AI Visibility

Does llms.txt Improve AI Search Visibility?

The data on llms.txt is in: 300,000 domains, zero statistical correlation with AI citations. Here is what the research shows, why the file cannot move the citation needle, and what actually drives AI search visibility in 2026.

Jaxon Parrott
Jaxon ParrottApr 7, 2026

No. A 300,000-domain study found zero statistical correlation between implementing llms.txt and AI citation frequency. Removing the llms.txt variable from the predictive model actually improved its accuracy.

The pitch for llms.txt sounds straightforward: place a file at the root of your domain that tells AI systems what your company does, which pages matter most, and how they should describe you. A structured identity signal, purpose-built for the age of large language models.

Multiple independent studies found no statistically significant relationship between implementing llms.txt and how often AI engines cite your brand. The file that promised to speak directly to machines turns out to have very little to say to them.

This post covers what the data actually shows, why the mechanism behind llms.txt is the wrong one for AI citation, and what the evidence says about how brands earn AI search visibility in 2026.

Key takeaways

  • The 300,000-domain study found zero correlation. SE Ranking's analysis found no measurable relationship between having llms.txt and citation frequency. Removing the variable improved model accuracy.
  • Google has not adopted it as a signal. Google's official AI search guidance makes no mention of llms.txt as an input for AI Overviews or AI Mode.
  • Only 10% of domains have implemented it. Low adoption across all traffic tiers shows the market has not treated it as a baseline requirement.
  • AI engines mostly cite earned media, sources brands neither own nor pay for, not brand-authored identity files. Muck Rack's What Is AI Reading? (May 2026) found 84% of cited links came from earned media, the sources brands neither own nor pay for — a category spanning journalism, third-party corporate content, Wikipedia, Reddit, government and academic sources, with journalism alone at about 27%.
  • Stacker reported a 239% median lift in AI citations for the stories it distributed. Stacker and Scrunch's March 2026 study across 30 brands and 8 AI platforms reported third-party editorial coverage going with measurable AI citation gains.

What is llms.txt?

Llms.txt is a plain text file placed in the root directory of a domain, designed to give large language models structured information about a brand. The format was proposed as an AI-era equivalent of robots.txt: a simple, machine-readable signal that tells AI systems what to prioritize when interpreting your domain.

A typical llms.txt file includes a brand description, a list of key pages, product/service information, and guidance on how the brand prefers to be described in AI-generated responses. The intent is to reduce the ambiguity that AI systems encounter when drawing on scattered or outdated content about a company.

The analogy to robots.txt and sitemap.xml makes the concept easy to explain. It also contains the flaw. robots.txt tells crawlers what they can access. sitemap.xml tells search engines what content exists. Both communicate directly with systems designed to receive those signals. llms.txt assumes AI citation systems work the same way. The data shows they do not.

Unlike search engine crawlers, which follow machine-readable directives, AI citation systems synthesize information from across the web. They evaluate sources based on signals that predate the file by years: publication authority, third-party corroboration, editorial credibility, and the structural weight of who is talking about you versus what you say about yourself.

What does the research data show about llms.txt effectiveness?

Five independent studies have tested llms.txt effectiveness. All five found no measurable benefit for AI search visibility. The largest study, spanning 300,000 domains, found that removing the llms.txt variable from its predictive model actually improved accuracy.

SE Ranking analyzed approximately 300,000 domains using both statistical correlation tests and an XGBoost predictive model. The core finding: the file was not just neutral. It added noise. SE Ranking concluded that llms.txt "doesn't seem to directly impact AI citation frequency. At least not yet."

ALLMO asked the same question from a different angle: do pages with llms.txt outperform their peers in AI visibility? They reviewed the top 50 domains across three high-performance rankings and checked how many had implemented the file. No advantage was detectable.

OtterlyAI took a behavioral approach, monitoring AI crawler activity. Their findings were consistent: the file did not receive preferential treatment from AI crawlers.

IndexLab ran its own test in late 2025. Their updated analysis found no measurable effect on AI citation rates.

Search Engine Land tracked 10 sites directly, monitoring changes in AI engine treatment after llms.txt implementation. No measurable benefit appeared.

