Short answer. An llms.txt file is a plain-text map of your website for AI assistants. No major AI provider has confirmed it reads the file, so build it as cheap insurance. The real payoff is the clean, structured product data you write along the way.
An llms.txt file is a short, plain-text map of your website written for AI assistants. For most manufacturers it takes a day or two to build and costs nothing to host. Whether it gets you cited more often is still unproven, so treat it as cheap insurance rather than a growth strategy.
Why this matters now
The reason to care is how buyers research. A maintenance engineer hunting for a replacement seal, or a buyer shortlisting contract machining shops, now often asks ChatGPT, Claude, Perplexity or Google's AI answers before opening a catalog. Those tools answer from whatever they can read about you.
If your specs live in scanned PDFs, a JavaScript configurator and a login-gated distributor portal, the assistant has little to go on. It may guess, cite a distributor's stale listing, or name a competitor whose data is easier to read. This guide covers what llms.txt is, what it can and can't do, and how to build one that fits an industrial catalog.
What llms.txt is, and what it isn't
llms.txt is a Markdown file at the root of your domain (yourcompany.com/llms.txt) that tells an AI model what your company does and which pages to read. Jeremy Howard of Answer.AI proposed it on September 3, 2024. It is a community proposal, not a ratified standard.
The format is deliberately simple. In order, a file contains:
- An H1 with your company or site name. This is the only required part.
- A blockquote summary: one to three sentences on who you are and what you make.
- Optional plain paragraphs with context a model needs to read the rest correctly.
- Any number of H2 sections, each a list of links written as [Page name](url): what's on it.
- An optional section titled exactly "Optional". A model short on space can skip it.
The proposal has two companions. Pages can offer a clean Markdown copy at the same URL with .md added, so /products/gate-valves also serves /products/gate-valves.md. And llms-full.txt, developed by Mintlify with Anthropic, bundles the full text of every listed page into one file.
It sits next to two files your web team already manages, and it replaces neither:
| File | What it does | Who reads it today |
|---|---|---|
| robots.txt | Says which paths crawlers may visit | Google, Bing, GPTBot, ClaudeBot and most well-behaved bots |
| sitemap.xml | Lists every indexable URL | Search engines, as a discovery aid |
| llms.txt | Curates your best pages, with descriptions, for AI models | No major AI provider has confirmed it reads the file |
One common confusion: llms.txt does not control or block AI access to your site. If you want to allow or refuse AI crawlers, that still happens in robots.txt (GreenGeeks, May 2026).
The honest state of play
Today, an llms.txt file does not measurably help you show up in AI answers. The crawlers behind ChatGPT, Perplexity and Google's AI results mostly skip it and read your HTML pages instead. Four independent studies, gathered by Optimizely in October 2026, point the same way:
| Study | Sample | Finding |
|---|---|---|
| SE Ranking | About 300,000 domains | 10.13% had the file; no clear link to being cited in AI answers |
| Limy | 500M+ AI bot visits over 90 days | 408 requests for llms.txt |
| Otterly | 62,100+ AI bot visits over 90 days | 84 requests, about 0.1% |
| Ahrefs | 137,210 domains | 97% of files got zero requests in May 2026 |
Google has been the most direct. Gary Illyes said in July 2025 that Google does not support the file. On June 15, 2026, Google's AI optimization guide added that Search, including AI Overviews and AI Mode, does not use it, and that having one neither helps nor hurts.
The file does have readers, just not the ones most marketers hoped for. Coding agents such as Claude Code fetch it to find the right page fast. Chrome's Lighthouse team added an experimental audit that checks for it as a sign a site is ready for browsing agents. And any assistant a user points at the file can read it.
For a manufacturer, that leaves a narrow but real case. The file is cheap. The buyers most likely to send an AI agent to read your site are engineers and technical buyers. And building it forces you to fix the pages every AI system reads anyway, which is where the payoff actually is.
Why manufacturers have more to gain than most
Industrial websites are unusually hard for AI to read, so the gap between "invisible" and "legible" is wider for you than for a software company. Most of what a buyer needs to know sits in formats a model handles poorly:
| Where the information lives | What an AI assistant gets | The fix llms.txt pushes you toward |
|---|---|---|
| PDF datasheets, often scanned | Fragments, or nothing if the PDF is an image | A text or Markdown spec page per product family |
| JavaScript configurators | Empty page shells; the part-number logic is never written down | A plain page explaining how part numbers are built |
| Distributor portals and login-gated catalogs | Nothing | Public summaries of what you make and who sells it |
| "Contact us for pricing" product pages | A product name and little else | Key specs, materials, ratings and lead-time ranges in text |
| Certifications buried in footers or PDFs | Missed or guessed | One page listing every certification with scope and expiry |
| Shop capabilities spread across brochures | Vague claims | One page with tolerances, materials, envelope sizes and volumes |
The questions buyers ask AI are specific: who makes a 316 stainless ball valve rated to 1,000 psi, which shop can hold ±0.0005 in on titanium, what replaces a discontinued competitor part. An assistant can only answer them in your favor if those facts exist somewhere in plain text. llms.txt is the index that points to that text, which means you first have to write it.
