What Analytics Couldn't Tell Us: Pairing Exit Intercepts with Behavioral Data at Topshop
Pairing exit intercepts with behavioral data at Topshop produced a 5.8% conversion lift that analytics alone could not explain.
TL;DR
Pairing exit intercepts with behavioral data at Topshop produced a 5.8% conversion lift that analytics alone could not explain.
- Problem
- Analytics showed a conversion gap in Topshop search, but could not explain why people were not using it.
- My role
- Analyst / Optimization Specialist
- Team
- Qubit · Topshop
- Timeline
- A/B/n testing cycle
- Methods
- Exit intercepts, funnel analysis, and A/B/n testing.
- Outcome
- +5.8% conversion from the winning search design. Search users converted about 10 times higher, which prioritized the work.
- Decision it informed
- Test visibility and affordance copy in four variants against the existing design, then ship the winner with a quantified case.
Overview
How intercept survey feedback and behavioral analytics were triangulated to find a hidden conversion lever, and how a four variant experiment turned it into measurable revenue.
- Role
- UX Analyst & Experimentation Designer
- Platform
- Qubit
- Methods
- Exit intercepts, funnel analysis, and A/B/n testing
- Outcome
- +5.8% conversion from search; 9–11% cumulative on product pages
My contributions on this project
- Variant UI Design
- Designed the search-field treatments, a border and new placeholder copy, as variations that ran in Qubit. Each cell isolated or combined one hypothesis, so a win could be traced to its mechanism.
- Test Architecture
- Set up the test in Qubit: the existing design as control, cells that isolated or combined border and copy, an even split across all cells, conversion as the primary metric, and a rule to reach statistical significance before any rollout.
- Visitor Opinion Survey
- Configured the exit intercept in Qubit's Visitor Opinion tool: open-ended, fired in-session at points of friction and on exit intent, with frequency caps on who saw it and how often.
- Product Page Micro-Changes
- Ran the same intercept, hypothesis and controlled-test loop on size selectors, delivery tabs, confirmation pop-ups and add-to-bag buttons, without committing IT resources to anything unproven.
- Analysis & Reporting
- Read each variant against control, then took the winner and its business case into IT prioritization for the sitewide rollout.
01: The Problem
Numbers tell you the scale of a problem. Feedback tells you how to deal with it.
Topshop's search bar sat on every user's path to purchase, but it was underperforming. We knew something was wrong from the behavioral data alone: search adoption was low relative to comparable retailers, and sessions that never touched search converted at a fraction of those that did.
Indexed so that sessions without search equal 1. Sessions were segmented by whether they used search, so the gap is a correlation: it priced the problem but could not explain it.
What the quantitative data couldn't tell us was why search adoption was low. Was it discoverability? Usability? Result quality? Pushing a sitewide redesign on a guess would have burned IT resources on an unvalidated hypothesis. We needed the user's voice before touching the interface.
02: The Analysis
Triangulating intercept feedback with behavioral data
Qualitative Signal: Exit Intercepts
Using Qubit's Visitor Opinion exit feedback tool, which was an on site intercept survey triggered as users left, we captured open ended feedback at the moment of abandonment. A recurring theme emerged: users struggled to both find and use the search bar. It read as a low contrast text element with no visual affordance signaling "type here".
Quantitative Signal: Analytics
Qubit's behavioral analytics quantified the stakes: visitors who used search converted roughly 10 times higher than those who did not. Search was not a convenience feature; it was the single highest intent behavior on the site, and friction there was directly suppressing revenue.
The intercept feedback gave us the diagnostic; the analytics gave us the business case. Neither alone would have justified the work, but together they made it the obvious next test.
What the intercept responses pointed to
The search bar didn't look like something to type into
It read as a low-contrast text element, with no visual cue saying "type here".
Paraphrased response “I can't find the search bar.”
Action taken: It became H1, visibility: test a border around the field.
The placeholder copy didn't invite a search
The second cause the feedback pointed to: the placeholder text did not invite action.
Action taken: It became H2, invitation: test new placeholder copy.
Responses are paraphrased. No names or identifying details are kept.
