What is llms.txt?
llms.txt is a Markdown file at the root of a site that lists the pages an AI system should read, with one line on each. Jeremy Howard proposed it on 3 September 2024 at llmstxt.org. The only required part is an H1 with the site name, followed by an optional summary and lists of links.
A shop’s file can be this short, shown here for an invented bike shop:
# Velo Atelier
> Bike shop in Lausanne. Road, gravel and e-bikes, delivered across Switzerland.
## Policies
- [Delivery](https://example.ch/delivery.md): carriers, costs and delivery times in Switzerland
- [Returns](https://example.ch/returns.md): 30-day returns, conditions and refunds
## Buying guides
- [Gravel bike sizes](https://example.ch/guides/gravel-sizes.md): frame size by rider heightOurs is at /llms.txt, with the full text in /llms-full.txt. llms-full.txt is a community convention and not part of the proposal.
Does Google use llms.txt?
No. Google’s guide to generative AI in Search, updated on 10 July 2026, says you do not need machine-readable files, AI text files or Markdown to appear in Search, that it is “completely fine” to keep an llms.txt file, and that Search ignores it (Google Search Central).
Google’s page on AI features says the same for AI Overviews and AI Mode: “There are no additional requirements”, and a page must be indexed and eligible for a snippet (Google Search Central, updated 10 December 2025). So llms.txt does nothing for Google rankings, and a shop should never buy it as a Google tactic. The GEO vs SEO guide covers what does count for Google’s AI answers.
Who does read llms.txt?
Tools and agents that fetch it on purpose. Chrome’s Lighthouse checks the file in its agentic browsing audits, and flags it only when the server returns an error; a missing file is marked not applicable (Chrome for Developers, 5 May 2026).
Beyond that, an agent browsing on a user’s behalf can use the file as a table of contents if it chooses to request it. We have seen no assistant vendor document that its crawler reads it, and no published study measuring its effect on citations. That is why we treat it as cheap and optional: it costs an hour, keeps an inventory of what the site publishes, and helps the agents that look.
What is the difference between llms.txt and robots.txt?
robots.txt tells crawlers what they may fetch. llms.txt suggests what is worth reading. Only the first one controls access, and it is the one that decides whether a shopping assistant sees the shop at all.
| robots.txt | llms.txt | JSON-LD | |
|---|---|---|---|
| What it says | Which crawlers may fetch which paths | Which pages matter, with a line on each | What the product is, its price, stock, delivery and returns |
| Who reads it | Every well-behaved crawler, including GPTBot and OAI-SearchBot | Agents and tools that request it | Google, Bing, Merchant Center and AI crawlers |
| Effect if missing | Everything is allowed | None for Google | The offer can drop out of agent comparisons |
| Priority for a shop | Check first | Optional, last | First |
A crawler blocked in robots.txt cannot fetch the product pages, whatever llms.txt lists, so check those rules before anything else.
Which product schema fields do agents read?
JSON-LD in a <script type="application/ld+json"> block states the offer as fields. Where the page says “CHF 3,890, in stock, delivered in two to three days”, the markup says price, priceCurrency, availability and a shippingDetails block.
An agent comparing five offers needs five prices in one unit and five delivery times in one format, and prose does not give it that. In our audit of 240 Romandie shops, 117 of the 211 read in full (55%) showed their price only in page text, with no Offer price in the structured data (Romandie audit, July 2026). Google also asks that structured data match the visible text, so the markup must come from the same record as the price on the page.
Google says structured data is not required for its generative AI search and that no special markup exists for it. That holds for Google ranking. Shopping agents and feed-based channels still compare on fields, which is why JSON-LD comes first for a shop.
Why publish a Markdown mirror of a page?
A Markdown mirror is the same page as plain text at the same address with .md added, for example /products/autobasket-discover.md. It strips navigation, scripts and cookie banners, so an agent spends its context on content.
The llms.txt proposal describes this pattern. Every page on this site has one, announced with <link rel="alternate" type="text/markdown"> and a matching HTTP Link header. Mirrors help most on size guides, return conditions and buying guides, and least on product pages, where the JSON-LD already carries the facts.
How do you keep the three layers from drifting apart?
Generate all three from one source. Here, the mirrors and llms.txt are built from the rendered HTML at every deployment. On a shop, render the JSON-LD from the product record that fills the visible price. The failure we find most often is a price updated on the page and left unchanged in the markup, and an agent that sees two prices trusts neither.
The Machine Readability Sprint sets up all three layers in seven days, and the Buyer Journey Check shows which of them your shop is missing today.