Your docs' newest readers are AI agents

Coding agents now read documentation on a developer's behalf, and on one docs host they make about as many requests as browsers do. An agent can't ask what a screenshot meant. It learns your UI from what the page gives it, so a stale picture or a vague description teaches it the old screen.

How much docs traffic is agents

The best public numbers come from Mintlify, which hosts documentation for other companies and can see who requests it. It has published two reports, measured two different ways. Read them as two snapshots, not one line on a chart.

Report	Finding	How it was counted
March 2026, published 3 April 2026	AI coding agents made 45.3% of about 790 million requests over 30 days. Browsers made 45.8%. Claude Code alone was 25.2%.	Each request sorted by its user-agent header, the name a client sends with every request.
Midyear report, published 29 July 2026	Agents made 66% of measured traffic in July 2026, up from 15.2% at the start of the year.	Agent requests are machine routes plus requests from known AI addresses or user agents. Human visits are page loads counted in the browser.

How far to trust those numbers

It is one vendor's customers. Mintlify is a platform for developer documentation, whose readers are the people most likely to use a coding agent. Other kinds of docs may look different.
The methods differ. The March report counts requests by user agent, and says agents that send no recognizable one, such as Codex, are missed, so the real share is likely higher. The July report counts agent requests against human page loads, and notes that a change in how agents fetch pages can raise the share even if the work they do stays flat.
July was not over. The midyear report was published on 29 July and describes July "so far".

So don't say "two thirds of docs readers are AI". Say that on Mintlify's sites, agents make roughly as many requests as people or more, depending on how you count. That is enough to change how you write.

What an agent gets from a screenshot

An agent reading your docs page can get three things from a screenshot:

The alt text, the description in the image's alt attribute.
The text around it, the step it illustrates and the caption.
The image itself, if the tool fetches it and can read pictures. Not every tool does.

Each of those can be stale. If the image shows last year's settings page, an agent that looks at it learns last year's settings page. If the alt text says "Settings screenshot", the agent learns nothing, and a screen-reader user doesn't either. Agents then walk your users through what they learned, with full confidence.

What to do

Write alt text that says what the screen shows. "The Billing settings page, with the Change plan button under Current plan" beats "Screenshot". WCAG already asks for a text alternative that serves the equivalent purpose (success criterion 1.1.1, Level A), so this helps people too.
Put the steps in words. Name the button and where it is. Then a reader that never looks at the picture still gets it right.
Keep the image at a URL that serves the current screen. An agent that fetches the image should see today's product.
Offer a plain-text version of each page. Every page on this site has a Markdown copy: add .md to the URL (index.md for one that ends in /), or send Accept: text/markdown. The site's llms.txt lists them. An agent gets the text, the alt text and the image URLs without the page's navigation.

How Screenery helps

Screenery keeps the third point true. Your tests capture the screenshots, and a push on the default branch moves each one's @latest URL, so an agent that fetches the image gets the build you just shipped. Alt text stays yours to write; put it next to the URL in your docs.

Agents can also work with Screenery directly. Its MCP server, a standard way for an agent to call a service's tools, lets a coding agent set up screenshots for a repository and compare a pull request's screenshots with @latest. See Screenery for AI agents.

Sources

Every page below was opened on 13 November 2026, and each number on this page was checked against it.