I configured the intercept in Qubit's Visitor Opinion tool. It fired in-session, at points of friction and on exit intent, and it was open-ended: we didn't yet know the hypothesis space, and fixed answer options would only have echoed our own guesses. Targeting mattered as much as the question. Frequency caps decided who saw it and how often, so heavy visitors weren't oversampled.
03: The Decision
Test the hypotheses, do not ship the guess
The feedback pointed at two candidate fixes: visibility (the field did not look like an input) and affordance copy (the placeholder did not invite action). Rather than allocating expensive IT resources to push a full sitewide change based on either hypothesis, we designed four search variations testing both dimensions: changes in copy, and the addition of a border to the search box.
The test was split evenly across all users, with the existing design as control, so any change would carry statistical evidence into the IT prioritization conversation, not opinion.
- H1Visual affordance to border
- H2Invitation to copy change
- H3Combined effect
| Existing copy | New copy |
|---|---|
| No border | |
| ControlThe existing design. Every variant was measured against it. | Copy onlyTests H2, invitation, on its own. |
| Border added | |
| Border onlyTests H1, visibility, on its own. | Border + copyTests H3, the combined effect, with a longer and a concise placeholder.Variant D, the concise one: +5.8% conversion, shipped sitewide. |
Traffic split evenly across all cells, with the existing design as control. Run to statistical significance before rollout.
The variants ran in Qubit as JS/CSS variations on the live site, so testing the hypotheses didn't need IT to build a sitewide change first. IT came into the conversation afterwards, with a winner and the evidence for it.
04: The Execution
Four variants, one control, evenly split
Explore the test cells below. Each variant isolates or combines the two hypotheses so the winning treatment could be attributed to a specific mechanism, not just the new one.
Design decisions I made
The calls that shaped the test, and the reasoning behind each.
Why four variants, not two
A single A/B test of a new search bar would only have shown that something changed, not which element drove it. Isolating border and copy in their own cells, and combining them in others, meant a win could be traced to its mechanism rather than to novelty.
Why an in-session exit intercept
We needed to hear from people at the moment search let them down. So the survey fired in-session, at points of friction and on exit intent, and stayed open-ended, because fixed options would only have echoed our own guesses.
Why an even traffic split
Every cell, control included, got an equal share of traffic. Visitors were assigned at random and kept in the same cell for the session, so each variant was compared with control on equal terms.
Why test in Qubit before asking IT
Qubit ran the variants as JS/CSS variations on the live site, so we could show which change worked before asking for IT time. The sitewide build came after, with a measured result attached.
05: The Impact
Evidence, not opinion, went to the roadmap
- +5.8%conversion upliftVariant D against control. It reached statistical significance before rolling out sitewide.
- ~10×higher conversion for search usersThe behavioral signal that made search worth testing.
- 9–11%cumulative uplift on product pagesThe same intercept-to-test loop, applied to size selectors, delivery tabs, confirmation pop-ups and add-to-bag buttons.
The winning variant shipped sitewide with a quantified business case attached. Just as importantly, the method stuck: the same intercept feedback to hypothesis to controlled test loop was applied to a series of product page micro changes, size selectors, delivery tabs, confirmation popups, and add to bag buttons, which together added a cumulative 9 to 11% uplift, without committing IT resources to anything unproven.
What this project taught me
Intercepts earn their place at the moment of intent
Intercept feedback captured intent, not general attitude. Targeting logic, who sees the survey, when, and how often, determined whether the signal was usable at all.
Triangulation is what makes feedback actionable
Qualitative feedback without behavioral data is a mystery. Joining exit responses with behavioral analytics turned a vague complaint into a prioritized, sized, testable hypothesis.
Test both factors at once, not one after the other
Putting border and copy into one test, each isolated and then combined, showed which mechanism mattered. Testing them in turn would have taken two rounds and still left the combination untested.
Small tests protect big resources
The business case required to settle an argument should always run before engineering commits. Evidence travels further with stakeholders than any deck of opinions.
Case study conducted with Qubit for Arcadia Group. Metrics as publicly reported.