Matching Bikes and Riders with Conversion Attribution in Search

Jan 18, 2024

At Riders Share, we optimize search to find the best match between riders and motorcycles. But once you have the basics like availability, location, price, rating, cancellation rates, reviews and rental history, how do you learn which bike each rider will love most from the data? How do we know that showing a bike in search will cause a rider to book a ride successfully?, our marketplace search partner, predicts the best match by carefully measuring what listing will be selected from a list and result in a sale. That is, showing which bike will most likely cause the rider to book. This is called “conversion attribution.” Attribution is best known from ad tech: when a sale is made, which ads shown get credit for that sale? You may have heard a version of this called “last click attribution” if you’ve ever run a Google ads campaign. Promoted uses the same idea, but with all listings in Riders Share search result. They optimize each listing as if it were a conversion optimization ad, even though it’s not since Riders Share has no ads.

Once we attribute all the bookings to which searches caused them, we use these as “labels” in a supervised machine learning (ML) system to predict which listings will most likely cause bookings in the future. This is called “conversion optimization.”

Getting Technical models the user path to purchase as a Markov model of impression, to click, to attributed purchase, using concepts from ad tech like Google AdWords to optimize marketplace search for conversions. optimizes search ranking in three steps:

  • See it: Riders must see a listing in search to be influenced by it. Sometimes, riders book bikes without searching for them. These don’t count for helping to optimize search. We borrow an ad tech industry standard for “saw a listing” called “IAB Impression,” which is at least 50% listing visibility for at least one continuous second on web and mobile.

  • Choose it: After seeing a listing, the rider must choose it by clicking on it in search. This indicates that the listing description was relevant enough to the query for further consideration. We use another ad tech concept called “MRC click,” which removes duplicates and attributes clicks to a specific search optimization decision. We track what position was clicked and model away any bias from riders clicking the first listing regardless of how good the first result may be.

  • Book it: After the rider has chosen the listing and has carefully reviewed all the details, availability, price, and other options, the rider must book it! Our system calls this a “Purchase” or “conversion.” 

Getting Fancy: Last Multi-Click Attribution

When you book a bike on Riders Share or purchase on any online marketplace or e-commerce site, there is rarely a direct identifier between the “listing view” and the final purchase. We must infer it! needs this connection between “listing delivery decision” (insertion) and the booking (purchase) to generate a training example for ML optimization. uses a user identity graph, contentIDs, and time to infer connections between insertions and purchases.

However, what if there are many potential views and clicks? How do you choose? Do you choose the first click? The last click? What if there is no click? uses a model called “last multi-click attribution.” First, we look back at a fixed period, for example, one week, for any insertions with an impression and a click. When there is only one click, it gets 100% “credit” in conversion optimization. When there are multiple clicks, they all get equal partial credit. For example, if there are 3 clicks in the window, they all get 33.3% credit. “Credit” is used for label weighting in a post-click conversion model used in search ranking. For apps like Riders Share, multi-click attribution is important because riders frequently navigate back and forth through the app, making many clicks on the same listing. Because uses real-time user engagement features like Checkouts, by the last click, it may be too easy to predict which bike the rider will book — the one in their cart! However, the first click was the click that started the path-to-booking for this rider. It needs some credit, too! saw a 1-3% increase in total purchases from moving from last-click to last-multi-click attribution labeling.

A diagram of how joins conversions like Purchases to past search decisions using last multi-click attribution

Learn More is always improving its unified search and ads ranking optimization algorithms to help match buyers and sellers for the top online marketplaces and scaled e-commerce. Learn more about

As a Host, what can I do to optimize my search rankings?

The main takeaway: your bike needs to be the most desirable for most people for it to rank at the top more often. The search engine will rarely show the same results twice, though, with past searches influencing newer ones. Different people will also get different results optimized for them, and the combination of bikes shown matters as much as the individual bike's performance.

The Performance section of the menu offers several metrics you can focus on to influence search results. Optimize these, and your bike should naturally rank at the top:

Thanks for riding with us!