Rewards were already there.
Factor sells weekly meals you can skip or pause. Customers are already paying every week.
The brief asked for a second paid membership on top of that one, built around premium meals and free shipping, the way CookUnity had bundled the same two things.
There was already a Rewards program too. Adding a membership beside it meant overlapping benefits, another subscription to explain, another set of economics to protect, and another thing customers would have to hold in their heads.
The target was 10% adoption. Before designing anything I went looking for the gap the program was meant to fill. I read customer behaviour, competitor programs, past product work, the financial model, Rewards as it stood, and what engineering could ship in the time we had.
None of it pointed where the brief pointed.
I argued to fold it into Rewards. We went the other way.
My recommendation was to put the premium benefits inside Rewards. One program. One set of economics.
The team went standalone. Standalone was the version the business could price and forecast, and someone could be held accountable to a number. Folding the benefits into Rewards would have made the result unreadable. Any lift could have come from the benefits or from Rewards itself, and we would not have been able to say which.
That was a hard note to take. It was also the right call. The point was never to win the argument. It was to know what we were choosing before we built it, and to design a test that could tell us if the choice was wrong.
A small team, and permission to throw things away.
Factor+ was built in the company’s Gen AI squad, outside the roadmap. A product manager, an engineering manager, six engineers, and me. No business as usual work, and a brief to ship daily.
That changed how I worked. We could build an idea, challenge it, drop it and start again, instead of getting every decision ready before moving. We used Claude through exploration, prototyping and implementation, which shortened the distance between an idea and something a customer could hold.
Speed only helped because we were deliberate about what we wanted to learn. Fast multiplies a bad question as readily as a good one. So we cut hard. Tiering could wait. Gifting could wait. Anything that assumed the model already worked could wait too.
The first version had one question to answer. Whether paying for this changed what anyone did.
A second subscription has to earn its place.
The membership had to make its value obvious without making anyone work to understand it. The hub put what was included, what had been used, what had been saved and when benefits expired in one place.
Benefits appeared in the cart, where they get spent, rather than filed away in an account area.
Cancellation stayed legible on purpose. A membership that is hard to leave gets noticed well before its benefits do.
The MVP was not trying to solve every possible version of membership. It was trying to make the proposition clear enough that the behaviour we measured afterwards meant something.
Adoption was the wrong number to watch.
A paid membership can look successful because people like getting a discount. That was not the thing we needed to know.
We built the measurement around a held-out control, read over fifty-two weeks and cut by loyalty level. The experiment was scoped to mid-loyalty customers.
The loyalty cut is the part that turned out to matter. A single blended number would have shown a healthy program and hidden that it was healthy in one place and expensive in another.
Adoption proves an offer is attractive. It doesn’t prove it’s worth making.
The members stayed, and that compounded.
By the July read, 7.77% of eligible customers had joined. At full rollout it cleared its 10% target, averaging 10.4% by the last read I had before moving off the program.
Weekly adoption sat around 10% and barely moved. Members stayed, so the stock compounded. Factor+ went from 1% of all weekly active Factor customers to 9% in about four months. At that same read, members were placing 13% of every order on the brand.
For customers in the middle loyalty bands the model did what we had hoped. Fifty-two-week net value rose 9.8%. Cancellation fell 5.3% against control. Those customers were behaving differently, and the held-out control is what let us say so.
At the highest loyalty levels the economics inverted. Those customers already ordered often enough that membership created almost no incremental behaviour, so we were subsidizing activity that was going to happen anyway.
The segmentation model is what the group kept.
The read changed the question. It was no longer how to scale Factor+. It was where Factor+ makes sense at all. The answer was not everyone.
The strongest return came from customers engaged enough to see the value with room left to change what they did. That is a segmentation model rather than a single adoption target.
Where is there enough behavioural headroom for a membership to pay for itself?
It has become the standard way membership is evaluated across the HelloFresh group, and it is informing the next brand launch. The next brand does not need to copy the Factor+ interface. It can start from the same question.
What we didn’t fix.
Finding the ceiling is not the same as doing something about it. The customers above it are still in the program on the terms we launched with. Deciding what to offer them instead is a pricing question more than a design one.
We also never went back to the question I opened with. Once standalone shipped and the numbers came in, the case for folding premium benefits into Rewards was stronger than it had been at the start, because we knew which customers it would work for. I did not reopen it, and neither did anyone else.
And there are still two programs to explain. Rewards and Factor+ sit beside each other, which was the objection I raised in the first week. It remains true.