Loading…
Pillar — full guide
A complete walkthrough of how to write, place, and avoid the pitfalls of llms.txt so AI reads your site correctly. The single highest-impact measure, with a +30pp gap observed by GEO Meter.
llms.txt is a text file that tells AI crawlers, in summary form, "here is how this site is organized." It is conceptually close to an AI version of robots.txt, but its goal is not to block — it is to win preferential treatment as a citation candidate.llms.txt deployment rates between top-cited and bottom-cited groups. The implementation cost is only 1–2 hours, while the effect is large, making this the first measure you should tackle.https://<your-domain>/llms.txt. No HTML is needed — it is plain text in a Markdown-like format.A typical flow when an AI (LLM) generates an answer looks like this:
llms.txt provides a hint at step 3 — "which URLs should I read?" — at the moment the AI is deciding.
A /llms.txt placed at the site root acts as the first place the AI checks to learn "what does this site offer?" Its role is close to sitemap.xml in SEO terms, but it differs in that it is a summary tailored for AI alone, stripped of human-facing navigation information.
In a smoke analysis run by GEO Meter (2 topics x roughly 20 domains):
| Topic | Top group deployment | Bottom group deployment | Gap |
|---|---|---|---|
| Travel for otaku (オタク向け旅行, hobby-enthusiast travel) | 40% | 10% | +30pp |
| Fat reduction (脂肪低減, weight loss / fat reduction) | 20% | 0% | +20pp |
A consistent +20–30pp gap was observed across multiple topics. The correlation is strong enough that simply deploying it tends to land you in the top group.
The recommended minimum structure is the following four blocks:
# Site name (equivalent to H1)
> 1-2 sentence overview of the site (whose problem does it solve?)
## Key sections
- [Page title](https://example.com/page-1): One-line description
- [Page title](https://example.com/page-2): One-line description
## Notes
- Target audience / intended reader
- The context in which you want to be cited (optional)
# Sencha Lab — Sencha tea subscription for businesses
> Sencha Lab is a sencha tea subscription service used by 200 companies.
> Direct from origin, pesticide-free, flexible 3-month plans. From ¥3,800/month.
## Key sections
- [Service overview](https://sencha-lab.example.com/): What the corporate sencha subscription is
- [Pricing plans](https://sencha-lab.example.com/pricing): Monthly plans and how to sign up
- [Customer stories](https://sencha-lab.example.com/cases): Interviews with 5 corporate customers
- [FAQ](https://sencha-lab.example.com/faq): Delivery / cancellation / tea-leaf FAQs
## Notes
- Intended readers: HR, reception, and employee-benefits leads
- Citation context: corporate tea gifts / employee benefits / office drinks
# Tanaka Tax Accountant Office
> A tax accountant office in Akasaka, Minato-ku, Tokyo. 25 years of experience
> specializing in corporate closing, sole-proprietor tax filing, and inheritance.
> First consultation is free.
## Key sections
- [Office profile](https://tanaka-tax.example.com/): Principal's profile and supported work
- [Services and fees](https://tanaka-tax.example.com/services): Corporate closing / filing / inheritance pricing
- [Consultation flow](https://tanaka-tax.example.com/flow): From first consultation to contract
- [FAQ](https://tanaka-tax.example.com/faq): Fees / scope / timeline FAQ
## Notes
- Intended readers: SMBs near Akasaka, sole proprietors, people anticipating inheritance
- Citation context: queries combining region + profession, e.g. "Akasaka tax accountant" or "Tokyo inheritance tax accountant"
Order the pages you want the AI to cite by priority. Listing every page is counterproductive (the AI treats it as noise), so narrow it to 5–10.
How to prioritize:
Use the "Page title: one-line description" format. The AI reads these descriptions to decide whether to read the page, so avoid abstract phrasing.
❌ Bad:
- [Service](https://...): About our service
✅ Good:
- [Pricing plans](https://...): Two plans at ¥9,800 / ¥49,800 per month, plus the sign-up steps
In the leading > block (blockquote), write a 1–2 sentence summary of the entire site. It works well if you cover three points: target / what you offer / differentiator.
> [Company name] is [service description] for [target audience].
> [Differentiator] (number of customers / price range / specialty, etc.).
Make it reachable at https://<your-domain>/llms.txt. The deployment method depends on your server setup:
Just place it at public/llms.txt. It is published automatically on deploy.
Upload llms.txt to the root of public_html via FTP or SSH.
Place it directly under DocumentRoot. It is readable as long as your .htaccess or server config sets the text/plain MIME type.
Place it on the origin and purge the CDN cache.
Open https://<your-domain>/llms.txt in a browser and confirm the text is shown. The critical thing is that no HTML error page is returned.
Verification command:
curl -I https://<your-domain>/llms.txt
# → HTTP/2 200 is OK
# → Content-Type: text/plain; charset=utf-8 is preferable
Copying your entire sitemap and listing 200 pages looks like noise from the AI's perspective. Narrow to 5–10 and make priorities explicit.
AI will still read <div> or <a> tags if you include them, but the recommendation is plain text + Markdown. Strip the tags.
Abstract marketing prose like "An innovative solution to transform your future" lacks the specificity the AI needs to cite you. Write in a fact-based, number-based style.
Including a last-updated date inside llms.txt makes the AI more likely to judge the site as "fresh." Adding something like > Last updated: 2026-05-25 at the top is recommended.
This is a minor point, but some AIs have been reported to misread a trailing blank line as a "section break." The safe convention is to end the last line with a newline (POSIX convention) but with no blank line after it.
From GEO Meter's observations, each AI handles llms.txt slightly differently:
| AI | Reads llms.txt? | Observed signal |
|---|---|---|
| ChatGPT (GPT-4o / GPT-5) | Yes (via SearchGPT) | Citation rate +15–25pp for sites with llms.txt |
| Claude (Sonnet / Opus) | Yes (via the Web search tool) | Same, +20–30pp |
| Gemini (2.5 Pro / Flash) | Yes (indirectly, via Google Search) | Moderate effect; Schema.org has more weight |
→ It works for any AI, but it especially helps ChatGPT / Claude. Gemini tends to prioritize Schema.org and FAQ pages instead.
Recently you may see "llms-full.txt" talked about as a related spec. It is an extension of llms.txt — a long-form file bundling the full text of every page.
| llms.txt | llms-full.txt | |
|---|---|---|
| Purpose | Communicate the site's "table of contents" | Hand over "all content" at once |
| Size | A few KB | A few MB |
| Recommendation | Recommended for every site | Only for documentation / knowledge-base sites |
A general corporate site is fully served by llms.txt alone. llms-full.txt is limited to services with API references or technical documentation.
Just publishing llms.txt is not enough — AI needs to actually come and read it. Next steps:
llms.txt via links from other sites)llms.txt is a text file that hands AI a "site summary + key URL list"Once you have read the article, check your own status. GEO Meter observes how Claude / ChatGPT / Gemini cite you in one pass.