StudySample sizeMethodFinding
SE Ranking (Nov 2025)300,000 domainsStatistical correlation + XGBoost modelNo correlation; removing llms.txt improved model accuracy
ALLMO (Jan 2026)Top 50 domains across 3 AI rankingsAdoption rate vs. AI citation rankNo advantage for domains with llms.txt
OtterlyAI (2025)Multi-domain crawler behaviorAI crawler monitoringNo elevated crawl priority for llms.txt
IndexLab (Oct 2025)Multi-site before/afterPre/post citation rate comparisonNo measurable effect on citation frequency
Search Engine Land (2025)10 tracked sitesLongitudinal AI engine monitoringNo measurable benefit post-implementation

Adoption is thin as well: in SE Ranking's study, almost nine out of ten of the domains it checked had no llms.txt file.

Do AI engines cite llms.txt files themselves?

Almost never, and when they do it is Perplexity. The five studies above ask whether having the file changes citations. A narrower question is whether an engine ever lists the file as a source, and our sister research publisher's answer ledger can count that.

Machine Relations Research's count of cited addresses in its Index answer ledger (Machine Relations Research is a Bramwell company, as AuthorityTech is) counted 201,266 cited-URL events from ChatGPT, Claude, Gemini, Google AI Mode and Perplexity on 1,008 buying questions, July 6 to November 9, 2026. An llms.txt file was listed as a source three times, all by Perplexity, for customer.io, docs.perplexity.ai and mastra.ai; 93.2% of the events were addresses with no file suffix at all. The ledger records the address an engine listed, not whether it read an llms.txt file.

Three events is not evidence for or against the file. An engine can read an llms.txt file and cite the page it helped it find, and the ledger cannot see that. What the count does show is where citations go: to ordinary web pages. The file that was built to be read by machines is almost never the thing a machine lists as its source.

Why can't llms.txt move the AI citation needle?

Llms.txt targets the wrong layer of the AI citation decision. AI engines evaluate external corroboration and editorial authority, not self-authored identity files. The mechanism behind the file assumes AI systems have an informational gap your text file can fill. The actual gap is reputational.

AI language models learn what to trust from the web as it existed during training. Citation behavior reflects patterns in training data: which sources appeared consistently, which were cited by other credible sources, which had claims independently corroborated. That signal landscape was built over years.

When an AI engine decides whether to cite a brand, it draws on everything it has internalized about that brand's web presence. Does the brand appear in coverage from high-authority publications? Do multiple independent sources describe the brand consistently? Has expertise been validated by journalists, analysts, and domain experts with no financial stake?

A text file at the root of the brand's own domain contributes nothing to those signals. It is self-declared.

Google's own guidance points the same way. Its documentation for AI features in Search says there are no additional requirements to appear in AI Overviews or AI Mode, and that you don't need to create new machine-readable files, AI text files or markup to appear in them. The SEO fundamentals that apply to Search apply there too.

What sources do AI engines actually cite?

Earned media, sources a brand neither owns nor pays for, accounts for the large majority of AI citations — 84% in the most recent Muck Rack edition, and between 82% and 89% across every edition since July 2025. Journalism alone is about a quarter of all cited links, a large and stable slice inside that category.

Muck Rack's What Is AI Reading? study, covering more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries, found that 84% of them came from earned media, the sources brands neither own nor pay for. That category is broader than press coverage: it bundles journalism with third-party corporate content, Wikipedia, Reddit, PubMed, government and academic sources. Journalism alone accounted for about 27%.

Stacker and Scrunch ran a distribution study across 30 brands, 87 stories, and 8 AI platforms (2,600+ prompts). Stacker reported a 239% median lift in AI citations for stories distributed through its network of outlets, against brand-owned content alone. Its smaller December 2025 pilot of eight stories had reported a 325% lift (about 8% of answers citing only the brand's original, about 34% citing some version of the stories); the larger study's median rests on 87 stories rather than eight. Median cross-platform coverage rose from 5.4% to 17.9%.

