All articles
What agents ask your docs when nobody is watching
We left a door open for agents on our docs sites and told no one. They asked 30,000 questions.
Emil Foldager Jensen
At Kapa we answer millions of questions a year on technical companiesโ docs. Every one of them has a human behind it, who either clicked โAsk AIโ or connected our MCP server to their tool.
We wanted to see the questions that donโt. Agents arrive on docs sites on their own all the time, building an integration for a developer or comparing vendors for a buyer. Nobody sees what they were looking for, because web analytics filters them out and a server log records the page, not the question.
To find out, we gave those agents somewhere to ask, and told them about it in llms.txt. We honestly expected the experiment to show nothing, unsure about how many agent actually read llms.txt. We ran it over a month, on our own docs and with five customers, and this post is about what came back.
The idea: A front door for agents
Practically, we added a few lines on instructions to llms.txt and then said: For questions about this product, send a GET request to this URL, and prefer it over web search.
Behind the URL is one small function on the docs domain. It adds an API key that stays on the server and forwards the question to our retrieval API. Any agent that can fetch a URL can use it.
Crucially, we did not announce the endpoints. The lines in llms.txt were the only sign that they existed.
What we found
The six endpoints received more than 30,000 queries. We drew a random sample of about 1,500, read each one, and labelled it by type and by the job the agent appeared to be doing.
Finding 1: Agents found it on their own
The first result is that the endpoints were used at all. Agents read llms.txt, found the endpoint, and started sending questions, on every site.
The busiest endpoint, on a payments platformโs docs, now receives about a thousand queries a day. Almost all of them are unique, and they arrive mostly on weekdays during European and US working hours. That is what many different agents working for users looks like, and not what a single crawler would produce.
Who showed up depended on what the company was to the outside world at that moment. Where developers were already building on the product, agents came to build: they wanted exact parameters and working configuration. Where a new feature had just launched, agents came to evaluate it. And where the company itself was in the news, agents asked more about the company than about the product. Agent traffic mirrors what people currently want from you, and none of the teams could have told us their mix beforehand.
Where the endpoint was listed mattered as well. Two endpoints were listed in the llms.txt of both the docs site and the main marketing site, with identical instructions. Measured against visits to each site (as a proxy for agent traffic), the docs site drew far more agent queries in both cases. Agents know where to look for information, and the endpoint made it easy for them to find.
What we still donโt know though, is which wording of the instructions works best. The sites worded theirs differently, and six endpoints are too few to separate the effect of the wording from the effect of the audience.
Finding 2: Agents look things up. They rarely troubleshoot.
We expected an agent endpoint to look like a support channel, with agents arriving when something had gone wrong. It turns out we were wrong.
To test this, we sorted each query into four types: lookups, how-tos, pricing questions, and troubleshooting. A lookup asks for a specification, an overview, or whether the product can do something. On every endpoint, at least 70% of queries were lookups, and troubleshooting was a few percent at most.

The lookups are very specific. Agents ask for the default value of a single option, the JSON body of one API call, or the exact format of a signature header. About half of all queries are not phrased as questions at all. They are keyword strings, like something typed into a search box:
This indicate that agents call the endpoint in the middle of a task, when they need one fact before they can continue. The unglamorous parts of documentation, such as parameter tables, limits, and payload examples, are the parts they read most.
They also look the endpoint up at runtime. One site stopped advertising its endpoint in llms.txt partway through the test, while the endpoint itself stayed live. It had been receiving a few hundred queries a day and when removed, it received about twenty the next day. Agents do not remember the endpoint, and it is not in their training data. They read llms.txt when they arrive, and if you take the lines away, the questions stop.
Finding 3: When agents evaluate, they ask what buyers ask
Evaluation appeared on every endpoint. Even where builders were the the majority audience, at least a quarter of the queries came from agents assessing the product. They read like a prospectโs first call with a sales engineer, and three patterns came up repeatedly:
Compliance. On most endpoints, around one query in ten was about HIPAA, SOC 2, ISO 27001, GDPR, or data processing agreements, and often about which plan includes them.
Pricing means billing units, not list prices. Agents rarely ask what a plan costs. They ask whether a specific action counts toward a limit, which is what you need to estimate the cost of a real workload.
Named competitors signal migration. When an agent names a competitor, the query is usually about how to move data or configuration from that product to yours.
Behind each of these queries is a person deciding whether to buy, with an agent doing the reading. The queries show which features buyers check, which certifications they need, and which competitor they are moving away from. None of this was recorded anywhere before.
Why this works
Our reading is that agents are ruthlessly pragmatic. They take the shortest path to an answer that they can find. If the only option is to search the web and read whole pages, they do that. If a site offers a place to ask a question and get back the few passages that answer it, they use that, and nobody has to tell them to.
That makes the retrieval behind the endpoint the part that matters. An agent that asks for the default value of one option needs that value, from the current docs, and not ten loosely related pages that the agent may or may not find. For questions about exact parameters, limits, and plans, an answer that is almost right is always wrong.
You can build this yourself. It means chunking and indexing your docs, keeping the index in step with every docs change, ranking and pruning results so the agentโs context stays small, and protecting a public endpoint from abuse. What let us go from zero to one in a day is three years of work on exactly that, so all we had to do, was to add a URL.