Agents Now Account for 80% of Traffic to the Documentation of 20,000 Companies — In January It Was the Other Way Around

The headline "AI agents read 80% of the internet" practically begs to be filed under clickbait. It is imprecise, but the story itself is more interesting than the headline suggests. It is about how documentation for developers has changed its primary reader.

What was actually measured

The figure does not apply to the entire web. It refers to traffic to sites with official developer documentation owned by roughly 20,000 companies. Among them are Anthropic, Microsoft and Amazon. Currently about 80% of this traffic comes from agents. In January the ratio was the opposite: humans predominated.

Over ten months, the majority of documentation readers shifted from humans to programs. That dynamic is what is worth remembering, not the round number from the headline. Note that we are retelling figures from a Substack post and from an analysis by Andrey Sebrant. How exactly the authors separate agents from humans is not clear from the retelling. Anyone who plans to cite this data should open the original source and check the methodology.

Why documentation specifically

Documentation is almost ideal content for an agent. It has a clear purpose: find out how to call an API method, which parameters a function accepts, what a specific error means. Previously a developer with such a question would open a page and read. Now they give the task to a coding assistant, which goes to the site itself, picks up the pieces it needs and writes the code.

Humans, meanwhile, read documentation less often and indirectly, through the agent's response. This is our interpretation, but it explains well why visit counters have shifted so sharply. A single developer request can generate several agent requests to different pages, so the shift in traffic looks more dramatic than the shift in the actual number of users.

What to do about it

If you are responsible for documentation or a product with an API, several practical steps follow from this.

• Separate your analytics. Look at what share of requests to your docs comes from agents. If you count all visits as human, your metrics are misleading you.
• Write for the machine as the first reader. Clear headings, complete code examples that can be copied and run, explicit API versions, understandable error descriptions. An agent does not guess from context; it takes what is written.
• Don't hide the substance. Key information in tabs, images or an interface rendered only by scripts risks not reaching the agent at all.
• Consider machine-readable formats. For example, versions of pages in plain markup or the community-proposed llms.txt file with a list of important materials. This is not a standard but a convention, and not everyone supports it, but the cost of adoption is low.
• Measure success differently. Page views are a poor indicator of whether the documentation helped. It is more useful to see whether the agent completed a task with your API on the first attempt.

Humans still need documentation, but a different kind: concepts, architecture overviews, learning from scratch. The reference layer is increasingly turning into an interface for agents.

Limitations

First, this is a sample of documentation sites, not of the entire internet. Blogs, news and forums are read differently, and the 80% cannot be transferred to them. Second, this is a share of visits, not a share of readers. An agent can make dozens of requests where a human would open one page. Third, as already stated, the details of the methodology are unknown from the retelling.

The bottom line: even with caveats, the trend is obvious. If developers use your product, a significant part of the first acquaintance with it today is made not by a human. Documentation should be tested the same way you test an interface: can a new reader, this time a programmatic one, make sense of it.