A documentation page can look complete in a browser while its original HTML lacks the text a reader needs. An AI agent fetching that page through a plain HTTP request without running JavaScript will miss that text. Agentic Engine Optimization (AEO) helps identify these barriers by checking content access, structure and ways to discover related material.
The AEO audit in milten.io examines a selected page and public site files. Its report shows which checks passed, what needs attention and where information is unavailable.
What is Agentic Engine Optimization?
Agentic Engine Optimization means preparing a website for AI agents to read and use. For documentation, that means an agent can retrieve an operation's description, find an example and follow a link to a related reference.
The audit is useful for integration guides, API references and pages with technical examples. It complements SEO by examining how a software client can access published content. The report does not measure search rankings or the likelihood of being cited in AI answers.
What does an AEO website audit check?
The milten.io score combines eight checks in five groups:
| Group | What is examined | Score weight |
|---|---|---|
| Content discovery | robots.txt rules and the llms.txt index | 20 |
| Page structure | Headings, semantic markup, examples, Markdown and signs of JavaScript dependency | 25 |
| Context size | Token counts and page metadata | 25 |
| Interaction rules | Access and constraint descriptions in agent-permissions.json | 5 |
| Sharing text with an agent | Content-copy controls and links to raw text | 10 |
Each check includes an outcome, points and findings. For example, the audit examines both the presence of llms.txt and its contents: links, descriptions of materials and sections. Having the file alone does not make documentation easy to use.
Token counts help estimate how much text a model would need to process. They do not calculate request costs, since models use different tokenizers and prices.
The audit inspects the original server response without executing JavaScript. It looks for structural signs of problems but does not run an AI agent on a user task. Checking whether a particular agent can use the documentation requires a separate attempt at that task.
How ARD helps agents discover site resources
Agentic Resource Discovery (ARD) describes the discovery of resources available to an AI client, such as an API, MCP server or workflow. The client can use a description to find a suitable resource, then access it through that resource's own interface.
The ARD section checks how these descriptions are published. It looks for /.well-known/ard.json, HTML and HTTP links, Agentmap directives in robots.txt, and embedded ARD JSON-LD. It also recognizes the older ai-catalog.json format to identify material that may need migration.
The report lists discovered resources, the addresses of their descriptions, field errors and any limits that prevented a complete check. A missing catalog, a failed fetch and an invalid document appear as different outcomes. This publication check does not confirm that the listed APIs or MCP servers work.
ARD results appear separately and do not affect the AEO score. This section previously belonged to the speed scanner; new checks run in AEO, while saved speed reports remain accessible.
How to improve your site after an AEO audit
Start with a page an agent needs to read for a specific task. For an integration guide, for example, check whether the operation's description, parameters and sample request are available.
- Check access to the main text. If the report flags a JavaScript dependency, compare the original HTML with the browser content. If the required text is missing from the HTML, consider server rendering or an accessible Markdown version.
- Review the structure. Use sequential heading levels, place examples beside the operations they explain and remove repeated explanations. A table helps where the reader needs to compare parameters or options.
- Review the volume. A large reference can be split into separate topics linked to one another. Keep the conditions and examples needed to understand each page.
- Check how supporting material is discovered. If you use
llms.txt, update its links and short descriptions. If you publish resources through ARD, fix unavailable catalog addresses and invalid entries. - Audit the same URL after publishing your changes. Compare the findings, then give your chosen AI agent the original task, such as finding an API operation's parameters and preparing a request.
How to read the AEO score
The eight checks have a combined weight of 85 points. Checks for AGENTS.md and SKILL.md are disabled. The score ring displays a percentage of 85; saved reports are also recalculated without these two checks.
The current scale has a limitation: the structure check awards at most 12 of its nominal 15 points. The attainable total is therefore 82 out of 85, or 96% after rounding. That percentage alone does not mean an unresolved problem remains on the page.
Review the findings and fetch errors when using the report. The rules rely on heuristics and may miss unfamiliar markup, wording or languages. A high score does not guarantee search rankings, citations in AI answers or support for every format by a particular agent.
Frequently asked questions about optimizing for AI agents
Does an AEO audit check the whole website?
One run examines the selected URL and the public site files retrieved for it. It does not crawl every page or read a private repository. Run a separate check for another guide or reference page.
Should I allow every crawler in robots.txt?
No. Blocking a particular crawler may be an intentional site policy. Compare the finding with your intended access policy and change the rules only for the clients you want to allow.
Is adding llms.txt enough?
No. An index of links cannot fix inaccessible text, fetch errors or incomplete examples. Make the required content available first, then decide whether llms.txt suits your agent workflows.
Run an AEO website audit on one guide or API page. Use its findings to choose your edits, then check the result against the same task.
