# Could a coding agent set up your product from your docs? Free docs check · EarnCitations

> Free: we read your developer docs as an AI crawler would and check what a coding agent needs to set you up, from the quickstart and auth to errors and llms.txt.

Source: https://earncitations.com/docs.html

> Every page on this site: https://earncitations.com/llms.txt (the whole site as one file: https://earncitations.com/llms-full.txt)

Free tool updated 2026-11-22

## Could a coding agent set up your product from your docs?

Developers ask Claude Code, Cursor and Codex to add a product to their app. The agent finds your docs, reads them as text and follows them. This checks what it would find there.

### The twelve checks

- Docs an agent can find. The homepage links to them (a docs subdomain, a /docs path, or a domain of the same name such as shopify.dev or dropboxapi.com), or you gave the docs address. If the first docs link has no quickstart (user help docs, or a developer landing page with a "Get started" heading and no code), we try up to three more, as an agent would: its own docs links, then the homepage's next docs link. A help centre for users (help., support.) ranks below developer docs, and counts only when the homepage links it as docs.

- Readable by AI crawlers. The docs' main page has text without JavaScript, isn't a sign-in page, and robots.txt lets GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot and PerplexityBot fetch it and the other pages we read (a robots.txt that answers with a server error counts as blocking everything, as crawlers treat it).

- A quickstart. A getting-started page inside the docs, not a sign-in page or a page outside them. If the docs list several and the first has no code, we try up to two more.

- Prerequisites stated. What you need first: an account, a key, a runtime version.

- Code as text. Code blocks in the quickstart's HTML, not screenshots or code added by JavaScript.

- An install or first-call command. npm install, pip install, go get, a curl request or similar.

- How to authenticate. The quickstart says how to get and pass a key, token or OAuth credential.

- What success looks like. An example response or output, so an agent can tell whether it worked.

- A machine-readable API contract. An OpenAPI file on your site, linked from the docs (we open it to check), at /openapi.json, or listed in an [RFC 9727](https://www.rfc-editor.org/rfc/rfc9727) API catalog at /.well-known/api-catalog.

- Errors documented. An errors or troubleshooting page linked from the docs, the reference or the quickstart.

- Versions named. A version near the instructions, or a changelog or migration guide linked.

- llms.txt points to the quickstart. If you have one, it links the quickstart.

Checks 4, 7, 8 and 11 read the page's own content (its <main> or article), not the sidebar or header, so a nav link called "Authentication" doesn't count as the quickstart explaining it.

Seven notes come with the score but don't count towards it. Each appears when:

- One name per credential: the docs use three or more names for keys and tokens.

- Sections that stand alone: a quickstart section's code needs a credential its text never names, or the text points back to "the key from setup".

- Errors an agent can act on: the errors page names no status or error code, never says when a retry is safe, or doesn't say what to do.

- Install commands that match npm: the quickstart installs a package npm marks deprecated, one pinned two or more major versions behind, or a name npm doesn't have.

- Where keys go, and how they change: nothing we read says whether a key may be used in a browser or only on a server, or how to rotate or revoke it.

- Quickstarts an agent can follow: a docs page names several quickstarts ("Node.js Quickstart", "Python Quickstart") as cards, and none of those names is the text of a link in the HTML (the cards may become links once JavaScript runs), so we set it aside and followed another docs link.

- The current path leads to current pages: the quickstart, the API reference or the keys page it links says at the top that it's deprecated, legacy or replaced, so an agent following the docs may copy old code.

See the questions below.

Add a competitor in the second box to see, check by check, where their docs give an agent something yours don't. The same check is earncitations docs <site> [--vs rival.com] (exits 2 when a check fails), the MCP tool check_docs and GET /api/docs?url=… ([docs](https://earncitations.com/developers.html#api)). To see everything a crawler reads on one page, use [what AI sees](https://earncitations.com/view.html).

### Where this fits

Jam splits "agent experience" into four stages: whether agents choose your product, understand it, can use it, and get it fixed. Here is what we have for each today, and what we don't.

| Stage | What we have | What we don't |
| --- | --- | --- |
| Chosen | The [question map](https://earncitations.com/map.html) and the [answer check](https://earncitations.com/claims.html) with your competitors: whether AI answers name and recommend you, per answer | What coding agents pick inside an editor; we measure answers people read |
| Understood | The [answer check](https://earncitations.com/claims.html) (each claim against the page it cites) and [what AI sees](https://earncitations.com/view.html) on any page | Running an agent over your docs to see what it gets wrong |
| Used | This docs check | Running the quickstart in a sandbox |
| Fixed | "Copy a prompt for your coding agent" on every result, and the [llms.txt generator](https://earncitations.com/llms-txt.html) | Pull requests to your docs repository (planned) |

Where this comes from: Jam (spreadjam.com) lists "agent experience" as coming, and has written about API docs for AI: real HTML rather than screenshots, no sign-in walls, consistent names for credentials, versions next to the instructions. This check is our deterministic take on the part we can measure without running an agent.

### Questions

#### Why would a coding agent read my docs?

Because developers now ask one to. "Add payments to this app" or "set up error tracking" sends an agent such as Claude Code, Cursor or Codex to choose a product and then follow its docs. If the agent can't find your quickstart, can't read it without JavaScript, or has to guess the install command or how to authenticate, it gets stuck or picks a product whose docs it can follow. Jam (spreadjam.com) calls this "agent experience": whether agents choose your product, understand it, can use it and can fix problems with it. This check covers the "use" part.

