llms.txt for Ecommerce: The Complete Guide

Every few weeks a merchant asks us whether they need an llms.txt file, usually because a competitor added one or an agency pitched it as the secret to getting recommended by ChatGPT. The honest answer is more boring than the pitch: llms.txt is a one-page markdown map for AI systems, it costs about an hour to write, it is cheap insurance rather than a ranking hack, and most stores that have one wrote it badly.

This guide covers what the file actually is, how it differs from robots.txt and sitemap.xml, exactly what an ecommerce store should put in it, how to keep it from going stale, and — because this matters more than anything else — what it does not do.

What is llms.txt?

llms.txt is a plain markdown file served at the root of your domain — yourstore.com/llms.txt — that gives AI systems a short, human-written orientation to your site: who you are, what you sell, and where the canonical pages live. It was proposed as an open convention in late 2024, modeled loosely on robots.txt but with the opposite job: instead of telling crawlers what to avoid, it tells them what to read.

llms.txt

A markdown file at /llms.txt containing a curated map of a site for AI consumption: an H1 with the site name, a one-paragraph summary in a blockquote, then H2 sections of annotated links to the pages that matter most. Voluntary convention — no assistant is obligated to fetch or honor it.

The format is deliberately simple. One # heading with your store name. One blockquote (>) with a two-or-three-sentence summary of what you sell and who you serve. Then ## sections — categories, products, policies, guides — each containing a short list of markdown links, one per line, each with a one-line description after the URL. That is the whole specification. If your llms.txt takes more than a page to scroll, you have misunderstood the assignment.

llms-full.txt

A companion convention: a single file containing the full flattened text content of the mapped pages, so an AI system can ingest everything in one fetch. Useful for documentation sites; rarely worth maintaining for an ecommerce catalog that changes weekly.

The reasoning behind the convention is sound. When an assistant like ChatGPT, Claude, or Perplexity fetches your site to answer a shopper's question, it lands on pages built for human browsing — navigation chrome, banners, popups, scripts. A curated markdown index gives the machine a clean starting point: here is what this store is, here are the twenty pages that answer most questions. Whether a given assistant actually uses it is a separate question, which we get to below.

llms.txt vs robots.txt vs sitemap.xml

These three files get conflated constantly, and the confusion leads merchants to write llms.txt files that read like sitemaps — hundreds of bare URLs with no annotation, which defeats the purpose. The three files answer three different questions.

The dependency runs one way: llms.txt assumes you have already decided to let AI crawlers in. If your robots.txt blocks GPTBot, ClaudeBot, and PerplexityBot, an llms.txt file is a welcome mat behind a locked door. Sort out your crawler access policy first — we walk through that decision in should you block AI crawlers — then curate the map.

And to be precise about sitemap.xml: keep it. llms.txt does not replace it. Search engines still want the exhaustive inventory, and AI crawlers use sitemaps too. The two files coexist — one is the phone book, the other is the list of numbers you would actually give a guest.

Does your store need llms.txt?

Short answer: probably yes, because the cost is close to zero — but with correctly-sized expectations, and not first. Here is the honest state of things as of mid-2026: llms.txt is a voluntary convention, no major assistant has publicly committed to it as a retrieval or ranking input, and adoption among AI crawlers is uneven. Some fetch it. You can check your own server logs for requests to /llms.txt and see exactly which bots in your traffic do.

So the case for writing one is not "assistants will rank you higher." The case is: it takes about an hour, it costs nothing to serve, the downside is nil if you maintain it, and if any assistant in your traffic does consult it, you have handed that assistant your best-foot-forward summary instead of letting it assemble one from whatever pages it happened to crawl. That is a good trade. It is just not a strategy.

What it should never be is your first move. If your product pages lack structured data and your catalog attributes are half-empty, an llms.txt file is a beautifully annotated map to pages that cannot answer questions. Sequence the work: schema and catalog first, llms.txt as the finishing layer. In our Commerce GEO engagements it is one of the last deliverables, not the first.

What goes in an ecommerce llms.txt

Most published examples of llms.txt come from software documentation sites, so ecommerce merchants copying them end up with the wrong shape. A store's llms.txt has five sections, in this order:

1. Identity block

The # heading and blockquote. Name the store, state what you sell, who it is for, and what you are known for — in the same plain terms a shopper would use. > Acme Grip makes weightlifting accessories for home-gym athletes: lifting straps, chalk, and belts, shipped from the US, with a lifetime warranty on stitched goods. An assistant that reads only this line should be able to categorize you correctly. Skip the brand poetry; this is a fact sheet, not a manifesto.

