vercel.com — llms.txt & GEO score

Vercel

Vercel is a cloud platform for building, deploying, and scaling web applications and AI workloads.

7 of 11 checks pass. What is left: descriptions say something concrete, points at clean markdown copies, list items are markdown links and enough written detail per link.

• Every link carries a description 24 of 24 described (100%)
  10/10
• Descriptions say something concrete Average 64 characters across 24 described links
  6/7
• Descriptions add to the title No description merely repeats its title
  4/4
• Points at clean markdown copies 20 of 24 end in .md or .txt (83%, detected by file extension)
  8/10
• Link titles are descriptive All 24 titles are meaningful
  4/4
• A workable number of links 24 links (12–50 is the sweet spot)
  4/4
• Skippable links marked Optional An ## Optional section is present
  3/3
• No duplicate links All 24 URLs are distinct
  2/2
• List items are markdown links 8 of 32 list items are plain text, not links
  3/4
• Enough written detail per link 22 words per link across 530 words
  2/3
• Links are absolute URLs All 24 links are absolute URLs
  1/1

All 7 checks pass in this group.

• Exactly one H1 project title “Vercel”
  3/3
• Summary directly under the title Present as a blockquote
  3/3
• Summary is a usable length 98 characters
  3/3
• Concrete language, not marketing filler No marketing filler detected
  3/3
• Grouped into balanced sections 9 H2 sections, largest group holds 5 of 24 links
  4/4
• Conforms to the published grammar Markdown only, headings used as the spec defines them
  2/2
• Section names say what is inside All 9 headings are descriptive
  2/2

All 5 checks pass in this group.

• At the root of the domain you typed Found at the standard /llms.txt path
  2/2
• Sent as plain text Content-Type: text/plain; charset=utf-8
  1/1
• robots.txt exists 629 B of crawler rules
  1/1
• AI crawlers are not blocked No AI user-agent is disallowed at the root
  6/6
• robots.txt points at a sitemap A Sitemap: line is declared
  1/1

5 of 6 checks pass. What is left: homepage has a meta description.

• Homepage title is the right length 31 characters, from title
  2/2
• Title is not padded or repetitive No repeated terms
  1/1
• Homepage has a meta description 45 characters, from meta description — thin for a snippet
  1/2
• Canonical URL declared https://vercel.com
  1/1
• Language declared lang="en"
  1/1
• Social card image og:image is set
  1/1

3 of 5 checks pass. What is left: server reports when it last changed and the index is kept current.

• Publishes llms-full.txt Companion full-text file found
  3/3
• The index points at the full file llms.txt links to llms-full.txt
  1/1
• Index is a sensible size 4.6 KB
  2/2
• Server reports when it last changed No Last-Modified header — clients cannot tell when this changed
  0/1
• The index is kept current Cannot be confirmed — the server sends no Last-Modified header
  0/2

• Points at clean markdown copies Content & links +2

  Right now: 20 of 24 end in .md or .txt (83%, detected by file extension)

  Publish a .md version of each page and link to that. The single biggest win here: the model reads your content instead of your navigation and footer.

  Where: Your CMS or docs build — this is the biggest single win here

  1. Publish a .md version of each page you link to, at a predictable URL (page.md next to page).
  2. Point the llms.txt links at the .md copies, not at the HTML pages.
  3. Strip navigation, cookie banners and footers from the markdown — that is the whole point: the model reads your content instead of your chrome.
  4. Declare the pair in your HTML with a link relation so clients can find it without guessing.

  In your file

  line 37

  - [Vercel REST API OpenAPI description](https://vercel.com/openapi.json): Machine-readable OpenAPI 3.0 description for the public REST API.

  line 62

  - [Agent resource catalog](https://vercel.com/.well-known/ai-catalog.json): Experimental catalog of agent-facing resources.

  line 63

  - [Product taxonomy](https://vercel.com/docs/taxonomy.json): Canonical product names, aliases, and deprecations.

  Placeholders in { } are yours to write — we do not invent text about your pages.

• The index is kept current Depth & upkeep +2

  Right now: Cannot be confirmed — the server sends no Last-Modified header

  Regenerate llms.txt when the docs change. A year-old index quietly points models at pages that have moved.

  Where: Your build pipeline

  1. Regenerate llms.txt in the same step that publishes content, so it cannot go stale silently.
  2. Treat it as a build artifact, not a hand-edited file — hand-edited indexes drift within a quarter.
  3. An llms.txt with discontinued products is worse than none: a model will repeat the wrong facts confidently.
• Descriptions say something concrete Content & links +1

  Right now: Average 64 characters across 24 described links

  Write 60+ characters that add information the title does not: numbers, limits, formats, names.

