Motorbike marketplace: how we counted, and how to set it up
For the analyst who wants to know how a result was read when the sale closes offline, what counted and what did not, and what was switched on in 22 days.
- Every definition behind the numbers on the case study
- The setup in the store, step by step
- What to check in your own data
Open the method and setup notes
Enter your work email and the notes open on this page. We also send you the link, so you can come back to it.
Thanks. The notes are open below.
We use your email to know who reads our research. Privacy policy
What was measured
Period and population
A two-sided used-motorbike marketplace in Malaysia and Vietnam, with about 500,000 page views a week. Two equal 22-day windows: a measured baseline from 4 to 25 June 2026, and Releva active from 26 June to 17 July. The channel shares use 1 July to 6 August. The event-level data was checked with the client.
What counts as a result
A booked test ride. Test riders per 1,000 visitors went from 1.65 to 2.32 (+40%) on 20% fewer visitors; bookings from 182 to 203 (+11.5%); WhatsApp intents per 1,000 visitors from 22.8 to 27.8 (+22%). In 22 days 5,048 people showed booking intent on WhatsApp and 400 booked; 81% of the intents were anonymous.
How an order is credited
There is no order. A click counts only if that person then books a test ride; WhatsApp intents carry no value in the report. First-party identity across devices ties the booking to the click. One in eight test rides was for the exact bike clicked in a recommendation, a median of about nine hours earlier.
Controlled or observed
Before and after, not a holdout. Rates are per visitor, so the result holds while traffic moves. The 6.5× test-ride rate for product viewers who saw a recommendation is exposed against not exposed; they chose to look, so part of that gap is intent.
What the numbers do not say
Recommendations reached about 4% of page views, and 1.9 million searches in nine weeks were not personal yet. Daily attributed value was 45% higher after the proof window than during it, as the model retrained.
The terms on the case study
| Term | How it is defined here |
|---|---|
| Test riders per 1,000 visitors | Test-ride bookings divided by visitors, times 1,000: 1.65 in the baseline, 2.32 with Releva active. |
| WhatsApp intent | A shopper showing booking intent in a WhatsApp chat. Counted, but worth zero in the report. |
| Attributed value | The value of the bikes on test-ride bookings that followed a Releva click: recommendations 47%, banners 31%, triggered email 19%, web push 3%. |
| Exact bike | A test ride booked for the same bike the buyer clicked in a recommendation: 1 in 8. |
| Exposed | A product viewer who saw a recommendation; they booked at 6.5 times the rate of viewers who did not. |
| Repeat share | Repeat customers' share of weekly revenue: from 28% to 35% over five weeks. |
What was switched on, in order
A measured baseline first
Twenty-two days, 4 to 25 June 2026, with nothing new switched on.
Only bookings count
First-party identity across devices, and attribution set so that only test-ride bookings carry value.
Recommendations on live listings
On product pages and in banners, retrained on the marketplace's own data in July.
Triggered email
Abandoned-browse and abandoned-sell emails replaced bulk sends to an old, quiet list.
Web push from zero
A way to reach anonymous visitors again: 7,315 subscribers in 22 days; push sends went from 1,093 to 14,763.
The read-out
Equal 22-day windows, rates per visitor, and the bike-level match between click and booking.
To check it in your own store
Pick the one event that is your sale (a booking, a call, a form) and count it per 1,000 visitors over equal windows before and after a change. A rate per visitor holds when traffic moves; a raw count does not.
Give clicks a value of zero unless that person then produces the event, and check how many events match the exact item clicked. That match is the part of the result no one can argue with.
See which of last week's visitors you missed
Thirty minutes, with your store and ad accounts open. Then 30 days free. A holdout group decides: if the agent doesn't add orders in 30 days, you don't pay.
Book 30 minutes
You'll talk to one of us.