From booked rides to real margin

Key figures

On a mature acquisition account, switching from booked-ride bidding to real-margin bidding produced a double effect, isolated in an AB test and then confirmed at the scale of the whole setup.

Allocab operates in a French ride-hailing market dominated by international platforms that heavily over-invest in paid acquisition. Competitive pressure on Search is structural and growing: the category's advertising investments in France have increased year after year since 2022, and Allocab's competitors spend substantially more than it does on the channel. The French ride-hailing market keeps growing in ride volume (+16% in 2024 according to ARPE), but for a challenger, every position on Search costs more to defend each year.

For a player whose acquisition still hinges on Search, the challenge is twofold: holding ground against competitors who keep outspending, and making every euro invested produce more value. When you cannot win the volume war, you win the value-per-conversion war. This is precisely the trajectory we have been driving with Allocab since late 2024.

This note covers a structural switch made in spring 2026: moving from bidding on booked rides to bidding on the real margin of each completed ride. Method, results, and what we take away for other mature acquisition accounts.

Until early 2026, Allocab was bidding on the booked ride: the most natural conversion event to track, the highest-volume one, the one that gives the Smart Bidding algorithm the most signal to learn from.

But bidding on booked rides means bidding on an intermediate event, not on real business value. The account had reached the limit of what that signal could deliver, for three operational reasons.

We designed the switch in three phases, with a rigorous test protocol before any account-wide rollout. Methodological discipline is what made the effect measured and reproducible.

Four months of upstream preparation: sending, for each ride, the margin and the completion status from Allocab's back office to BigQuery, enriching that signal with Google Analytics 4 (gclid and consent status of each ride), then feeding it back into Google Ads through the Data Manager API as a value-based offline conversion import; the margin of each completed ride sent as conversion value, matched on hashed email and gclid. On the site side, server-side tagging and advanced Consent Mode secure a durable, consent-compliant measurement. The signal becomes usable as a conversion value by January 2026: Smart Bidding now has a measure of business value per conversion, rather than a mere counter of intermediate events.

This stage is more demanding than it looks. Bidding on completed rides means bidding on an event that happens later in the funnel, in lower volume, with extra loss points: rides not honoured, CRM data not sent back in time, attribution gaps between the booked and the completed event. The error tolerance is very low: any hole in the feed immediately degrades learning.

Controlled test launched on 10 February 2026 on four mirrored campaigns: two test campaigns duplicated on Margin tROAS bidding, two control campaigns kept on the existing tCPA. Three successive measurement windows over eight weeks, with cross-checked sources: Google Analytics 4 (volume and cost) and Allocab's CRM (real margin and completed rides). At each window, we read the gap across four dimensions: volume, acquisition cost, unit margin and total margin.

Across the first two windows, the test lagged the control on almost every metric in absolute terms. This is where discipline matters most: not concluding too early. When you ask the algorithm to optimize an event it has never targeted, the learning phase is mechanically long.

What we observed over the eight weeks of testing:

• Windows 1 and 2. The control stays ahead on most indicators, but the test's trajectory shows clear learning, with the gap narrowing at every measurement.
• Stabilization. After a few weeks, parity between control and test on the key indicators.
• Window 3 (verdict). The test pulls clearly ahead. At comparable spend across the four aggregated campaigns, +42% margin per completed ride and +19% total margin, while spending less than the control.

On the strength of the test verdict, the full switch was carried out in April 2026. The switch was also an opportunity to simplify the account structure. With a unified margin objective, the case for multiplying campaigns by segment type collapses: margin being intrinsically comparable, the target is unified across the account. Everything can be consolidated into large Search and PMax campaigns, with segmentation at ad-group level, which mechanically simplifies the algorithm's learning.

In practice, the migration followed two principles.

Expected and observed effect: the later waves learned much faster than the initial AB test. Where the test had needed nearly eight weeks to pull clearly ahead of the control, the switch waves on the rest of the account stabilized within two to three weeks. The reason is mechanical: the account had by then accumulated enough historical data on the downstream event (completed ride and its margin) for the algorithm to no longer start from scratch at each wave. The upstream signal-reliability work pays off a second time here, by speeding up subsequent switches.

At near-stable investment vs the previous month, the first full month after the switch (May 2026) delivered +26% margin on Google Ads non-brand Search and Performance Max. Beyond the figure, the striking point is the absence of any volume-for-value trade-off.

That is a rare configuration in mature paid acquisition, and it is exactly what margin bidding is meant to produce: aligning the algorithm's optimization objective with the real business value of each conversion, without sacrificing volume.

In a market where competitors keep investing more in paid, this quality of unit-level arbitration becomes a key competitive advantage, with no volume trade-off.

Real-margin bidding is not the right entry point for every account. To get full value from it, three conditions must be met.