitched in 2024 as an "AI-friendly" equivalent of a sitemap, and by 2025 it had become a checkbox item on AEO checklists across the industry. Teams that had never looked closely at robots.txt or structured data suddenly had a new, simple-looking task, and simple tasks get done regardless of whether they work.

The appeal was speed. Unlike entity disambiguation or a genuine content restructure, an llms.txt file can be written and deployed in under an hour. For a marketing lead who has noticed AI answers citing competitors instead of their own company, that speed felt like progress. The problem is that speed of implementation and evidence of effect are two different things, and the industry moved on the former for a year before anyone checked the latter.

Does llms.txt improve AI citation? What the evidence shows

llms.txt does not currently have any verified effect on AI citation, based on the largest study of live adoption to date. Ahrefs analysed 137,210 domains that received traffic in May 2026 and found that 28% had published an llms.txt file, but of the roughly 38,000 valid files, 97% received zero requests in that month. Of the small fraction that were requested at all, only 19.5% of that traffic came from named AI tools; the rest came from SEO audit software, generic bots, and profiling tools like BuiltWith checking whether the file existed, not reading it for retrieval.

That is the practical answer to the llms.txt AI citation question most B2B teams are actually asking: will adding this file get us cited more often. The honest answer, based on the current evidence, is no measurable effect, because the crawlers that actually feed AI answers are barely present in the request logs of the sites that publish the file.

What do OpenAI, Google, Anthropic and Perplexity actually say about llms.txt?

None of the major AI providers have publicly confirmed that they read or act on llms.txt files in production. OpenAI, Anthropic, and Perplexity publish their own guidance for site owners, and that guidance covers crawler user agents and robots.txt directives, not llms.txt. Google has gone further and explicitly said the file is not needed to appear in AI-generated search results, and Google's John Mueller has described it in public comments as, at best, a minor convenience for AI coding assistants reading developer documentation, not a citation mechanism for commercial content.

That silence from the companies that would benefit most from a simple, universal standard is itself informative. A file that materially improved which sources got cited would be an easy thing for an AI provider to promote, because it would reduce their own retrieval costs. None have done so, more than a year after the format was proposed.

Why do most llms.txt files never get read?

Most llms.txt files never get read because retrieval crawlers discover and select content through search indexes, not through a manifest file sitting at a URL nobody told them to check. ChatGPT's browsing and Perplexity's retrieval both lean heavily on live web search and their own crawl indexes, meaning a page has to already be indexed, well-structured, and authoritative enough to surface in a query before an llms.txt entry would ever be relevant. The file assumes an AI system visits your root domain looking for a curated reading list. In practice, these systems arrive at a specific URL because it ranked for a specific query, not because a manifest pointed them there.

This is the same mechanism problem that shows up in AI citation signals more broadly: engines cite what their retrieval layer already trusts and can parse, not what a site self-nominates as important.

What actually improves AI citation instead of llms.txt?

The tactics that verifiably move citation rates are less novel than llms.txt, which is part of why they get skipped. In priority order:

1. 1. Confirm AI crawlers can actually reach your site. Check robots.txt and any CDN or WAF rules for GPTBot, PerplexityBot, ClaudeBot, and Bingbot. A blocked crawler makes every other tactic irrelevant.
2. 2. Structure content for extraction. Answer-first paragraphs, FAQPage and HowTo schema, and clear H2 questions give retrieval layers something they can lift cleanly into a generated answer.
3. 3. Fix entity disambiguation. A Wikidata entry, consistent business details across your site and third-party listings, and a clear About page let engines confirm who you are before they cite you.
4. 4. Build the off-site citations engines actually pull from. Review sites, industry directories, and earned press coverage are where retrieval-augmented engines find corroborating sources, which matters more than any single on-site file.

This is the order SourceRank AI works through on every account, because it matches how citation is actually earned rather than how easy a tactic is to implement. The SourceRank AI how-it-works page walks through the full methodology, and the services page covers what we execute on your behalf once the audit identifies the gap.

Is there any legitimate use for llms.txt?

llms.txt may have a narrow legitimate use for developer-facing documentation sites, where AI coding assistants sometimes reference a curated file list to reduce token usage when summarising an API. That is a different use case from B2B commercial content trying to be cited in a ChatGPT or Perplexity answer about vendors, tools, or services. Publishing the file is not harmful, and it costs little once it exists. The mistake is treating it as an AI citation strategy, or worse, as the reason a citation programme has stalled elsewhere.

How should a B2B team prioritise technical AEO work?

A B2B team should prioritise technical AEO work by fixing crawler access first, structured data second, entity signals third, and off-site citations fourth, because each layer is a prerequisite for the next one to matter. Spending a sprint on llms.txt before confirming crawler access, for instance, produces zero measurable improvement, since a file sitting next to a robots.txt block does nothing for an engine that was never allowed in.

The SourceRank AI score measures your current citation rate across ChatGPT, Perplexity, Gemini, and Copilot and flags which of these layers is your actual gap, so the next engineering sprint goes to the tactic with evidence behind it rather than the one that shipped fastest last quarter.

Frequently asked questions