Should You Really Spend Time on llms.txt in 2026?

llms.txt in 2026: Should you spend time on it? The direct answer

No, except in specific cases. For 8 out of 10 sites, the llms.txt file will not trigger any additional citation in ChatGPT, Claude, Gemini or Perplexity. The data converges. Here is my concrete decision tree, validated against the 3 reference studies available at end of May 2026.

Before copy-pasting a template found on llmstxt.org, identify your site profile. That decides, not the hype.

Site profile	Verdict	Real GEO priority
B2B SaaS with API/technical documentation (Stripe, Vercel, Cursor style)	Yes, do it	High (useful for third-party AI tools)
Media, blog, editorial site	Not a priority	Low (measured gain: zero)
E-commerce, B2C SaaS, local business site	No, waste of time	Very low (no impact demonstrated)
Agency, freelancer, consultant	Optional (client education signal)	Medium only as a demonstration

The logic behind this table rests on the key distinction John Mueller publicly drew on Bluesky on May 22, 2026: llms.txt serves AI agent “functionality”, not search engine “discoverability”. In other words, it helps an agent once it lands on your site, it does not bring it in.

If your business does not depend on AI agents parsing your documentation to execute tasks, the cost/gain ratio collapses.

What is llms.txt and what does it really do?

llms.txt is a Markdown file placed at the root of a domain (example.com/llms.txt) that gives large language models a curated index of a site’s priority content. The spec was proposed by Jeremy Howard of Answer.AI in September 2024 on llmstxt.org.

Its original goal is simple: reduce HTML noise for LLMs. A standard web page contains navigation, scripts, pop-ups and trackers that models must filter before reaching the actual content. llms.txt offers a clean Markdown summary. Token savings, faster parsing.

Jeremy Howard’s original proposal (Answer.AI, 2024)

The spec defines two variants. The llms.txt file contains a table of contents with one-line descriptions of each important resource. The llms-full.txt file embeds the full content of those resources in Markdown.

Howard frames the intent clearly: “A proposal to standardise on using an /llms.txt file to provide information to help LLMs use a website at inference time.” The key phrase is “inference time”, not “ranking time”. The author promises nothing on visibility in AI answers.

Anthropic (Claude), Cursor and Mintlify officially support it. OpenAI, Perplexity and Google do not use it as a citation signal. Neither do Meta or Mistral.

The difference with robots.txt and sitemap.xml

These three files are often confused. They play distinct roles.

• robots.txt controls crawler access (allow or block GPTBot, ClaudeBot, etc.).
• sitemap.xml lists URLs to be indexed by search engines.
• llms.txt proposes an optimized summary to LLMs that want to quickly consume your content.

Comparing llms.txt to robots.txt is misleading. robots.txt is a de facto standard respected by 100% of major crawlers since 1994. llms.txt is not standardized, not mandatory, and recognized by less than 10% of AI platforms.

The 3 studies showing that llms.txt does not impact AI citations

Three independent studies, published between November 2025 and April 2026, total 456,000 domains and over 156,000 AI citations analyzed. None finds a positive effect of llms.txt on visibility in AI answers. The convergence is what makes the conclusion solid.

SE Ranking: 300,000 domains, no measurable correlation

SE Ranking published in November 2025 the largest study to date. The team analyzed 300,000 domains, cross-referenced llms.txt presence with citation frequency by major LLMs, then applied classical statistical analysis and an XGBoost model.

Main result: removing the llms.txt variable from the predictive model improved accuracy. The file introduced noise, not signal. Measured adoption: 10.13% of domains. Among the 50 most AI-cited domains, only one had a llms.txt.

The study identifies the real predictors of citation:

• Sites with more than 32,000 referring domains: 3.5x more cited by ChatGPT
• Reddit and Quora presence (millions of mentions): 4x more chances to be cited
• Profiles on Trustpilot, G2, Capterra, Yelp: 3x more chances for ChatGPT
• Pages with FCP under 0.4 seconds: 6.7 citations on average vs 2.1 for slow pages

OtterlyAI: 84 bot visits out of 62,100 (0.1%)

OtterlyAI ran a more field-grounded experiment in February 2026. The team deployed a valid llms.txt at the root of a site and monitored server logs for 90 days, isolating AI user-agents.

Out of 62,100 AI bot visits recorded during the period, only 84 targeted /llms.txt. That is 0.1% of total AI agentic traffic. A standard content page on the same site averaged 265 bot visits. llms.txt underperformed by a factor of 3 compared to a regular page.

OtterlyAI’s conclusion: “From the bots’ perspective, llms.txt is nearly invisible.” The publisher even removed llms.txt from its public GEO audit following this study.

ALLMO: 94,000 cited URLs, llms.txt present in less than 1%

ALLMO took the opposite angle. Rather than tracking bots, the team analyzed 94,000 URLs actually cited in 11,000 AI responses collected between August and December 2025. Models included: ChatGPT, Claude, Gemini, Grok and Perplexity.

