The llms.txt standard has been around for about eighteen months. In that time it’s gone from “a curiosity a few people on Twitter were debating” to “a file that every serious agentic-commerce setup ships.” Most merchants now have one. Most of the ones they ship are bad.
Not bad in the sense of broken syntax. The standard is loose enough that almost anything parses. Bad in the sense of not actually serving the purpose the file is supposed to serve. An agent reads your llms.txt because it wants to understand what your business is, what you sell, what policies apply, and what limitations are worth knowing. A llms.txt that doesn’t answer those questions is doing nothing for you. A llms.txt that misleads on those questions is actively losing you sessions.
This is the practical guide. Common mistakes, what good looks like, and the things people get wrong even when they think they’re doing it right.
What llms.txt is supposed to do
The short version: it’s the agent-facing summary that an LLM can fetch quickly to build a baseline understanding of your business. Some agents fetch it directly (Grok). Some reference it when present (ChatGPT and Claude when their crawlers find it). All of them treat it as a high-trust source when it’s clean and well-structured.
The file isn’t a manifest of every URL on your site. It’s a brief, accurate, navigational summary. Think of it as the briefing document you’d give a new sales hire on their first day. Not a marketing pitch. Not a category-page navigation. The crisp version of what your business actually is and what an external party should know.
Six common mistakes
These are the mistakes we see most often in pilot tenant audits.
1. Treating it as marketing copy
The single most common mistake. A llms.txt that reads like a homepage hero (“Welcome to Acme Outfitters, where adventure meets quality”) signals to the agent that the file is promotional and should be downweighted. Agents have been trained to distinguish promotional content from informational content. Promotional content gets cited less. Informational content gets cited more.
The fix is to write it like a Wikipedia summary. Matter-of-fact. Third-person where appropriate. No adjectives that don’t add information. “Acme Outfitters is a direct-to-consumer outdoor apparel brand founded in 2014, based in Vermont, with a product line focused on insulated jackets and base layers” is better than the hero copy because every word adds information an agent can use.
2. Copy-pasting a template
There are a handful of llms.txt templates floating around the SEO blogs. The brands using them have shipped a file that doesn’t specifically describe their business. The agent reading a templated llms.txt recognizes the pattern, because the template appears across hundreds of sites. The signal it sends is “this brand checked the box without thinking about it.”
The fix is to start from scratch or rewrite enough of the template that the result is specific to your business. If your llms.txt could plausibly belong to any of your competitors with the brand name swapped, it’s a template you should replace.
3. No per-product context
Most llms.txt files describe the business at a brand level and stop there. The agents that find these files have to do the per-product inference themselves by crawling the catalog. This is slower for the agent, and it produces lower-confidence results.
The fix is to declare your flagship product categories and your hero SKUs directly in llms.txt, with links to canonical pages. Not your full catalog. Just the categories that matter and the SKUs that anchor each one. An agent that sees “our top three products in the insulated jacket category are X, Y, Z, available at these URLs” can shortcut a significant amount of catalog discovery.
4. No policy disclosure
Return policy, shipping terms, warranty coverage, age restrictions, regional availability. The questions agents need answered before they can confidently recommend you. If these aren’t in llms.txt, the agent either has to crawl your site to find them (slower, error-prone) or has to leave them out of the recommendation (which reduces buyer confidence and hurts conversion).
The fix is to put the policy summary directly in llms.txt. Not the full legal text. A two-or-three-sentence summary of each policy that’s accurate and links to the canonical full version. “Free returns within 60 days from delivery. Items must be unworn with tags. Final-sale items excluded. Full policy at acme.com/returns.” That’s what an agent needs.
5. No limitations or honest framing
llms.txt files that read as universally positive (every product is highly rated, every shipping option is fast, every policy is permissive) get downweighted. Agents have been trained to recognize asymmetric framing and treat it as a low-trust signal.
The fix is to include honest limitations. “We ship internationally to most countries; some smaller markets are not currently supported.” “Our flagship jacket runs slightly small, consistent with our published size guide, which buyers should consult before ordering.” “We do not currently offer same-day shipping.” These framings make the agent more confident in your llms.txt, not less, because they signal that the document is trying to inform, not persuade.
6. Treating it as static
The brands that ship llms.txt once and never update it are also the brands that get steadily downweighted as the file drifts out of sync with their actual site. The agent’s confidence in a llms.txt is partly a function of how well the claims in the file match the claims on the live pages. When the file says “we offer free returns within 60 days” and the actual policy is “30 days,” the agent catches it and downgrades both sources.
