What Is /llms.txt and What Should Your Store Put in It?
A founder-voice guide to the llms.txt standard for ecommerce: what it is, what to include, where to host it, and how to keep it useful for AI assistants.
If you have shipped Schema.org Product markup, an OpenAI Commerce feed, and a clean direct answer in the first 150 words of every PDP and your brand still gets ignored when a buyer asks ChatGPT for a category recommendation, the gap is often discovery. The AI assistant does not know your site exists in a usable form, or it cannot find the right resources fast enough to cite you. The fix is a small file at the root of your domain called llms.txt.
What is /llms.txt and what should your store put in it?
/llms.txt is a markdown file at the root of a website that AI assistants read to get oriented to the brand, products, and key resources without having to crawl every page. The llms.txt proposal, authored by Jeremy Howard at Answer.AI, defines the standard: a single markdown file at /llms.txt that opens with the brand name and a short description, then links to the most important resources on the site in a structure the model can read in one fetch.
For ecommerce, the file should include: the brand pitch in 1 to 2 sentences, the ICP (who the store is for), pricing tiers if relevant, links to article clusters and key marketing pages, programmatic access points (sitemap, JSON feeds, Agent API endpoints), and a contact channel. Total length is usually 1 to 2 pages of markdown. The file is text-only, no HTML, no images, no styling.
Why /llms.txt matters for AI shopping
AI assistants do not crawl websites the way Google does. They make targeted requests when a buyer asks a question, often with limited time and a small token budget for retrieval. A brand site with 200 product pages, 40 articles, and a complex navigation menu is a needle-in-haystack problem for the assistant. The model fetches the homepage, sometimes a sitemap, maybe one or two indexed pages, and then has to compose an answer.
/llms.txt collapses that into one file. The assistant fetches /llms.txt, reads the brand pitch and the ICP, sees a list of clustered article links, finds the merchant feed and Agent API, and can compose a useful answer without crawling further. Brands shipping a clean /llms.txt get cited more reliably because the model has less to figure out.
This matters more in 2026 than it did a year ago. ChatGPT Shopping, Perplexity Shopping, and Claude all prefer brands with discoverable structure. A site that ships /llms.txt plus a clean sitemap plus Schema.org markup signals that the brand has thought about AI consumption. A site without those signals is harder to recommend.
What an ecommerce /llms.txt should contain
The structure that works for premium DTC is six sections.
Section 1: Brand intro. Two sentences. What the brand sells, who for. Skip the marketing voice; the AI is reading for clarity, not personality. "Brand X sells premium leather backpacks for daily commute at $300 to $600 average order. The brand ships globally and operates as a direct-to-consumer business since 2018."
Section 2: What the brand does. A short bulleted list of capabilities the AI should know about. Materials, build quality, manufacturing location, sustainability claims, warranty terms. Each bullet is one fact, not a paragraph.
Section 3: Who the brand is for. The ICP in 2 to 4 lines. Premium DTC brands tend to have specific buyer profiles; spell them out. "Daily-commute professionals at $80,000+ annual income, often in finance, design, or tech. Buyers value materials over brand logos and prefer mid-tier price points to luxury markups."
Section 4: Pricing. If the brand has clear tiers (subscription, plans, bundles), list them with prices and what each unlocks. For single-product DTC, list the average order range and any subscription option.
Section 5: Key resources. Links to article clusters, product feed URLs, Agent API base URL, sitemap, and any other programmatic access points. Group by topic so the AI can pick the right link without parsing the whole list.
Section 6: Contact. Where to reach the brand. Email, a contact form URL, or a support channel. Skip social media unless the brand wants to be cited there.
The full file usually fits in 1 to 2 pages of markdown. Longer is rarely better; the AI is looking for orientation, not depth.
Where to host the file
/llms.txt lives at the root of the domain. For most stores that means https://www.yourbrand.com/llms.txt, served as plain text with Content-Type: text/markdown or text/plain. Subdirectory locations or alternate filenames will not be discovered.
For Next.js or Hydrogen stores, the file goes in the public/ directory and ships with every build. For Shopify themes without static-asset support, the file usually ships as a custom page route or through the brand's CDN.
Verify the file is reachable with curl https://your-store.com/llms.txt. The response should be the markdown content as plain text, not an HTML page wrapping the markdown.
What llms.txt does NOT replace
Three signals that /llms.txt does not replace and brands sometimes confuse:
Schema.org markup. The structured data on individual product pages still does the heavy lifting for AI extraction at the page level. /llms.txt is site-level orientation; Schema.org markup is page-level data. Both should exist. The product schema markup guide covers what to ship per page.
robots.txt. The crawler-control file at /robots.txt tells AI crawlers which paths they may fetch. /llms.txt is content; /robots.txt is policy. Both files are useful and they serve different roles.
Sitemap.xml. The sitemap lists every indexable URL on the site. /llms.txt curates the most important resources for AI orientation. Sitemap is exhaustive; llms.txt is selective.
The OpenAI Commerce feed. The merchant feed at OpenAI Commerce ships product data for ChatGPT Shopping discovery. /llms.txt does not include product-level data; it points to the feed and to the article and marketing pages. The agent-readable feed guide covers the feed side.
How to keep the file useful
A stale /llms.txt is worse than no file. The most common failure modes:
- Outdated pricing. When tiers change, update the file before the next billing cycle. Brands that update the pricing page but forget llms.txt give the AI conflicting signal.
- Dead links. When article slugs change or pages are deprecated, prune the llms.txt links. AI assistants drop brands whose key resources return 404.
- Marketing drift. The brand pitch in section 1 should match the homepage and the about page. When marketing rewrites the homepage, also update llms.txt.
The simplest pattern is to generate /llms.txt from the same source of truth that powers the marketing site. If the article index is generated at build time, the llms.txt link section can come from the same data.
What to expect after shipping
Brands that ship a clean /llms.txt typically see two effects within 2 to 4 weeks. AI assistants cite the brand more accurately when buyers ask about the category, because the model has the right context in one fetch. Buyers who are doing AI-assisted research arrive at the site already pre-qualified, so the conversion rate from those sessions tends to run higher than average.
Neither effect shows up in Google Analytics directly. Track it instead by asking ChatGPT, Claude, and Perplexity the category question your brand should rank for, and watch how often the response cites your brand and how accurately it describes you.
Where to start
If you have not shipped /llms.txt yet, the work is roughly an afternoon. Write the six sections above, save them as a markdown file, deploy to the root of the domain, verify with curl. The AEO playbook for ecommerce covers where llms.txt fits in the broader signal stack.
If you already ship Schema.org markup and a merchant feed but the brand is still invisible in AI search, llms.txt is the most likely missing piece. The model has the page-level data and the feed-level data, but no site-level orientation. The file is small. The fix is fast.
The buyer is asking the assistant. The assistant has 30 seconds and one fetch budget for orientation. The brand whose /llms.txt answers the question first is the brand cited.