llms.txt is an emerging convention: a plain-text (often Markdown-flavored) file at your site root that lists canonical pages worth reading for AI systems and agentic browsers. Think of it as a librarian’s shortlist, not a sitemap dump of every tag page you ever shipped.
It complements robots.txt and XML sitemaps. Robots still govern crawling permission. Sitemaps list URLs at scale. llms.txt says: “If you only read a handful of documents about us, start here.”
Why B2B sites benefit
B2B sites accumulate pricing PDFs, outdated landing clones, and thin location pages. Models that sample randomly may quote an old package or a thin ngành template. A curated llms.txt points agents at your current pricing philosophy, service pillars, and policy pages.
Example themes to include: company overview, core services, GEO/SEO stance, outsourcing FAQ, security/contact, and one or two deep technical articles. Exclude thank-you pages, parameter URLs, and experiments.
How to write a useful llms.txt
- Start with a one-paragraph company summary in plain English.
- List absolute HTTPS URLs with short annotations.
- Prefer apex canonicals (
https://vestadesign.com.vn/...) — avoid www duplicates. - Update when pricing philosophy or service lines change; show a last-updated date in your process docs.
- Keep it short enough that a human would actually maintain it.
Common mistakes
- Pasting the entire XML sitemap into llms.txt.
- Linking to noindexed or staging hosts.
- Contradicting on-page pricing.
- Treating llms.txt as a substitute for real FAQ and schema work.
- Promising that “llms.txt gets you into ChatGPT” — no public standard guarantees that.
How it fits a GEO program
GEO still needs entity consistency, answer-shaped pages, and technical SEO. llms.txt is a helper file. On vestadesign.com.vn we publish /llms.txt alongside Vietnamese and English service pillars. Read more on our GEO page and the Vietnamese explainer ecosystem under SEO & GEO.
Implementation notes
Serve text/plain (or markdown as plain text). Ensure CDN/cache does not forever pin a stale file after you change packages. Reference only pages that return 200 with a single clear H1. After publishing, spot-check with curl and with an AI assistant by asking factual questions about your company.
Need a review of your B2B information architecture for AI search? Contact info@vestadesign.com.vn.
Example annotation style
Good line: “https://example.com/pricing — Current package philosophy and what is excluded (ads spend, third-party licenses).” Bad line: “https://example.com/page?id=22 — click here!!!”. Annotations should remove ambiguity a model might otherwise hallucinate. If a page is locale-specific, say so. If a page is Vietnamese-only, label language so English agents do not mis-cite.
When you retire a service, remove it from llms.txt in the same release as the on-page change. Stale machine hints are worse than no hints. Store the file in git next to your sitemap generation script so marketing and engineering share ownership.
Finally, remember optional conventions evolve. Re-read community guidance yearly, but do not delay shipping clear human content while waiting for a perfect standard.
Security and privacy angle
Do not list internal runbooks, staging credentials pages, or customer-only portals in llms.txt. The file is public. Treat it like a press kit for machines: helpful, bounded, non-confidential. If you use bot management, make sure legitimate research agents can still read the URLs you listed — otherwise you published a map to locked doors.
Related: GEO agency · Outsource to Vietnam · MVP development · Technical SEO · Paid ads.
Vesta Design — Ho Chi Minh City · +84 356 413 333 · info@vestadesign.com.vn · Clutch profile.