--- title: "See how AI agents use your product · Agentrail" description: "See which AI agents use your product, what they try to do, where they drop off, and which products they choose instead." canonical: "https://www.agentrail.co/use-cases/agent-analytics" --- Agent analytics for product # See How AI Agents Use Your Product. See which AI agents use your product, what they try to do, where they drop off, and which products they choose instead. [Check your site’s agent readiness](https://www.agentrail.co/agent-readiness/checker) $120K Revenue recovered 2M+ Agent sessions ![A raccoon in a hoodie looking through binoculars, with a note: Found a suspicious drop-off.](https://www.agentrail.co/assets/use-cases/agent-analytics/hero-raccoon.png) The problem ## Your analytics shows a spike. The questions are still open. Product analytics counts clicks and page views from people. Agents don’t click. Many never run JavaScript, so the script never sees them. When one does show up, it looks like a bot, a person, or nothing. 1. Page views · /docs 48.2k312% Last 14 days **Nov 12 · 12.4k views**Who visited: Unknown 1. Nov 1 2. Nov 5 3. Nov 9 4. Nov 14 01 A spike in docs traffic ### How many agents came, and which ones? 2. 02 Sign-ups that don’t match the traffic ### Where did they stop: docs, MCP, sign-up or first API call? Week over week Last week This week Visits312% **11.7k** **Last week**11.7k visits **48.2k** **This week**48.2k visits Sign-ups2% **210** **Last week**210 sign-ups **214** **This week**214 sign-ups **Traffic ×4\. Sign-ups flat.** Where did they go? 3. Events across tools 3 records0 linked 10:00 – 10:15 Analytics **Visited /docs** anon\_8f2c Sign-up **Account created** user\_4417 API logs **POST /v1/events** key\_ab91 **10:14 · API call**Linked to docs visit: No 1. 10:00 2. 10:05 3. 10:10 4. 10:15 03 No shared story from docs hit to API call ### When a developer asks an agent, does it pick you? Why now ## Your next customer might not be human. People see your homepage. Agents see your docs, API, CLI and MCP server, and your analytics can’t see them. - **$50B**At risk if just 1% of agent sales fail - **$120K**Net new revenue at Akto in one month, after fixing agent sign-up Every week you can’t see agents is a week they choose someone else’s docs. ![A raccoon in sunglasses holding a fan of cash, with a note: $5 trillion by 2030? I want my cut.](https://www.agentrail.co/assets/use-cases/agent-analytics/why-raccoon.png) The agent funnel ## Discover, docs, MCP, register, claim, activate. Each stage shows how many journeys reached it, how many stopped there, and the main reason. Crawlers such as GPTBot and ClaudeBot are left out, so the count is agents doing tasks. ### Agent journeys, this week Each bar shows how many agents reached that stage. The red notes show how many stopped, and why. 1. 1,284 2. 940 3. 520 4. 180 5. 112 6. 64 1. Discover Found your product 1,284 agents · 100% ↓ 344 stopped · Answered from memory, never visited 2. Docs Read your docs 940 agents · 73% ↓ 420 stopped · Firewall challenged /docs 3. MCP Used your tools 520 agents · 40% ↓ 340 stopped · Tool call returned an error 4. Register Created an account 180 agents · 14% ↓ 68 stopped · Sign-up needed a browser 5. Claim Approved by a person 112 agents · 9% ↓ 48 stopped · Waiting for a teammate 6. Activate Made a first API call 64 agents · 5% Agent preference ## Are agents choosing you? Agentrail asks agents the questions developers ask. “Set up pull request review for my team. Pick the fastest tool.” It measures how often they mention you, recommend you, and choose you. These runs are simulations, so they don’t need your traffic. - ### A leaderboard Your place among the products agents recommended and chose, and the gap to the leader. - ### Head to head Your record against each alternative, with the agent’s own reasons for passing on you. - ### Memory vs research How often an agent answered without reading a page. Useful when a docs fix takes time to show up. ![Agent pick share for code-review tasks. CodeRabbit leads, then GitHub Copilot, Greptile marked as you, and Qodo.](https://www.agentrail.co/assets/use-cases/agent-analytics/preference-card.png) Agent experience ## Do agents actually get set up? For runs where agents set your product up in a real repository: how far they got, what they claimed against what a check showed, what a person still had to do, and time and tokens per agent. 1. Read docs 2. Installed 3. Configured 4. Integrated 5. Checked ### How far did it get? See the exact step each agent reached, from reading your docs to a passing check. Setup report Claude Code - Installed - Configured - Integrated Status**Complete** ### Did it really work? Agents say setup worked. Agentrail checks your product and shows if that’s true. Waiting on a person Authorize OAuth app CancelApprove ### Where did it need a human? The clicks, approvals and downloads an agent couldn’t finish on its own. In your product**3** - Env var missing from docs - MCP tool returned 500 - Sign-up needs a browser Everything else**2** - Network timeout - Rate limited by agent host ### Whose problem was it? Every failure, split into issues in your product and everything else. How data gets in ## The same sources as journey tracking. Without the website script, agents that drive a real browser look like people. The Command Centre says which surfaces are sending data, and which to connect next. [How journey tracking works](https://www.agentrail.co/docs) ![Code, docs, chat, GitHub, GitLab, Slack, Linear, Jira and other product surfaces feeding into Agentrail.](https://www.agentrail.co/assets/use-cases/agent-analytics/sources.png) FAQ ## Questions teams usually ask. How is this different from Google Analytics, PostHog or Amplitude? Those tools count clicks and page views from people. Agents don’t click, and many never run JavaScript, so the script never sees them. Agentrail reconstructs the agent’s own task: what it was trying to do, the path it took, and where that experience failed. How is this different from agent observability tools? Observability watches agents you run. Agentrail watches agents using your product: the journeys already in your traffic, and simulated questions developers ask when you don’t have that traffic yet. Do you count crawlers? Crawlers such as GPTBot and ClaudeBot are left out, so the count is agents doing tasks. How do you know which agent it was? Agentrail uses request, behavioral, temporal and product-level signals to fingerprint agent activity and stitch related interactions into one task journey. We don’t have much agent traffic yet. Is this useful? Yes. Early journeys show where a product built for people is hard for an agent. Preference runs are simulations, so they don’t need your traffic.