Skip to main content
The results page always shows every visitor in the test. Sometimes the question you really want to answer is narrower: “does the winner still win on mobile?” or “is the lift real for paid traffic or is it organic?” Segments narrow the whole report to that slice.

​
Adding a segment

On the results page, click Add segment. Pick a dimension and a value. The rest of the page recomputes, verdict, per-variant performance, confidence chart, speed panel, for the filtered slice. You can add multiple segments; they combine with AND. “Mobile” plus “United States” gives you mobile US visitors only. The filter also shows up in the URL. Copy-paste the URL into Slack and your colleague sees the exact same slice.

​
The dimensions

  • Device, Desktop, Mobile, Tablet.
  • Browser, Chrome, Safari, Firefox, Edge, Opera, others.
  • Operating system, Windows, macOS, iOS, Android, Linux, others.
  • Country, pulled from actual traffic on this experiment; countries with no visitors do not appear in the picker.
  • New vs returning, 30-minute session gap.
  • Traffic source, direct, organic, paid, social, referral, other.
  • Entry path, the first URL the visitor landed on, bucketed.
  • UTM campaign, pulled from actual traffic (data-driven picker).

​
What “the whole report” means

When you filter, everything on the page recomputes for the filtered slice:
  • Visitors count is the count in the segment.
  • Verdict, primary metric, guardrails, all recomputed on segment-only data.
  • Confidence chart plots the segment’s history.
  • Speed panel shows the segment’s speed data.
  • Segments cascade, if you filter by Mobile and then by US, the visitor pool is mobile US visitors, not mobile OR US.

​
Sample size shrinks

A test with 100,000 total exposures filtered to “France, tablet” might have 300 exposures. The verdict copy honors that, a segment that lacks the volume for a call says so, even if the whole test has a clear winner.

​
What to do with a segment win

A finding like “variant B loses overall but wins on mobile” is real information, but it is not a license to ship the variant to mobile only. Two reasons:
  1. Segment sample sizes are smaller, so segment verdicts are less reliable. A confident segment finding on 300 visitors is less trustworthy than the overall finding on 100,000.
  2. You did not pre-register the segment. The p-value on a post-hoc segment is inflated relative to a segment you named before the test began.
Take the finding as a hypothesis for the next test, not the answer to this one. “Mobile visitors might respond to B” is a great next experiment; “ship B to mobile now” is jumping the gun.

​
Segments and sequential engines

If you are using the sequential engine, segments still get an anytime-valid confidence sequence, you can peek at segment results as often as you like without inflating the false-positive rate. Frequentist segment reads have the usual multiple-comparisons caveat.

​
Live-shareable URL

Every segment filter serializes into the URL as filter.<dim>=… parameters. That means:
  • Share a slice in Slack, colleague sees the same view.
  • Bookmark a slice, comes back with the filter applied.
  • Post in a doc, the reader jumps straight to the slice, no clicking.

​
Data-driven pickers

Country and UTM campaign values are pulled from the traffic this experiment has actually seen. If your test has zero visitors from Brazil, Brazil does not appear in the picker. That prevents “the filter that filters everything out” as a class of confusion.