Everyone has a slide predicting AI agents will become major users of the web. We don't have a slide. We have logs. We put Rover on our own site, rtrvr.ai, and counted, including the parts that aren't flattering.
Your site has two kinds of visitors. We measured ours.
One embedded agent handled 17,063 interactions. Here's who was on the other end:
213 interactions came from autonomous AI agents across 37 distinct tools, arriving with someone's intent to get something done. And they didn't come to read: 4 out of 5 agent interactions (171 of 213) came through the agent endpoint itself, carrying a task to execute, not a page to browse. That's ~1 in 80, on a B2B site that isn't even optimized for agents. On commerce, payments, or API sites the share is higher, and it's revenue arriving through a door nobody built.
What Rover actually did
Not just chat. Work:
- 121 real multi-step tasks run end-to-end, 77% completed, averaging 5.8 steps and under two minutes each, no human in the loop.
- 573 prompts across 94 task runs: visitors and agents hand Rover jobs, not just questions.
- 92% of everything flowed through the embedded widget: one front door for humans and agents, measured.
The honest parts
We report the live number and the lab number. Live task success was 77% on a 121-task sample; our published Halluminate Web Bench score is 81.4%, state of the art.
And the number the logs could not give us: revenue. We had measured everything except the metric that matters, because no conversion outcome was ever defined for our own site. Most analytics stacks share that blind spot: rich activity data, no outcome attribution. It became a rule we apply to every customer on day one, and the reason step one of every Rover pilot is: pick the one number this has to move, and wire it before the agent goes live.
What this means for your site
- If 1 in 80 of our interactions is already an agent, your site needs a front door that serves agents (discovery, auth, execution), not just humans. That's what our agent endpoint speaks.
- 77% of multi-step tasks finishing with nobody in the loop is what "onboarding completed" and "ticket resolved by doing" look like in practice; it's the same capability compared against support agents in Rover vs. Sierra & Decagon.
- One script tag produced all of this data. The alternative is a year-long build and a $100M arms race.
Frequently asked questions
How do you know a visitor is an AI agent?
Rover classifies actors by how they arrive and behave, from declared agent protocols and signed requests down to behavioral heuristics, and records the attribution (tool, vendor, trust tier) per interaction. The 213 figure counts only confirmed autonomous-agent channels, not "likely bot" traffic.
One more finding from the logs: every one of the 213 arrived anonymous. Not a single agent declared who it was or who sent it. The web has no working identity layer for agents yet, which is exactly why sites need a front door that can authenticate them, not just detect them.
What do AI agents actually do on a website?
They arrive with a task: check a capability, extract an answer, complete a flow, book or buy on a user's behalf. Served well, they convert like intent-rich visitors, because that's what they are.
How do I make my site agent-ready?
Expose discovery (well-known endpoints), authenticate and meter agent traffic, and give agents a way to complete tasks, with human handoff for payments. Rover ships all of it in the same script tag that serves your human visitors. The manifesto has the full argument.
The agentic web isn't a forecast. It's in the logs.
We're onboarding our first five pilot companies now: one flow, one metric you choose, live the same day. Claim a founding slot →