2. Category map

Your top five to fifteen collections, each as a link with a one-line description of what is in it and who it is for. Not every collection — the ones a shopper would name. If you sell coffee gear, that is brewers, grinders, kettles, and accessories, not your "Summer Refresh Edit." The annotation matters: a line like Grinders: burr grinders for espresso and filter, $80-$600 linked to the collection tells a machine far more than a bare URL.

3. Hero products

Ten to thirty products maximum — your best sellers and the SKUs you most want an assistant to know exist. Each entry gets a direct link and one differentiating fact: the thing that would make an assistant pick it for the right shopper. These should be the same hero SKUs you enriched first in your catalog enrichment pass — the map and the territory should agree on what matters.

4. Policies and logistics

Shipping regions and times, return window, warranty, sizing guidance. This section earns its place because policy questions are among the most common things shoppers ask assistants — "can I return it," "do they ship to Canada," "is there a warranty" — and the answers change purchase decisions. Link the canonical policy pages, and put the one-line answer in the annotation itself: Returns: 60-day free returns, US and Canada.

5. Guides and evidence

Buying guides, comparison pages, care instructions, FAQ pages — the content that answers pre-purchase questions. These pages tend to be exactly what assistants quote when they cite a merchant, so pointing at them directly is the highest-leverage part of the file. If you have nothing to put here, that is a finding: it means you have no citable content, and no map fixes that.

How to write and ship it

1. 1

   Draft the identity block

   One heading, one blockquote, three sentences. Write it like you are briefing a smart stranger who will answer questions about your store all day. Test: could someone categorize your store from this paragraph alone?

2. 2

   Pick the categories

   Five to fifteen collections a shopper would name unprompted. One line each: what is in it, roughly what it costs, who it is for.

3. 3

   Pick the hero products

   Ten to thirty SKUs. Best sellers, flagship items, and anything with a clear "best for X" position. One differentiating fact per line — the fact an assistant would repeat.

4. 4

   Add policies and guides

   Link the canonical shipping, returns, warranty, and sizing pages with the one-line answer inline. Then list your buying guides and comparison content.

5. 5

   Publish at the root and verify

   Serve the file at /llms.txt with a text content type. Fetch it yourself, click every link, and check your logs over the following weeks to see which crawlers request it.

Maintenance discipline

Here is the failure mode we see most: a merchant ships a decent llms.txt in a burst of GEO enthusiasm, then never touches it again. Nine months later it points at three discontinued products, a returns policy that changed, and a collection URL that 404s after a navigation redesign. A stale map is worse than no map — you are actively feeding wrong answers to any system that trusts the file.

• Review quarterly, minimum. Put it on the same calendar as your merchandising refresh. The review takes fifteen minutes: click every link, check every claim.
• Update on trigger events. Policy changes, hero-product swaps, category restructures, and URL migrations all invalidate lines in the file. Add llms.txt to the checklist for each.
• Keep it curated, not generated. Resist the app that auto-dumps your full catalog into llms.txt. Three hundred lines of unannotated product URLs is a sitemap with a markdown extension — the editorial judgment is the entire value.
• Keep it short. If the file grows past what a person would read in two minutes, cut it back. Curation means saying no.
• Assign an owner. Files without owners rot. Whoever owns your PDP copy or your merchandising calendar owns this file too.

What llms.txt does and does not do

Setting expectations precisely is the difference between a sensible hour of work and a disappointment. What it does: gives any AI system that consults it a clean, accurate, merchant-authored summary of your store; concentrates crawler attention on your best pages; costs almost nothing to create and maintain; and signals baseline AI-readiness. What it does not do: guarantee that any assistant reads it, improve rankings in AI answers on its own, fix weak product pages, or substitute for structured data and catalog work.

The mechanics of actually getting recommended — being retrievable, being parseable, having evidence an assistant can cite — are a bigger system, and we have mapped it in how products get recommended by AI. llms.txt is one small, cheap component of that system: the courteous gesture of handing the machine a well-organized brochure instead of making it wander the store. Make the gesture. Just do not confuse it with the store.