# Watching the recordings of people who saw variant B > A generic link to your recordings tool is useless. Being able to watch only the people who received the losing variant is where the answer actually is. - Source: https://bixlow.com/en/blog/watch-sessions-from-variant-b - Language: en - Published: 2026-10-25T09:00:00.000Z - Updated: 2026-10-06T09:28:18.400Z - Author: fernando - Topics: session recordings, Microsoft Clarity, experiment design, A/B testing, mobile CRO, CRO, experiment metrics - Publisher: Bixlow --- Here's a result: experiment 32, sticky add-to-cart on mobile. Add-to-cart +8.4%. Purchase: no significant difference. Revenue per session: flat. What do you do with that? The primary metric moved, the guardrail didn't degrade, but the thing you actually wanted — more revenue — didn't happen. The numbers have told you everything they can. The rest is in the sessions. ## Why filtering by variant is the whole point Without variant tagging, opening Clarity gives you every session on the site. Most of them are control. The behaviour you want to understand — what people did with the sticky button — is diluted among thousands of sessions where the sticky button didn't exist. With tagging, you filter to experiment 32, variant B, mobile, and you have the 237 sessions where the thing happened. Now twenty minutes of watching is worth something. This is the difference between "we have session recordings" and "we can see what our experiment did". Most companies have the first and think they have the second. ## A workable routine **Watch ten sessions that converted and ten that didn't.** Not twenty random ones. The contrast is where the information is: if converters and non-converters behave identically up to the point of decision, your change isn't where the difference is being made. **Watch at 2x and skip the idle parts.** You're looking for hesitation, repeated scrolling, taps that don't do anything, and rage clicks. Clarity surfaces those automatically, which is the fastest place to start. **Write down what you see before interpreting.** "Six of ten scrolled past the button without using it" is an observation. "People don't like sticky buttons" is a theory you've attached to it. Keep them separate, because the observation stays true and the theory usually doesn't. ## In our example Watching variant B sessions for experiment 32 might show something like this: people tap the sticky button, the cart drawer opens, and they immediately close it and keep scrolling. They're using it to check the price, not to buy. That reframes the whole result. Add-to-cart went up because the button made an action cheap, not because more people decided to purchase. The extra carts were exploratory. Which explains exactly why purchase didn't move. And it generates a real next hypothesis: if people are tapping the button to see the price, the price isn't visible enough when they decide. That's a different experiment, with a different mechanism, that you'd never have arrived at from the numbers alone. ## Compare against control The single most useful comparison is the same routine run on control sessions. Ten from B, ten from A, same page, same device. Differences that look alarming in isolation often turn out to be how everybody uses your site. People scrolling past your CTA is only meaningful if they scroll past it *more* than they do in control. This is the step people skip, and skipping it is how a recording session turns into a list of things that have always been true about your site being blamed on your variant. ## Don't let recordings overrule the experiment Watching people struggle is vivid and persuasive in a way that a confidence interval is not. That asymmetry is dangerous: a team that has just watched five painful sessions will override a statistically solid result without noticing they've done it. The hierarchy holds. The experiment decides whether the change helped. The recordings explain the mechanism and generate the next hypothesis. If the recordings contradict the numbers, you've found something worth testing — not a reason to disbelieve the measurement. ## Key takeaways - Filtering recordings to one variant is what turns session replay from decoration into diagnosis. - Watch ten converters and ten non-converters, not twenty at random. The contrast carries the information. - Always run the same routine on control. Most alarming behaviours are just how your site works. - Recordings are vivid and non-representative. They explain and generate hypotheses; they don't overrule a result. --- ## Meet our AI CRO agent What you just read, put to work by an agent: it understands your visitors, decides what to test or personalize, builds it on your site without a developer and tells you when a result is real. https://bixlow.com/en/ai-cro-agent