Connecting Bixlow with Microsoft Clarity

There's a version of Bixlow where we build session recording and heatmaps ourselves. We decided against it, and the reasoning is worth explaining because it shapes how the integration works.

Microsoft Clarity is free, unlimited, and genuinely good. Rebuilding it would consume months and produce something worse, and it would add a second recording script to your page — meaning two sets of DOM observers, two upload streams and a measurable cost to the page speed you're trying to optimise.

What didn't exist was the connection. Clarity can show you a thousand sessions. It cannot show you the sessions where the visitor saw variant B of experiment 32. That gap is the whole reason quantitative and qualitative analysis stay separate at most companies, and closing it is a small amount of plumbing with a large payoff.

When a visitor is assigned to an experiment, Bixlow tags their Clarity session with the experiment identifier and the variant identifier as custom tags.

From then on, Clarity's own filtering does the rest: you filter by those tags and you're looking at exactly the population that saw one variant. No new UI to learn, no data leaving Clarity, and no duplicate recording.

From the Bixlow side, each experiment result carries a link straight into Clarity pre-filtered to that variant. You read the number, you click, you watch the behaviour behind it.

You need Clarity installed on your site and its project ID. In Bixlow you paste the project ID into the project settings and enable the integration. That's the configuration.

Two implementation details worth knowing.

The tagging is best-effort: if Clarity hasn't loaded when the assignment happens, Bixlow queues the tag and applies it when the script is ready. If Clarity never loads — the visitor blocks it, or it fails — your experiment is unaffected. The experiment engine does not depend on the analytics integration, which is the correct dependency direction and the reason a Clarity outage can never distort your results.

And the tags are identifiers, not content. We pass an experiment id and a variant id, not page contents, not customer data, not anything a visitor typed. The integration adds no personal data to what Clarity already collects under your own configuration and your own privacy policy.

Diagnosis of losses. The variant lost by 9%. Watch twenty sessions from that variant and the reason is frequently obvious within five minutes: an element overlapping on a specific viewport, a sticky bar covering the price, a tab that requires two taps because the hit area is wrong.

Explanation of neutral results. A flat test with no obvious cause is usually a test where the change wasn't noticed. Recordings show you whether anyone scrolled far enough to see the thing you changed.

Confidence about wins. A variant won and you don't know why. Watching how people used it tells you whether the mechanism was the one you hypothesised — which determines whether the learning generalises to other pages or was specific to this one.

That last one is underrated. A win you understand is worth several wins you don't, because it tells you what to do next.

Recordings are not evidence. Twenty sessions are not a sample, they're anecdotes with video attached, and it is extremely easy to watch five people struggle with something and conclude it's the problem when it affects 2% of visitors.

The correct sequence is always: the experiment tells you whether, the recordings suggest why, and the next experiment tests the why. Using recordings to decide is how teams end up redesigning something because one person on a recording hesitated.

• We integrate with Clarity rather than rebuilding it: it's free, excellent, and a second recorder would cost you page speed.
• Bixlow tags each Clarity session with the experiment and variant, so you can filter to exactly one variant's population.
• Tagging is best-effort and passes identifiers only. A Clarity failure never affects the experiment.
• Recordings explain, they don't prove. The experiment says whether; recordings suggest why; the next experiment tests it.