If llms.txt played a role, we would expect to find it on roughly 10% of cited sites, proportional to the global adoption rate. In reality, it was present in less than 1% of the 120 sites analyzed in depth. The effect is negligible and reproducible across 5 different models.

The 3 studies converge on a single finding: llms.txt is neither a ranking signal nor a discovery signal for major AI engines in 2026.

Why does Chrome Lighthouse audit llms.txt if Google Search says otherwise?

This is the most discussed contradiction of spring 2026. Google Search published in mid-May 2026 an official guide stating that sites need neither llms.txt, nor Markdown versions, nor specific markup to appear in Google Search AI features. A few days later, Chrome shipped Lighthouse 13.3 which audits llms.txt presence at the root of the domain.

The coherence appears as soon as you separate two distinct missions: being found (Search) versus being usable by an agent (Chrome/Lighthouse).

John Mueller’s “Discoverability vs Functionality” distinction

Lily Ray called out John Mueller on Bluesky on May 22, 2026, pointing out the irony: Google itself uses llms.txt and Markdown pages on its own properties (developers.google.com) while telling SEOs it is useless. Mueller responded by drawing a distinction he considers fundamental.

• Discoverability: being found by a search engine. That is classic SEO. llms.txt plays no role there.
• Functionality: helping an agent or AI tool perform a task once on your site. That is where llms.txt can help, on technical documentation consumed by coding assistants.

Mueller was explicit: creating a Markdown version of a shoe’s specs will not generate more sales. The useful scope is limited to technical documentation sites. He even calls the need a “temporary crutch”, as LLMs get better at parsing raw HTML.

What Lighthouse 13.3 changes (May 2026)

Lighthouse 13.3 shipped on May 7, 2026 the Agentic Browsing category out of its experimental status. It is now enabled by default. PageSpeed Insights received it within 2 weeks. Chrome 150 DevTools followed.

The audit does not produce a weighted 0-100 score like Performance. It shows a fractional pass ratio with pass/fail/warning markers on each criterion. The checks cover:

• Presence of a machine-readable summary at the root (llms.txt)
• WebMCP integration (exposing site actions to agents)
• Accessibility tree quality (a11y tree)
• Layout stability (Cumulative Layout Shift)

A Lighthouse audit flagging the absence of llms.txt is not a Google Search ranking signal. It is an agentic web readiness indicator. The two scopes do not overlap.

When llms.txt really deserves 30 minutes of your time

Three concrete cases justify implementation. Outside those, I arbitrate against.

Case 1: Technical documentation consumed by coding assistants. If your /docs pages serve as reference for Cursor, Continue, Claude Code or internal AI copilots, a llms.txt with llms-full.txt brings real token savings. Stripe, Vercel, Mintlify, Cursor have implemented the file in this precise logic. Third-party AI tools gain in speed and inference cost.

Case 2: B2B SaaS platform with API-driven product or integrations. If your customers build automations querying your docs, their AI agent is your target. Providing a clean entry point reduces friction and improves integrator satisfaction.

Case 3: Specialized editorial site with ambition to be cited in vertical tools. Rarer case, but valid if you publish reference data (industry statistics, business knowledge base) that a niche AI tool might ingest. The benefit remains limited, but the cost is minimal.

In all cases, the marginal creation cost is low (15 to 30 minutes for a well-structured site). The gain is not in mass-market AI visibility, it is in experience quality for agents that already touch your content.

When llms.txt is a waste of time: 4 affected profiles

Four profiles have no measurable return to expect. I spend my time elsewhere.

Generalist e-commerce site. Your product pages, descriptions and reviews are already parsed by LLMs via standard HTML. A llms.txt pointing to /collections/shoes will not boost your recommendations in ChatGPT. Your real levers: Trustpilot, Reddit, Quora, Google Business profiles, specialized press.

B2C mass-market SaaS. Spotify, Netflix, Airbnb do not need llms.txt to be cited. LLMs know them through the mass of existing reviews, articles and citations. The brand weighs more than the file.

Media or editorial blog site. Your individual articles carry the value, not a root file. Better to invest in content freshness and adapted Schema.org markup.

Local business or small company site. ROI is zero. Your Google Business profile, Google reviews and local linking do 100% of the AI visibility work.

The 5 real GEO priorities to execute before llms.txt

If you have 4 hours to invest in your AI visibility this week, here is the order I use them. All these priorities are sourced from the same studies that dismantle llms.txt.

1. Work on your referring domains. SE Ranking’s 32,000 backlinks threshold separates sites “3.5x more cited” from the rest. A targeted digital PR campaign beats any technical file.

2. Build a Reddit and Quora presence. Domains with millions of mentions on these platforms have 4x more chances to be cited. LLMs find the real voice of users there. Useful answers, not spam.

3. Create full G2, Capterra, Trustpilot profiles. 3x more chances to be cited by ChatGPT for sites present on these platforms. ROI is immediate and measurable.

4. Optimize First Contentful Paint under 0.4 seconds. 6.7 citations on average for fast pages, against 2.1 for slow ones. Performance is not just a Core Web Vitals signal, it is an AI prioritization signal.