  Where: The text after ": " on each link

  1. Write 60 characters or more, and make them add information the title does not.
  2. Cover every link: this is measured per LINK, not per description, so an undescribed line counts as zero.
  3. Name the concrete things on the page: numbers, limits, formats, model names.
  4. If the description could be pasted under any other link, it is not a description yet.

  In your file

  line 56

  - [Vercel Academy](https://vercel.com/academy/llms.txt): Courses and learning paths.

  - [Vercel Academy](https: {what is on the page — numbers, limits, formats}

  line 58

  - [Vercel Changelog sitemap](https://vercel.com/changelog/sitemap.md): Product updates and release notes.

  - [Vercel Changelog sitemap](https: {what is on the page — numbers, limits, formats}

  Placeholders in { } are yours to write — we do not invent text about your pages.

• Enough written detail per link Content & links +1

  Right now: 22 words per link across 530 words

  A list of names is not documentation. Aim for 30+ words per link across the file — summary, section context and descriptions together. For an exhaustive map what matters instead is that the list is grouped so a model can skip whole branches.

  Where: The descriptions and section context in llms.txt

  1. Count the words in your file against the number of links — aim for 30 or more per link.
  2. A list of page names is an index, not documentation; the descriptions are what a model actually reads.
  3. Add a line of context under each "## " heading saying what that group covers.
• Homepage has a meta description Search-engine basics +1

  Right now: 45 characters, from meta description — thin for a snippet

  Add a 70–160 character meta description. It is the snippet under your result, and assistants reuse it verbatim.

  Where: The <meta name="description"> tag on your homepage

  1. Target 70–160 characters; beyond that Google truncates it mid-sentence.
  2. Write it as the sentence you would want quoted, because assistants reuse it verbatim.

  Instead of

  <meta name="description" content="Welcome to our website.">

  Do this

  <meta name="description" content="Industrial spare parts for food processing lines, shipped across Europe with same-week dispatch and two-year warranty.">

• List items are markdown links Content & links +1

  Right now: 8 of 32 list items are plain text, not links

  A file list needs `- [Title](url): description`. A plain line gives a model something to read but nowhere to go — it cannot follow it, quote it, or check it.

  Where: Every "- " line in llms.txt

  1. Turn each plain line into a markdown link: "- [Title](https://your-site/page.md): description".
  2. A line without a URL gives a model something to read but nowhere to go — it cannot follow, quote or verify it.
  3. If a fact has no page behind it (an address, a phone number), put it in the free-text body above the sections, not in a file list.
  4. Note that "#" is a heading in markdown, not a comment — lines starting with # become H1/H2 headings.

  In your file

  line 9

  - Deploy and scale web applications, APIs, and AI workloads from a Git repository or the Vercel CLI.

  - [Deploy and scale web applications, APIs, and AI workloads from a Git repository or the Vercel CLI.]({https://your-site/path}.md): {what this page covers}

  line 10

  - Automate projects, deployments, domains, and account resources through the Vercel REST API or SDK.

  - [Automate projects, deployments, domains, and account resources through the Vercel REST API or SDK.]({https://your-site/path}.md): {what this page covers}

  line 11

  - Connect an authenticated AI assistant to Vercel projects, deployments, logs, analytics, and documentation through Vercel MCP.

  - [Connect an authenticated AI assistant to Vercel projects, deployments, logs, analytics, and documentation through Vercel MCP.]({https://your-site/path}.md): {what this page covers}

  Placeholders in { } are yours to write — we do not invent text about your pages.

• Server reports when it last changed Depth & upkeep +1

  Right now: No Last-Modified header — clients cannot tell when this changed

  Let the server send Last-Modified on llms.txt. Without it a client has to re-download the whole file to notice a change.

  Where: Your web server configuration

  1. Serve llms.txt as a real file on disk so the server can derive Last-Modified from its mtime.
  2. If it is generated by an application route, set the header yourself from the generation time.
  3. Without it a client has to re-download the whole file to notice a change, and cannot answer "is this current?".

  Instead of

  HTTP/1.1 200 OK\nContent-Type: text/plain

  Do this

  HTTP/1.1 200 OK\nContent-Type: text/plain\nLast-Modified: Wed, 10 Sep 2026 08:14:00 GMT

This score is Devsly’s own heuristic, not an official standard — llms.txt has no scoring body. It measures how well a file serves an AI assistant, and it does not predict whether any assistant will cite the site. The reading is a snapshot from 20 September 2026; run it again for the current state, or check a different site.