What a manufacturer's llms.txt should contain
Organize the file around the questions buyers ask, not around your website's menu. The summary and the product list matter most; everything else supports them.
The summary blockquote should fit in three sentences and carry hard facts: what you make, for which industries, where you make it, your key certifications, and whether you sell direct or through distributors.
The context paragraph below it is where you explain anything a model would otherwise get wrong. Typical examples are how your part numbers are built, which product names are trademarks, regions you don't serve, and work you don't take on.
Then the linked sections, in rough order of importance:
| Section | What to link | Write the description as |
|---|---|---|
| Products | One page per product family | Spec ranges, materials, ratings and standards met |
| Specifications | Text spec pages, part-number guide, competitor cross-reference | What the page lets a buyer look up |
| Capabilities | Processes, equipment list, tolerances, materials, part envelopes | Numbers with units, not adjectives |
| Industries | One page per market served | Which products or processes apply there |
| Quality and compliance | Certifications, export status, RoHS, REACH, conflict minerals | Certificate, scope and expiry |
| Buying | Request a quote, distributor locator, lead times, minimums, warranty | How the buying process works, in one line |
| Support | Manuals, maintenance guides, CAD and drawing downloads | Which products each covers |
| Optional | History, news, careers, blog | Short or none; models may skip this section |
Keep the whole file short enough to read in a minute, typically 20 to 80 links. It is a curated index, not a sitemap. Link to public, text-based pages, not to anything behind a login or form.
A sample llms.txt
Here is a complete file for a fictional valve maker. Swap in your own products and pages; keep the shape.
# Northbrook Valve & Controls
> Northbrook Valve & Controls makes industrial ball, gate and check valves in carbon steel, 316 stainless and duplex, sizes 1/2 in to 24 in, ASME Class 150 to 2500. Plants in Cleveland, Ohio and Monterrey, Mexico. ISO 9001:2015 and API 6D certified; sold direct to OEMs and through authorized distributors in North America.
Notes for AI assistants:
- Part numbers follow Type-Size-Class-Material-End. Example: BV-4-300-SS-RF is a 4 in, Class 300, 316 stainless ball valve with raised-face flanges.
- "NorthSeal" is our trademarked seat material, not a separate product line.
- We do not make valves for potable water or sanitary (3-A) service.
- Prices are quoted per order; we do not publish list prices.
## Products
- [Floating ball valves](https://northbrook.example/products/floating-ball.md): 1/2-8 in, Class 150-600, carbon steel and 316 SS, API 608, fire-safe to API 607.
- [Trunnion ball valves](https://northbrook.example/products/trunnion-ball.md): 2-24 in, Class 150-2500, API 6D, for pipeline and midstream service.
- [Gate valves](https://northbrook.example/products/gate.md): 2-24 in, Class 150-1500, bolted bonnet, API 600.
- [Swing check valves](https://northbrook.example/products/swing-check.md): 2-24 in, Class 150-900, API 594.
## Specifications
- [Part-number guide](https://northbrook.example/specs/part-numbers.md): How to decode or build any Northbrook part number.
- [Pressure-temperature ratings](https://northbrook.example/specs/pt-ratings.md): Ratings by material and class per ASME B16.34, as a text table.
- [Cross-reference](https://northbrook.example/specs/cross-reference.md): Northbrook equivalents for discontinued and competitor part numbers.
## Quality and compliance
- [Certifications](https://northbrook.example/quality/certifications.md): ISO 9001:2015, API 6D and API 607 certificates with scope and expiry dates.
- [Materials compliance](https://northbrook.example/quality/materials.md): RoHS, REACH, conflict minerals, and material test reports (MTRs) by product.
## Buying
- [Request a quote](https://northbrook.example/rfq.md): What to include in an RFQ and typical response time.
- [Where to buy](https://northbrook.example/distributors.md): Authorized distributors by region.
- [Lead times](https://northbrook.example/lead-times.md): Standard lead-time ranges by product family, updated monthly.
## Support
- [Manuals](https://northbrook.example/support/manuals.md): Installation and maintenance instructions by valve type.
- [CAD models](https://northbrook.example/support/cad.md): STEP files and drawings for every standard configuration.
## Optional
- [About Northbrook](https://northbrook.example/about.md): Company history since 1962.
- [News](https://northbrook.example/news.md): Product launches and new certifications.
Three things make it work. The summary alone answers "what do they make and can they meet my spec?" The notes stop a model from misreading part numbers and trademarks. And every link description carries numbers a buyer would filter on.