Ahrefs studied 75,000 brands and measured the correlation between visibility signals and AI Overview brand visibility. Brand web mentions correlated at 0.664. Traditional backlinks correlated at 0.218. Those are correlations in Google AI Overviews, not proof of cause. Ahrefs study.

WorldCom PR Group, a global network of independently owned PR agencies, says that research shows up to 90% of citations driving brand visibility in LLMs come from earned media sources.

Hard Numbers, a communications analytics firm, found that 61% of LLM responses reference earned editorial media.

StudySampleEarned media citation rate
Muck Rack, What Is AI Reading? (May 2026)25M+ links cited by ChatGPT, Claude, Gemini; 17 industries84% earned media (non-owned, non-paid) (journalism, third-party corporate content, Wikipedia, Reddit, government, academia); journalism alone about 27%
WorldCom PR Group (2025)160-agency global analysisUp to 90% from earned sources
Hard Numbers (2025)LLM response audit61% reference earned editorial
Muck Rack, What Is AI Reading? (July 2025)Links cited by AI engines, first editionAbout 89% earned media (non-owned, non-paid); journalism about 27%
Stacker + Scrunch (Mar 2026)30 brands, 8 AI platforms, 2,600+ prompts239% median lift in citations for distributed stories (Stacker-reported)
Ahrefs (2025)75,000 brandsBrand mentions correlate more strongly than backlinks with AI Overview brand visibility

Full synthesis of this research body and source documentation here.

What is the citation signal hierarchy for AI engines?

AI citation signals fall into three tiers. llms.txt sits at the bottom. Understanding this hierarchy explains why the file cannot substitute for editorial authority.

Authority signals (highest tier): Signals produced by editorial decisions from third parties with no financial stake in your brand. Journalists who covered you, editors who published your analysis, researchers who cited your data, publications that included your company in a comparison. These carry maximum weight because they represent external validation.

Technical signals (middle tier): Structured data, schema markup, E-E-A-T implementation, FAQ schema, and well-formatted content. These help AI systems parse and extract content accurately. They function as a multiplier on top of authority signals, not a substitute.

Identity signals (lowest tier): Your website copy, About page, product documentation, and llms.txt. These communicate what you want AI systems to know. They do not communicate what independent observers have verified. AI citation systems prioritize the latter.

This hierarchy reflects how AI systems are trained to evaluate credibility, the same way a human evaluates a brand differently based on an independent review versus the brand's own marketing copy. Firebrand Marketing reached this directly: earned media and PR are not supplementary to AI visibility — they are the foundation.

What actually builds AI search visibility?

Three categories of action produce measurable AI citation gains, and none of them involve implementing a text file at your domain root.

Earn coverage in publications AI engines already trust

AI engines have implicit trust hierarchies built into their training data. Publications with high editorial authority appear with higher frequency in training corpora and carry more citation weight. A placement in a publication the AI engine has encountered thousands of times produces a more durable citation signal than ten blog posts on your own domain.

RankEdge's synthesis of the March 2026 research body concluded that traditional SEO tactics (backlinks, keyword-optimized blog posts, internal linking) barely affect AI citation share. What moves the needle is whether credible sources independently describe your brand.

Distribute through third-party networks, not just owned channels

The Stacker and Scrunch study is the most operationally useful data point. Stories on brand-owned channels alone were the baseline. Stories distributed through third-party editorial networks showed a 239% median lift over it. Stacker calls its data observational, correlations rather than definitive causation, so read the gap as what the distributed stories showed, not as the isolated effect of the channel.

Build brand mention density, not just backlinks

The Ahrefs finding inverts the SEO playbook. In that study, backlinks correlated at 0.218 with AI Overview brand visibility and brand web mentions at 0.664. It measured mentions across the web, not only credible editorial sources, and it shows association, not impact. Technical SEO remains useful alongside editorial authority and is unlikely to substitute for it; Ahrefs did not test that interaction. Understanding earned media ROI in the AI visibility context starts here.

How does Machine Relations explain AI citation behavior?

The llms.txt story reveals a category error that runs through much of AI visibility strategy in 2026: treating AI search visibility as a technical problem rather than an editorial one.