5. Implement clean chunking and updated Schema.org markup. LLMs extract autonomous blocks, not entire articles. Paragraphs of 150 to 300 words, question-based titles, structured lists raise extraction probability.

How to create a clean llms.txt in 15 minutes (if you decide to go)

You fall into one of the 3 cases where it is worth it? Here is the minimal version that respects the llmstxt.org spec without overloading your site.

Step 1: List your 10 to 20 priority pages. Documentation, critical product pages, reference guides, API. Not the whole site.

Step 2: Write a one-line description per resource. 80 characters maximum. Precise, factual, no marketing.

Step 3: Structure in Markdown using the official format. An H1 with the site name, a blockquote paragraph for context, then H2 sections with link lists.

Step 4: Deploy at the root. The file must be accessible at yourdomain.com/llms.txt with HTTP 200, MIME type text/markdown or text/plain.

Step 5: Submit the URL in Google Search Console to speed up indexing. A dev5310 case study showed this submission can trigger OAI-SearchBot and ChatGPT-User crawls within 4 days.

Step 6: Monitor server logs for 30 days. Filter AI user-agents. If you see fewer than 10 hits on /llms.txt, it matches OtterlyAI’s numbers. No problem, you did it for integrators, not for Google.

A complementary llms-full.txt (Markdown export of actual content) only makes sense for technical documentation. For the rest, the table of contents is enough.

My final verdict on llms.txt in 2026

llms.txt is not a GEO lever in 2026. The data converges across 3 independent studies and Google Search’s official position confirms it. The file serves AI agent functionality, not discovery by mainstream AI engines. 30 minutes invested in a G2 profile or a useful Reddit reply will produce more citations than any llms.txt. For technical documentation sites consumed by third-party AI tools, the interest exists and the cost is minimal. For 80% of sites, it is a waste of time.

To verify the real impact of your GEO optimizations on your AI citations, try Cockpyt AI free for 14 days.

FAQ: your questions on llms.txt in 2026

Is llms.txt a Google ranking factor?

No. Google Search confirmed in May 2026 that llms.txt is not used in AI Overviews, AI Mode, or traditional rankings. Gary Illyes compared it in July 2025 to the deprecated meta keywords tag.

Does ChatGPT read my llms.txt?

OpenAI has not communicated officially on the topic. Log experiments by OtterlyAI and other independent actors show that OpenAI bots (OAI-SearchBot, ChatGPT-User) rarely access the file directly. Claude (Anthropic) is the only major LLM to have explicitly integrated support.

Do you need a llms.txt to pass Lighthouse’s Agentic Browsing audit?

Yes for that specific criterion, no for your SEO or GEO. The Lighthouse Agentic Browsing audit signals AI agent readiness, with no link to Google Search ranking or visibility in mainstream AI answers.

What is the difference between llms.txt and llms-full.txt?

llms.txt is a table of contents with short descriptions. llms-full.txt embeds the full Markdown content of listed resources. The latter is useful for technical documentation, useless for a classic editorial site.

Does WordPress automatically generate a llms.txt?

Yoast SEO and Rank Math offer automatic generation since late 2025. The feature builds a file based on the existing XML sitemap. The result is often too long and not selective enough. A 15-minute manual write produces a more useful file.

How much does setting up a llms.txt cost?

15 to 30 minutes of work for a well-structured site. No technical dependency, no plugin required. The hidden cost is the opportunity cost: that time invested elsewhere (digital PR, G2 profiles, content freshness) produces measurable ROI.

Will llms.txt become mandatory in 2027?

No signal in that direction. The standard is not validated by W3C nor adopted by OpenAI, Google, Meta or Mistral as a citation signal. Lighthouse 13.3 integration does not foreshadow a SEO Search obligation. The situation may evolve if the agentic web takes off, but nothing guarantees it today.

Sources

• SE Ranking, “LLMs.txt: Why Brands Rely On It and Why It Doesn’t Work” (November 2025) — study on 300,000 domains. seranking.com/blog/llms-txt
• OtterlyAI, “Llms.txt Experiment: What Marketers Get Wrong about llms.txt” (Thomas Peham, February 5, 2026) — analysis of 62,100 AI bot visits over 90 days. otterly.ai/blog/the-llms-txt-experiment
• ALLMO, “LLMs.txt for AI Search Report 2026” (January 2026) — 94,000 cited URLs analyzed. allmo.ai/articles/llms-txt
• Abondance, “Le fichier llms.txt ne sert à rien, mais Lighthouse le vérifie quand même” (Johan Sellitto, May 22, 2026) — John Mueller’s position. abondance.com
• Chrome for Developers, “Lighthouse Agentic Browsing scoring” (updated May 5, 2026). developer.chrome.com
• Search Engine Journal, “LLMs.txt Shows No Clear Effect On AI Citations, Based On 300k Domains” (Matt G. Southern, November 2025). searchenginejournal.com
• llmstxt.org — original specification by Jeremy Howard, Answer.AI (September 2024). llmstxt.org