#### Does this run an agent?

No. It reads your docs the way an AI crawler does (server HTML, robots.txt honoured, no JavaScript) and checks that what an agent needs is on the page, as text. That makes it quick, free and repeatable, but a pass doesn't prove an agent will succeed. Try the quickstart from a clean machine too, and note where you had to guess.

#### Why does "code as text" matter?

An agent copies commands and code from the page. Code shown in a screenshot, or added to the page by JavaScript after it loads, isn't in the HTML a crawler gets, so the agent can't copy it. We look for <pre> blocks, or fenced code in Markdown docs, on the quickstart page.

#### What counts as a quickstart?

A page your docs link to as "Quickstart", "Getting started", "Installation", "Set up" or similar, inside your docs (sign-up, sign-in, pricing and dashboard links don't count). If your docs' main page is the quickstart, we use that. If the getting-started page is only a hub of links, we follow its likeliest quickstart (up to four tries, stopping at the first with code; "no-code" pages are skipped), as an agent would. If the docs list several quickstarts and the first has no code, we try up to two more. A marketing page with a "Get started" heading but no code and no quickstart link isn't one: we follow its docs links instead.

#### What about pages behind a sign-in?

They fail "Readable by AI crawlers". A docs link that redirects to a sign-in page, or a page that's a password form and little else, gives crawlers and agents the form, not your docs. Keep setup docs public; put account-specific values (a key, a project ID) behind the sign-in, and say where to find them.

#### What is "one name per credential"?

A note, not scored. When the docs call the same thing an "API key" on one page, a "secret key" on another and a "token" on a third, an agent (and a person) can't tell whether they are one credential or three. If they are the same, say so where it first appears; if they differ, say how. Jam's posts on API docs make the same point.

#### What is "sections that stand alone"?

A note, not scored. People and agents can land on one heading of a quickstart, not the top of it. We split the quickstart at its headings and, for each section with code, look for two things: code that needs a key or token while the section's own text never mentions one, and text that points back to something it doesn't name ("the key from setup", "as shown above"). The fix is a few words beside the example: which credential, where the code runs, and a link to the setup step by name. The idea comes from Jam's post on keeping context in developer docs. It is a reading check, not a claim about how any engine splits pages.

#### What is "errors an agent can act on"?

A note, not scored. A linked errors page passes check 10; this note reads it. An agent that hits an error needs to know which error it is (an HTTP status or a stable code, or an error type in a typed SDK), what to do about it, and whether retrying is safe. We note when the page has none of the first, never mentions retrying, or never says what to do. Jam's post on documenting API errors goes further: say whether the request may have taken effect, and respect Retry-After. Those are worth writing even though we don't check them.

#### Does a good score mean a coding agent can set us up?

Not by itself: the check reads your docs, it doesn't run an agent. Each result has "Copy a test to run with your coding agent": a fixed starting state (an empty directory, a new session), the allowed tools, the task to paste, what counts as success and what to record. Run it with your own agent on a test account. Run the named version (the prompt names your product) apart from the unnamed one (the agent has to choose), and never add the two together. Compare the agent's guesses with the steps the check flagged.

#### What is "where keys go, and how they change"?

A note, not scored. When the docs use a credential, we read the page that explains it (an "Authentication" or "API keys" page linked from the pages we read; pages about signing in to the dashboard, such as two-factor or SSO, don't count), plus one linked page about managing or rotating keys if that page doesn't say it all. The note appears when none of them, nor the quickstart, says where a key may be used (on a server only, or safe in a browser or app) or how to rotate, revoke or expire it. An agent writing code needs the first to know where the key goes, and the second for what to do when one leaks. The idea comes from Jam's method post on documenting API authentication for AI-assisted work.

#### What is "install commands match npm"?

A note, not scored. The quickstart's npm install, yarn add, pnpm add or bun add commands are compared with the npm registry, one small request per package (up to four): a package npm marks deprecated, a pinned version two or more major versions behind the latest, or a name npm has no public package for (a typo, or a private registry). Agents copy install commands exactly, so a stale one installs stale code. Python installs aren't checked: PyPI's robots.txt disallows its JSON API, and we follow robots.txt. Jam's playbook for API and SDK companies makes the wider point: the package registry, README, docs and quickstart are separate surfaces that can disagree.

#### How does my result compare with other companies?

Each result says where it sits among the software companies' docs in our [docs study](https://earncitations.com/research.html#docs): how many passed fewer checks and how many passed more. Each check also says how many of them passed it. Only counts are published, not which company scored what.

#### Can I compare my docs with a competitor's?

Yes. Put their site in the second box (or add --vs rival.com to the CLI). We run the same check on theirs, after yours, and list the scored checks theirs passes and yours doesn't, and the reverse. Notes aren't compared. A step one side never reached (no quickstart, so nothing to read on it) shows as "not checked". It compares what each site gives an agent to read, not which product an agent would choose.

#### What do you fetch?

Your homepage (to find the docs link), the docs' main page, the quickstart (and up to four pages it links to, if it is only a hub), the API reference, the errors page it links, the npm registry entry for up to four packages the quickstart installs, robots.txt for each host (checked before every page, for us and for the AI crawlers), the OpenAPI file it links or a guessed /openapi.json or /openapi.yaml (or the one an [RFC 9727](https://www.rfc-editor.org/rfc/rfc9727) catalog at /.well-known/api-catalog lists), the authentication or API keys page and one page about managing keys, and llms.txt: up to about twenty-five requests within 45 seconds, with our identified user agent ([EarnCitationsAudit](https://earncitations.com/bot.html)). The result is kept for 30 minutes so a repeat check is quick; a check cut short by a slow site is not kept.
