Skip the haystack. Start at the struggle.

Session replay, ranked by need

Hundreds of recordings. Which one do you open?

A list sorted by time is good for storage. It is bad for choosing what to watch. So you play recordings fast, or you filter for what you already suspect, such as rage clicks or errors. A problem that nobody suspected stays hidden.

Watch is a tab on Sessions. It ranks the judged sessions in your date range and filters. For each session it judges, Sequential reads a short text summary of what happened: pages, clicks, and error counts when they are captured. It never reads the recording.

Sequential answers your project’s own questions about the session. What did the person come to do? How did it end? How hard was it? Each answer says how sure it is.

Watch puts sessions that broke for the person first, then sessions where the person struggled, then the rest. In each group, the session with the highest Watch score comes first.

You review the ranking. You do not build it.

Example: Northlight Outfitters, a fictional demo shop

On Tuesday morning, the top row says Watch now. The sentence is "Severe struggle — probably switched to support". The intent is "wanted to buy or subscribe · 99%". The signals are Rage clicks, Form left unfinished and Went in circles.

Watch from 0:11. A shopper types a postcode with a space, and the shipping form rejects it. They type it three more times. Then they open the help chat.

Sequential had already counted the failed request behind it as a finding: 38 people. In plain words, people get stuck at the shipping address. The replay shows one of them.

Nobody looked through 400 recordings. One person watched one replay, for less than a minute.

How it works

1. 01

   Judged once, after the session ends.

   The summary holds the events in order. It also holds the device, the browser, the entry page, the duration, and error counts when they are captured. Sequential never watches the recording.

2. 02

   Your questions, your weights.

   The answers blend into one Watch score. By default, the score uses friction, errors, likely product bugs and "worth a look". Owners and admins edit the questions and set what counts most. A weight change re-ranks every judged session at once. Nothing is judged again, so the change costs nothing.

3. 03

   Press play near the moment.

   Watch from m:ss opens the replay close to where the person struggled. If there is no struggle, it opens where they gave up. If there is neither, it opens at what they came to do.

Doubt is a result, not a gap

Watch shows its doubt.

• It hedges. "Probably" and "possibly" stand for real ranges. Signals under 40% stay off the row.
• It never shows 0% or 100%. Percentages stop at 1% and 99%.
• It says when it cannot tell. Sometimes Watch is not sure what the person came to do or how it ended, and no signal fired. Then the session goes to its own lane: too close to call. The lane is collapsed below the list.
• It shows its track record. Each row can end in thumbs up or thumbs down for one question: was this session worth watching? The answers add up to "Right X of Y you checked".
• Calibration. It compares how often your team said yes with how sure Watch was.
• A thumb changes no answer. A thumb is saved and counted in your score. It never re-judges a session and never changes the ranking.

Confidence that you can check is better than confidence that you must trust.

Your agent reads the same list

Ask your agent: "What were the worst sessions on Northlight yesterday?" The agent gets the top rows, with the doubt included. It gets a replay link when there is a recording.

In the fictional example, Northlight’s agent filed WEB-231 in their tracker. Sequential ranks the sessions. Your agent does the follow-up work. You decide.

For developers: any MCP client with the sessions:explore scope can call list_watch.

Good to know

• It reads events, not video. Unlabelled clicks read as "element". Mislabelled events can make it confidently wrong.
• It will be wrong sometimes. There is no published accuracy figure yet. A 90% answer should be wrong about one time in ten. A Maybe badge or a missing signal does not mean that nothing went wrong.
• Near the moment, not always on it. The Watch from time counts from the session’s first event, not from the start of the recording.
• About half an hour behind. A session usually shows 30 to 35 minutes after its last event. It takes longer on busy projects.
• A daily allowance. Each project has a daily judging allowance. When it is spent, new sessions wait. It resets at midnight UTC.
• Recent sessions only. Judging needs a session id, at least 3 events and a last event within 7 days. System events do not count. Older history is never judged automatically.
• Rewording applies to new sessions. A reworded question applies to new sessions only.
• On the Dashboard. The Worth watching tile shows the top three rows of the last 7 days, with the same badges.

Questions