AI agents visit your site. Do they really find the answer?
AgentRadar puts a question of your choice to two AI agents from different providers on your site, ten nominal runs per agent. Each answer is judged by a model from a different provider than the agent being tested; you receive the number of compliant answers out of the number measured.
Not a marketing simulation. A count, with its confidence interval.
✓ One site and a question of your choice · ✓ Two agents from different providers, ten nominal runs per agent · ✓ Single report sent by e-mail
An agent tries once, free, right here
This free test: an agent tries once, just to see. AgentRadar (paid measurement, described further down): two agents from different providers try ten times each, another AI checks every answer — a complete measurement, not a preview.
2 free runs per site · ~10-20 seconds · no signup
A loss of revenue no classic statistic ever shows
More and more real decisions — booking, visiting, choosing a route — now pass through an AI agent before they ever reach a human. When that agent fails to get the information, because it's on the wrong page, badly worded, or missing, the potential visitor leaves. Probably toward a competitor.
Nothing shows up in classic analytics: no measurable bounce, no exploitable trace. An invisible loss — one that grows with AI traffic.
An audit can tell you a page contains Schema.org.
A crawler can tell you an llms.txt file exists.
A visibility tool can tell you your brand appears in some answers.
AgentRadar answers a different question: "When a real AI agent comes looking for this exact fact, does it find it?"
A four-step measurement, from the question to the report
Choose
You give your site and the question you want put to AI agents.
Establish the facts
From at most five pages of your site, chosen by the service, the service determines the facts an agent must be able to restate. Each fact is tied to a quote taken word for word from one of those pages.
Ask and judge
Two agents from different providers receive the question: ten nominal runs per agent. Each answer is judged by a model from a different provider than the agent.
Receive
A single report is sent to you by e-mail: the number of compliant answers out of the number actually measured, with its 95% confidence interval, the detail of the runs and the protocol. If a technical or cost cap is reached before the ten runs, the report states the number of runs actually measured.
Our own measurements, published with their protocol
Research measurements on our pilot sites: these are not client results, and not what the AgentRadar Measurement delivers (see "What you get" below). We put real questions to an AI agent before, then after a change to the site. The results are only published with their protocol and their limits, on the Lab page.
Calling the tool, not just reading the page
An emerging standard (WebMCP) lets a site directly declare tools an agent can call to get a structured answer, instead of interpreting an HTML page the way a human would.
On the pilot site, a real agent receives the actual mission — "Le sentier de l'Imbut, dans les gorges du Verdon, est-il actuellement ouvert ou fermé ? Si fermé, depuis quand et jusqu'à quand ?" (question measured in French) — without being told to use the tool. Ten runs, judged fact by fact, dated journal.
Measurement of 6 September 2026, our pilot site, tool declared by us (agent gpt-5.4-mini, judge Claude Haiku 4.5). Third version of the mission: the first two were discarded. Exploratory result.
On the 3 remaining runs, the tool answered correctly but the agent called the same tool again instead of concluding — a limit of the model being tested, not of the tool.
A single report, sent by e-mail
The result
The number of compliant answers out of the number actually measured, with its 95% confidence interval.
The verifiable facts
The facts identified on your site, each with its quote taken word for word from a page.
The detail of the runs
The detail of the runs carried out, with your question.
The protocol
The protocol used for the measurement.
Find out what AI agents actually find on your site.
One site and a question of your choice, two AI agents, ten nominal runs per agent: you receive a single report by e-mail.
A question of your choice, put to two AI agents from different providers on your site as it stands at the time of the runs.
- One site and a question of your choice: two AI agents from different providers, ten nominal runs per agent
- Every answer judged by a model from a different provider than the tested agent
- The number of compliant answers out of the number actually measured, with their 95% confidence interval, and the verifiable facts with their quote taken from your site
- The detail of the runs and the protocol, in a single report sent by e-mail
- No correction, no recommendation and no second measurement: you apply your own fixes, Livada does not intervene on your site
Livada AgentRadar — A count, with its limits.
