Metrics tell you what. Recordings show you why.
Confidence Session Recordings replays real user sessions with the flags and variants each user saw. Record any audience you choose, or record straight from an experiment to compare variants. The Confidence Agent watches those sessions and reports the bugs, friction, and opportunities behind a metric change.
Add image here (HERO)
A session recording in Confidence with the experiment variant and an AI finding shown next to the replay.
Unlimited recordings on paid plans. 100,000 a month free on the free plan. You only pay for the recordings the Confidence Agent analyzes.
A guardrail moves, and the dashboard can't tell you why
Your experiment wins on its primary metric, but a guardrail metric gets worse. A guardrail metric is a metric you watch to make sure a change doesn't cause harm somewhere else, such as cart abandonment or error rate. Now you have to explain the regression before you can ship.
The numbers tell you which step got worse, but not what users ran into there. Finding out usually means days of slicing dashboards, adding logging, and guessing. Watching sessions would answer it faster, but a standalone replay tool doesn't know which variant a user saw unless you send it yourself, so comparing treatment with control takes extra work.
Add image here (ILL-1)
Illustration: an experiment result with a winning primary metric and a red guardrail metric, with a question mark over the guardrail.
How do Session Recordings work with feature flags and experiments?
Session Recordings is built into Confidence, so every recording shows the flags and variants the user saw. Open any recording to see them, or filter the whole list to one variant and compare it with another.
Record straight from your experiments
You decide what to record using the experiments and rollouts you already have. Recording rules choose who gets recorded; they never change which variant a user receives.
A/B test rule
Rollout rule
Conditional targeting rule
You watch the sessions behind a result, split by variant, without setting up a separate replay tool.
Add image here (B)
The Recordings list filtered to sessions in the treatment variant of a checkout experiment.
Add image here (C)
An A/B test recording rule that records 10% of sessions in each variant of an experiment.
What does the Confidence Agent find in recordings?
The Confidence Agent watches hundreds of sessions and tells you what they have in common. It transcribes each session, then reports patterns across them: bugs, points where users get stuck, and opportunities to improve. Each finding links to the recordings behind it, so you can watch the evidence in a few clicks.
Inside an experiment, the agent compares variants. It tells you what users do in treatment that they don't do in control, which is often why the metric moved.
You can ask follow-up questions, too. The agent has the full transcript of every session it analyzed, so it can answer questions like “Do users who leave try to find the shipping cost first?” and link to the sessions that show it. To check a pattern yourself, search transcripts by meaning: type “surprised by shipping cost” and get the sessions where that happened, even if no event or error was logged.
You go from “the guardrail moved” to a specific cause without watching every session, and you have the recordings to show your team.
Add image here (D)
An AI finding that compares treatment and control sessions in a checkout experiment, with links to recordings from both.
Add image here (E)
A question to the Confidence Agent about a checkout experiment, answered with a count and links to the matching sessions.
Recordings feed Confidence Loop
Session recordings are one of the signals Confidence Loop uses out of the box. Loop turns what users run into during sessions into observations, groups them into issues, and investigates them. Recordings explain one experiment. Loop watches your whole product and fixes what it can.
Narrow hundreds of sessions to the few that explain your result
Filter recordings by the flag variant a user saw, or by signals: behaviors Confidence detects in every session. Frustration signals include rage clicks, when a user clicks the same spot over and over, and dead clicks, when a click does nothing. Error signals include error messages, failed network requests, and console errors. You can also filter on how often a signal happens, such as three or more rage clicks.
Combine signals with page URL, duration, device, browser, OS, country, and custom tags. For example: every treatment session on the payment page with at least one failed request. One toggle narrows the list to sessions the Confidence Agent has analyzed.
Add image here (A)
A recording's detail panel showing the flags and variants evaluated during the session.
Private by default
The Session Recordings SDK masks every input field by default, so typed values like names, emails, and card numbers aren't captured. You can mask or block more elements on top of that, and recording rules let you record only the sessions you need, at the sample rate you choose. Recordings are kept for 30 days, then deleted. AI findings and transcripts stay, so you can still search and read them after the replay is gone.
Add image here (F)
A session replay with form inputs masked.
How do I set up Session Recordings?
01
Install the package
@spotify-confidence/session-recording npm package.02
Initialize the recorder
03
Create a recording rule
04
Open the Recordings tab
Frequently asked questions
How much do Session Recordings cost?
Which platforms do Session Recordings support?
Which flag evaluations show up in a recording?
Do recording rules change which variant users get?
Can I record only users in a specific experiment?
What does the Confidence Agent do with recordings?
How long are recordings kept?
Is user input recorded?
From a red guardrail to a fix
A store tests a two-step checkout against its current three-step flow. Orders go up 4%, but cart abandonment on the payment step rises 9%. The team adds an A/B test recording rule with AI analysis on.
Without anyone filtering, the Confidence Agent reports the difference between the variants: in the new flow, shipping cost first appears on the payment step, and 31 of 40 abandoning users leave within 15 seconds of seeing it. In the old flow, shipping shows in the cart. The PM asks a follow-up and learns that 18 of those users went back to look for the price, then watches five of the linked sessions to confirm.
The team adds a shipping estimate to the cart and reruns the experiment. The conversion gain holds, and abandonment is back to normal.
Read more about Session Recordings
Introducing Recordings
Confidence Session Recordings lets you replay actual user sessions, with AI-powered analysis to surface bugs and insights right from your experiments.
Confidence Loop: Building products that continuously improve
Confidence Loop turns signals from real users into product understanding and improvements, with teams choosing how much of the loop runs autonomously.