Building and publishing it, step by step
Most of the work is writing, not engineering. Budget a week of part-time effort for a mid-size catalog, mostly spent on steps 2 and 3.
- Collect the questions. Pull 20 to 30 real questions from sales, inside sales and recent RFQs. Ask each one in ChatGPT, Claude, Perplexity and Google's AI Mode, and save the answers as your baseline.
- Find the pages that answer them. For each question, note which public page holds the answer. Flag any answer that only exists in a PDF, a configurator, a portal or someone's head.
- Fill the gaps in plain text. Write the missing spec pages, the part-number guide, the certifications page and the capabilities page. This step does more for AI visibility than the file itself.
- Draft the file. Write the summary, the notes and the linked sections using the structure above. Every description should state a fact, not a slogan.
- Generate it if your catalog is large. If products change often, build the file from your PIM, ERP or CMS so it can't drift out of date. WordPress sites can use the llms.txt features now in Yoast SEO and Rank Math, then hand-edit the descriptions.
- Publish it at the root. The file should load at https://yourdomain.com/llms.txt as plain text. Optionally publish .md copies of key pages and an llms-full.txt.
- Check that AI crawlers can reach you. Review robots.txt and your CDN or firewall bot settings. A perfect llms.txt is pointless if the pages it links to are blocked.
- Test it. Render the Markdown, click every link, then paste the file into two or three assistants and ask your baseline questions. The llms_txt2ctx tool from llmstxt.org expands the file and its links into one context file for this kind of test.
- Give it an owner. Review it whenever a product launches, a certification renews or a page moves, and rerun the baseline questions each quarter.
Common mistakes
The most expensive mistake is expecting traffic from the file itself. The rest are about what you put in it.
- Linking to PDFs, gated pages or the configurator. If a model couldn't read the destination before, a link to it won't help.
- Writing slogans instead of specs. "Industry-leading quality" tells a model nothing; "ISO 9001:2015, 0.3% first-pass reject rate" does.
- Listing every SKU. Thousands of links turn an index into a sitemap. Link product families and let the family page carry the detail.
- Letting it go stale. A file that lists discontinued products or an expired certification is worse than none, because the few tools that read it treat it as true.
- Claiming what the site doesn't show. Statements in llms.txt should be backed by the linked pages. AI systems increasingly check pages rather than trust a site's description of itself.
- Leaving plugin output untouched. Auto-generated descriptions are generic. Edit them to carry the numbers buyers filter on.
- Exposing what should stay private. Leave out negotiated pricing, customers under NDA and anything export-controlled.
- Using it to block AI. llms.txt grants and restricts nothing. Access rules belong in robots.txt.
The real work: clean, structured product data
llms.txt is a table of contents. What gets you recommended is the content it points to, and for manufacturers that content starts as product data scattered across ERP records, spreadsheets, PDFs and the memory of senior engineers.
AI assistants answer one page at a time: they fetch whichever page best answers the question and cite it. So the durable investments are the same ones that make a good catalog for humans:
- One source of truth for specs. Keep product attributes in one structured system and generate web pages, datasheets, distributor feeds and llms.txt from it. Hand-maintained copies drift apart, and AI will happily repeat whichever version it finds.
- Answer-first product pages. Lead with what the product is, its key ratings and what it's for, then the detail. That is easier for a buyer to skim and for a model to quote.
- Honest structured data. Standard schema.org Product markup, with part numbers and key specs as properties, describes the page in a format search systems already use.
- Measure citations, not file requests. Track whether AI answers mention and cite you for your baseline questions, not whether a bot fetched llms.txt.
If AI agents start sourcing parts and shortlisting suppliers on buyers' behalf, as many expect, they will favor suppliers whose specs are complete, consistent and machine-readable. llms.txt is a small, cheap step in that direction. The data underneath it is the bigger one.
Bottom line and checklist
Build an llms.txt, but build it for the right reason. On its own it won't win you AI recommendations. The process of writing it will expose every spec, capability and certification that AI systems can't currently read, and fixing those gaps is what moves the needle.
- Collect 20 to 30 real buyer questions and record today's AI answers
- Map each question to a public, text-based page
- Write the missing spec, part-number, certification and capabilities pages
- Draft the summary, notes and linked sections with factual descriptions
- Generate the file from your product data if your catalog changes often
- Publish at /llms.txt and confirm it loads as plain text
- Check robots.txt and CDN settings let AI crawlers reach your pages
- Test the file in two or three assistants with your baseline questions
- Assign an owner and rerun the questions every quarter
Sources
- The /llms.txt file proposal, Jeremy Howard, llmstxt.org
- What Is llms.txt: Does It Work & How to Add One to Your Site, GreenGeeks, May 2026
- Does llms.txt work? What the AI search data says, Optimizely, October 2026