AI engines do not respond to technical signals the way crawlers respond to robots.txt. They respond to trust signals the way editorial systems respond to source reputation. The question they answer when deciding whether to cite your brand is not "has this brand implemented the correct file format?" It is "has this brand's claims been independently verified by sources I trust?"

That question has a name: Machine Relations. The discipline of earning AI citations through third-party editorial authority — securing the coverage, placements, and independent validation that trains AI systems to treat your brand as a citable source — is structurally closer to public relations than to technical SEO.

Muck Rack (a PR analytics company) produced the data GEO practitioners now cite as the clearest evidence of what AI engines cite: overwhelmingly earned media, sources brands neither own nor pay for. Ahrefs (an SEO data company) published research showing brand web mentions correlating more strongly than backlinks with Google AI Overview brand visibility. Stacker (a content distribution platform) ran a study reporting a 239% median citation lift for the stories it distributed, against brand-owned content alone. When a PR analytics tool, an SEO data company, and a content distribution platform each report a pattern pointing the same way, the direction is hard to ignore.

Todd Ringler, Head of U.S. Media at Edelman, described the implication: "So-called generative engine optimization is going to be front-and-center in any successful brand or reputation campaign. Unlike SEO, GEO focuses on authoritative content to give it a leg up on discoverability within AI platforms. Earned media and content strategies need to be savvy to where and how AI search is finding and structuring its answers."

As Jaxon Parrott has detailed in his analysis of the 86% problem in AI search, the brands that appear in AI-generated answers are overwhelmingly those with earned media presence, not those with every available technical file standard. The strategic framework for building that earned media foundation starts with understanding which publications AI engines trust and securing consistent placement in those sources.

Frequently asked questions about llms.txt and AI visibility

Does llms.txt hurt AI search visibility?

No. The file appears to be neutral. Implementing it does not improve visibility, and not implementing it does not create a penalty. SE Ranking's 300,000-domain analysis found no positive or negative effect.

Do AI engines ever cite an llms.txt file?

Rarely. In the Machine Relations Index answer ledger, an llms.txt file was listed as a source three times across more than two hundred thousand cited-URL events, all by Perplexity. The ledger records the address an engine listed, not whether an engine read the file, so the count says where citations go, not whether the file is used.

Are there any scenarios where llms.txt provides value?

Internal documentation systems, enterprise workflows where AI tools reference specific domain content, and developer tools that explicitly support the format may benefit. For mainstream AI search visibility — whether your brand gets cited when someone asks ChatGPT or Perplexity about your category — current evidence shows no benefit.

If technical signals don't drive AI citations, why does structured data matter?

Structured data helps AI systems parse and extract content accurately. The distinction is between extractability and citability. Structured data improves extractability. Editorial authority is a larger influence on citability. Both matter, and we think they act at different points in the decision chain, but no study cited here has tested that.

How long does earned media take to improve AI citations?

Stacker and Scrunch's March 2026 study, which Stacker says ran over 30 days, reported a median 239% citation lift for the stories it distributed. It took one post-distribution snapshot and does not say how soon after distribution the lift appeared. For broader AI training integration, the timeline depends on publication authority, coverage frequency, and brand mention consistency across multiple sources.

Which publications should brands target for AI search visibility?

Publications with long editorial histories, high traffic, and strong third-party credibility signals: national business media, industry trade publications, and data journalism outlets. A breakdown of which specific publications appear most often in AI engine citations is available at which publications get cited most by AI search engines in 2026.

Start your visibility audit

Correction, November 11, 2026

This page gave WorldCom PR Group a count of member agencies that WorldCom's page does not state, and said WorldCom found the earned-media share; WorldCom attributes it to research it does not name. The count is removed and the sentence now says what WorldCom says.

Correction, November 12, 2026

This page linked a SearchSignal blog post as Google's AI search documentation, and said Google had named domain authority, E-E-A-T, third-party coverage and validated structured data as the signals AI Mode and AI Overviews rely on. Google's documentation says something different: there are no additional requirements to appear in those features, and no need for new machine-readable or AI text files. The paragraph now cites Google's page and says what it says. The adoption figure now cites SE Ranking's study, which measured it, in place of SearchSignal's summary.