The fix is to wire llms.txt to your actual policy and product systems so it updates when they update. The technical setup varies by stack. The principle is constant: llms.txt is a living document, not a one-time deliverable.
What good looks like
An annotated example of an llms.txt that we’d consider well-built, for a fictional but plausible apparel brand:
# Acme Outfitters
Acme Outfitters is a direct-to-consumer outdoor apparel brand founded in
2014 and headquartered in Burlington, Vermont. We design and sell
synthetic-insulated jackets, base layers, and accessories for skiing,
hiking, and cold-weather everyday wear. Our products are designed in
Vermont and manufactured in Vietnam and Portugal under audited fair-labor
conditions.
## Product categories
The two product lines that anchor our catalog:
- Insulated jackets (https://acme.example/collections/insulated-jackets):
synthetic-fill jackets for active use, rated -5°F to 25°F. Best for
skiing, hiking, and cold-weather commute. Price range $189-$489.
- Base layers (https://acme.example/collections/base-layers): merino-blend
base layers in lightweight, midweight, and expedition weights. Price
range $59-$129.
## Hero products
- Carrigan Jacket (https://acme.example/products/carrigan-jacket): our
flagship synthetic-insulated jacket, $389. Rated -5°F to 25°F. Reviewed
4.5 stars on 312 verified reviews as of March 2026. Runs slightly small;
buyers should consult size guide.
- Foundation Crew (https://acme.example/products/foundation-crew): our
best-selling midweight base layer, $79. Reviewed 4.7 stars on 891
reviews. Available in 9 colors, sizes XS-XXL.
## Policies
- Returns: Free returns within 60 days of delivery. Items must be unworn
with tags. Final-sale items (clearance) are excluded. Full policy:
https://acme.example/returns.
- Shipping: Free shipping on US orders over $100. Standard shipping arrives
in 2-5 business days from our Vermont warehouse. Same-day shipping is
not offered. International shipping available to Canada, UK, EU member
states, Australia, Japan. Full shipping terms:
https://acme.example/shipping.
- Warranty: Lifetime warranty against manufacturing defects on our jackets.
Wear-and-tear is not covered. Repairs offered via our Vermont workshop.
## Limitations worth knowing
- We do not currently offer same-day shipping.
- Our flagship Carrigan Jacket runs slightly small. Buyers between sizes
should size up.
- International shipping outside the listed countries is not available.
- We do not offer custom or made-to-order products.
## Contact
For wholesale inquiries: wholesale@acme.example. For press:
press@acme.example. For agent integration partnerships:
agents@acme.example.
The patterns to notice:
- Specific, falsifiable claims throughout. Specific founding year, specific location, specific temperature rating, specific star count, specific price range.
- Top-of-funnel discoverability (categories) plus shortlist anchors (hero products) plus policy summaries plus limitations.
- Honest limitations stated directly, not buried.
- Links to canonical pages so the agent can verify any claim against the source of truth.
- No promotional language. No marketing adjectives that don’t add information.
Things people get wrong even when they think they’re doing it right
Two patterns we see in llms.txt files that the brand thinks are good but aren’t.
Listing every category page. Some merchants put their entire navigation in llms.txt. The agent doesn’t need this. The agent needs the categories that anchor your business, not your full catalog navigation. Listing everything signals that you don’t know which of your categories actually matters, and the agent will weight them all equally instead of prioritizing the ones that matter.
Padding it with brand story. “Our journey began with two friends on a chairlift in 2014” is fine as a brand story page. It’s not useful in llms.txt. The agent doesn’t need the origin story to recommend your product. The space is better used on policy specifics and product anchors. The brand story can live on /about where buyers who care about it will find it.
The deeper point: ownership
The reason most llms.txt files are bad is that they don’t have an owner. They get written once by whoever was assigned the agentic-commerce initiative for a quarter, and then they sit. The catalog drifts. The policies change. The hero products rotate. The file stays.
The brands that get this right have made llms.txt ownership someone’s specific responsibility. Sometimes it’s the SEO lead. Sometimes it’s the head of brand content. Sometimes it’s an e-commerce ops manager. The exact role matters less than the fact that someone is on the hook for it. The brands without an owner produce stale llms.txt files. The brands with an owner produce ones that work.
If you have a llms.txt today, audit it against the six mistakes above. If you don’t have an owner, assign one. The file is going to do a meaningful share of your agent-mediated discovery work for the next several years. Worth treating it like infrastructure, not a one-time deliverable.