Filtering Results

The aggregate result of an experiment can hide important patterns. A flat overall result might mask a strong mobile win and a desktop loss that cancel each other out. The filter bar on the Results tab lets you look at the same test for one slice of its visitors at a time.

The bar has five controls: Devices · Visitors · Exposure · Advanced · Dates. Every panel below the bar follows them, with the exceptions listed under What the filters do not change.

Devices

All devices, Desktop, Mobile or Tablet. The device is recorded when a visitor is bucketed: phones are detected from the browser’s User-Agent, tablets from the User-Agent or a viewport between 600 and 1024 pixels wide, and everything else is desktop.

Device views are precomputed alongside the main view, so switching between them is quick, and refreshing an unfiltered view refreshes them too.

Visitors

All visitors, New visitors or Returning visitors. The pixel decides which a visitor is once per browser session: a visitor is returning when their durable visitor id already existed before the session began, and new otherwise. A first-time visitor stays new on every page of that first visit and flips to returning on their next session.

The visitor type has only been recorded since the feature shipped, so a segment view sees a subset of the test’s events. The note under the bar says since when: Audience: New visitors — segment data since Sep 4, 2026. A segment where any variation has fewer than 100 sessions shows a Low sample in this segment warning and withholds the winner.

Exposure

All visitors, Single-variant only or Mixed only. This control only appears when the test’s stickiness is per session or per page, so that one visitor can be shown more than one variant over the test’s lifetime, or once such visitors exist.

• All visitors counts a mixed visitor once in every variant they saw, and credits each conversion to the variant shown at the time. A banner says how many visitors saw more than one variant. It shows the first time you see it and is listed in the Alerts tab after that. Close it with × and it moves to the Alerts tab’s Dismissed (n) group; choose Show again there to bring it back. The low-sample and out-of-date warnings close the same way and return when their condition comes back. Dismissals are remembered per experiment in this browser.
• Single-variant only keeps the visitors who saw exactly one variant: the clean comparison. Panels that cannot be narrowed this way carry an All visitors tag.
• Mixed only replaces the whole body with a descriptive table of the visitors who saw more than one variant, grouped by the set of variants they saw. There is no statistical verdict, because these visitors can’t be credited to a single variant, and the other controls are disabled while it shows.

The mixed set is defined over the test’s whole lifetime, so the Devices, Visitors and Dates filters never change who counts as mixed. Exposure is not remembered: opening another experiment starts on All visitors again.

Advanced

Advanced opens a popover of conditions on visitor attributes: Traffic source, Browser, Operating system, Language, Sessions, Landing page and UTM campaign. Each condition is an attribute, an operator and a value; rows are joined with and or or, and you can stack up to 20. Nothing changes until you click Apply filters; Clear all drops every condition at once.

Once applied, the control shows the number of conditions and the bar explains that attribute data is only recorded from the day the pixel began collecting it: Filtered results include data from Sep 4, 2026. A rule set that matches nobody shows No visitors matched these conditions instead of zeroed cards, because older sessions never match. Advanced filters are not remembered between visits.

Dates

Whole test, Last 7 days, First 7 days or Custom range…, which opens a calendar. Presets are whole days in the workspace time zone, clamped to the day the test started and either today or the day it ended. The trigger shows the span you picked, for example Sep 4 – Sep 17.

The date range is remembered per experiment on this browser, so the next time you open the same test it starts where you left it.

What the filters do not change

Three panels always show the unfiltered data, and wear an All visitors tag when Single-variant only is selected:

• AI Analysis, which reads the whole test.
• The ShopifyQL cards (Order Insights, Geographic Breakdown, Product Performance), which are store-wide figures.
• Advanced DOM Insights, which has its own Exposed sessions only switch instead.

Only views with no visitor segment, no Advanced conditions and the whole-test date range are cached. Any other view is computed live every time, so it can take a few seconds to appear.

On a narrow screen

The bar always stays on one row. When there isn’t room for every control, Devices, Visitors, Exposure and Advanced fold behind a Filters button that shows how many are active, while Dates and the refresh stay in the bar; on a phone Dates folds into the panel too. Tap Filters to open the panel, pick a value and the panel folds again. Clear filters at the foot of the panel puts every control back to its first choice and the dates back to the whole test.

Reading a filtered view

Every filter shrinks the sample, and the engine’s ten-conversions-per-variation floor applies to the view you are looking at, not the whole test. A filtered view that reads Not enough data yet is telling you exactly that.

What to look for:

• A consistent lift across devices and visitor types is the cleanest result. The variant works for everyone.
• A big lift on one slice, flat elsewhere is a partial win. Consider applying the change only where it works, using device targeting on a re-run.
• Mixed direction (Variant B wins on mobile, loses on desktop) means the aggregate averages out a real opportunity, or a real risk. Investigate further.
• One slice with very few visitors has a fragile probability. Don’t read too much into a “winner” with 50 mobile visitors.

Next Steps

• Run an experiment scoped to a single device class up front: Device Targeting.
• Push variant data into GA4 for segmentation the bar does not offer, such as country: Google Analytics 4.
• Make a confident decision once the slices line up: Declaring a Winner.
• Limit an experiment to the visitors you care about before it runs: Audience Targeting.