# Can an AI agent find its way through your docs?

44 of 143 products' docs give an AI agent all four things it needs to find its way. The pages themselves hold up best. The code samples and the index are where docs fall short.

Drives: 12 Oct 2026 to 1 Nov 2026 · 143 products · Counted 2 Nov 2026 · Source: https://agentdrives.com/lab/docs-structure · CC BY 4.0

## What we found

| Part | Holds up | Products |
|---|---|---|
| An index that lists the docs | 78 | 143 |
| One topic per page, under headings | 141 | 143 |
| Code on API reference pages | 84 | 123 |
| A quickstart with a first request | 100 | 143 |

- 141 of 143 keep one topic per page, under headings.
- 120 of 143 publish an llms.txt, and 7 more only a sitemap.
- 69 of the 120 llms.txt files list at least half of the docs pages we read. 51 list nearly all of them.
- 78 of 143 have an index that holds up: an llms.txt that lists at least a third of the docs pages we read.
- 51 of the 123 products with API reference pages put a code sample on nearly every one.
- 100 of 143 have a quickstart with a first request to copy. 27 have a quickstart without code, 16 none.
- Beyond the 44 that have all four, 89 get there in part.

## For docs owners

- List every docs page in your llms.txt, each with a line on what it covers.
- Put a request an agent can copy on every API reference page.
- Make the quickstart go from a key to a first working request, as a code block.
- Keep one topic per page, with a heading for each step.

## For agent builders

- Read llms.txt first, then the sitemap: 51 of the 120 llms.txt files list less than half of the docs.
- Look for the quickstart's first request before the API reference: 100 of 143 products have one to copy.

## How we tested

- 143 of the 204 products we list: every one whose drive read its docs, drives from 12 Oct 2026 to 1 Nov 2026.
- Each drive reads the product's docs from its llms.txt, else its sitemap, else its navigation, and keeps each page's title, headings, length and code blocks.
- Four parts, each scored from 0 to 1: the index (an llms.txt by how much of the docs it lists, else a sitemap), pages with one topic under headings, API reference pages with a code block, and a quickstart with a code block.
- A part holds up at 0.6 or more. A product passes with an average of 0.85 or more and no part under 0.6.

## Limits

- Each product's docs as they were on the day of its drive.
- For 62 products the drive read part of the docs. Their scores are for the pages it read.
- Code samples are counted on API reference pages. Whether docs pages link to the llms.txt isn't checked.
- Counts under 5 are left out.

## Sources

- [The standard: docs an agent can navigate (DOCS-11)](https://agentdrives.com/standard.md#docs-11)
- [The llms.txt format](https://llmstxt.org/)
- [How we drive](https://agentdrives.com/method.md?via=md)
- [The AgentDrives Lab](https://agentdrives.com/lab.md)