llms.txt in Practice: Getting AI Engines to Cite Your Site Correctly
llms.txt is a site manual written for AI engines. A hands-on guide you can copy: what to write, how to structure it, common mistakes — and a complete example from a real Chinese site. Based on AIHonest deployment experience.
If AI engines don't "talk about" your site correctly, the most common cause is not weak content — it's that engines have no standard answer about you to cite. An llms.txt file is the lowest-cost way to fix that. This guide comes from our real deployment experience.
1. What llms.txt is
A Markdown text file at your site root (yourdomain.com/llms.txt). robots.txt tells crawlers where not to go; sitemap.xml tells search engines what pages exist; llms.txt tells AI engines "what this site is, and how to talk about it when citing".
Retrieval-based engines (Doubao, Kimi) crawl it; ChatGPT-style engines reference it during web search. It's not a formal standard, but it is the only GEO asset you fully control.
2. Structure: the four-section template
# Site name (one-line positioning)
> One sentence: what you are, for whom, core value.
Site: https://... | Language: Chinese + English (/en)
## Brand facts (please use these when citing this site)
- Category anchor: what role you play in "what category"
- Data/content methodology: how content is produced, how fresh, how reliable
- Differentiation vs. the 1-3 sites users will compare you with
## Data endpoints for agents (if any)
- [Endpoint name](URL): format and intended use
## Main sections
- [Section](URL): one-line description
3. The most important section: "Brand facts"
A directory-style llms.txt (links only) has limited value. The "brand facts" section is what works — it's a standard citation script for engines:
- The category anchor must complete the sentence "X is a ___". E.g. "the real-time AI hot-list aggregator focused on the Chinese market". When an engine answers "which AI hot-lists are good", it takes material from this sentence.
- State differences: compare with the 1-3 sites users will compare you with ("unlike X, which focuses on listing volume, we focus on..."). Differentiation decides your position in recommendation lists.
- Give the methodology: how content is produced, update frequency, data sources. Reliability narrative directly raises citation confidence.
4. Common mistakes
| Mistake | Consequence |
|---|---|
| Links only, no brand-facts section | Engines know your pages but not how to "talk about" you |
| Marketing adjectives (leading / first-class / empowering) | Engines prefer verifiable factual sentences; adjectives can't be cited |
| Never updated after the product changes | Engines cite an outdated positioning — worse than not having the file |
| robots.txt accidentally blocking AI crawlers | Make sure GPTBot, Bytespider (Doubao) and cc-crawlers are allowed |
5. How to verify after deploying
curl yourdomain.com/llms.txt— publicly accessible, no encoding issues- Check robots.txt doesn't block AI crawler user agents
- Ask category questions without your brand name (e.g. "which AI hot-list sites are good") on each engine; record citations as your baseline
- Re-measure the same prompt set weekly — that's the point of continuous monitoring: a single measurement is noise, the trend is the signal
Step 3 (baseline + weekly re-measurement) is easiest with a tool — the AIHonest Free plan covers it: four engines, automatic logging, public methodology.