AI Replay Scanner
Describe what to look for in plain language. The scanner reads each new recording, scores it, and when it finds real friction it files a feedback item with the replay attached at the worst moment. From there it clusters, counts toward demand, and can go to an AI-drafted pull request.
Nobody watches session recordings. The frustrating sessions are in there, rage clicks and all, and they scroll away unwatched.
Session replay tools solved capture years ago. What nobody solved is attention: a product with even modest traffic produces more hours of recording per day than anyone will ever watch, so the recording of the user who hit the same error five times and left is sitting in a list, unwatched, labeled only by duration. AI scanners are the obvious answer and the analytics tools are starting to ship them, but their version emits an analytics event: a score you can chart, a summary in a digest, and it is on you to turn any of it into a bug report, a backlog item, and eventually a fix. The distance between "a scanner noticed something" and "someone is fixing it" stays exactly as long as it was.
How it works
Disclosure: I build Usero, so weigh that. The scanner here is built to end in the inbox, not on a chart. You write a plain-language prompt ("watch for users failing to complete checkout") and set a score threshold. Each new replay is turned into a deterministic activity transcript, every click, error, rage-click run and navigation with typed values redacted, and the AI reads that transcript, never a video it can narrate loosely. Every moment it reports must cite a real transcript entry or the claim is dropped before you see it, the same anchor-or-drop rule our evidence quotes use. Above your threshold, the finding files as a feedback item with the replay attached and a marker at the worst moment. It clusters with related feedback, counts toward demand, and can trigger an AI-drafted pull request. You review everything; nothing auto-merges.
Every claim in a finding is anchored to a real recorded event and seeks the player to that moment. If the model cannot point at the event, the claim is dropped server-side. No invented clicks, no plausible narrative dressed up as observation.
A scanner result above threshold is a feedback item in your inbox, with the replay attached at the friction moment. It behaves like any other feedback: clustering, demand counts, PR eligibility. Not a metric that needs a human to re-type it into an issue tracker.
Scanning runs on the same flat-rate AI pipeline as the rest of Usero. No per-observation credit meter, no per-PR fee on acting on what it found.
A session with no friction produces no finding, and a replay the scanner has not read says so in one plain line. No confidence percentages that are not calibrated against anything.
The honest objection
Because this one is not allowed to narrate. The model never sees pixels; it reads a transcript of recorded events extracted deterministically from the replay, and every moment in its finding must reference a specific transcript entry. Claims that do not anchor are dropped before display. We built it this way after watching a competitor scanner report a button click that never happened, at 100% stated confidence, and repeat it in three surfaces. The transcript is also stored with each scan, so you can check exactly what the model was shown.
FAQ
A deterministic activity transcript extracted from the rrweb recording: clicks with their targets, rage-click runs, navigations, console errors, scrolls and idle gaps. Typed input values are redacted; only field labels are included. The transcript is capped in length and stored alongside the scan result for audit.
When the score meets the threshold you set (default 7 of 10). The feedback item carries the scanner summary, the anchored moments, and the replay linked at the first friction moment. Below the threshold, the scan result stays on the session page and nothing is filed.
On every new completed replay for clients that enabled the scanner, via a periodic sweep, plus a "scan now" button on any session. One scan per replay; re-scans do not duplicate feedback.
Not directly. A filed finding is ordinary feedback, so it participates in clustering and the demand threshold like anything else in the inbox. The path to a PR always runs through the same review gates as the rest of Usero, and nothing auto-merges.
Free tier. No credit card. Two-minute install. The AI opens the PR, you merge it.
Keep reading