Not Everyone Needs 30% Off

A restaurant runs 30% off its whole menu. Every customer who opens the app sees 30% off.

The regular who orders from them every Friday gets 30% off an order they were going to place anyway. The person who has never tried that cuisine sees the same 30%, which might not be enough to get them to take a chance. One number, doing two opposite jobs badly.

That was how most of the discount budget went out. Untargeted, the same depth for everyone. And much of it was funded by restaurant partners themselves, so every unnecessary discount came straight out of their return.

Show each customer the discount they actually need. For the same restaurant: nothing for the loyal regular, a little for someone who has tried it but orders around, more for someone who has never tried the cuisine but is open to it.

The obstacle was obvious once you looked for it. If every customer saw the same discount, the history can’t tell you what anyone would have done at a different one. There was nothing to learn from. We had to create the data first.

So I designed the work in three phases, each one earning the right to the next.

For full-menu discount campaigns, the control group saw the partner’s original discount. In the variant, each customer and partner pair was randomly assigned a different depth: the original, 10 points deeper, 10 points lighter, or 20 points lighter. The same customer always saw the same depth for the same partner, so the experience stayed consistent.

Guardrails mattered as much as the design. Discounts were capped at 50%. If a discount fell to zero, the customer could still see the partner’s other deals. Sales were tracked against a pre-agreed guardrail throughout, so we would know early if the test was costing too much.

Partner return was defined simply: sales generated, divided by what the discounts cost. Measured the same way for control and variant, partner by partner.

The hardest part wasn’t the maths. It was keeping partners in. Some worried that random discounts would undermine trust in how deals work, especially brands new to promotions. A test that loses its participants halfway through proves nothing, so a lot of the work was with commercial teams, explaining why the test mattered and what partners would get out of it.

The experiment gave us a read on how different customers and partners respond to discount depth. I turned that into a rule-based model: customers grouped by how much a discount changes their behaviour, partners by basket size and how much they rely on deals.

The logic is simple to say and hard to get right. Go deeper where a discount changes behaviour. Pull back where it doesn’t. And protect the margin of mature brands whose regulars would order anyway.

Rules came before machine learning on purpose. They are easy to explain to partners and account managers, easy to check against a control, and they give the model a baseline it has to beat.

The experiment did what it was built to do. It gave us the sensitivity data the model needed, segment by segment.

It also showed something worth knowing on its own: many customers who had been getting deep discounts converted at lighter ones, or with none at all.

And it surfaced the tension the next phase had to solve. Deeper discounts drive more orders but lower partner return. Lighter discounts improve return but can cap volume. Partner return and platform sales don’t always move together, so the goal isn’t to maximise either one. It’s to find the point that works for both.

The rule-based model went live in three pilot markets, chosen because menu discounts were a large share of their deals. The machine learning phase was designed to follow: partners commit a fixed discount budget, and the model decides which customers see which discount.

If everyone saw the same price, your data can’t tell you what anyone would have paid.

Sometimes the most valuable thing a project produces is the dataset that makes the next one possible. Spending a little to find out beats guessing at scale.