Give every row on the home screen an owner the listener can name
Recommendation teams are often asked to meet commercial and cultural commitments by adding weights to a ranking model. Each obligation should instead have its own labelled, capped place on the screen, so that listeners and boards can see who decided what.
A home screen in a cultural business carries more obligations than its design admits. Ours at Lisandre Audio carries at least three: agreements with publishers who pay for visibility, a public commitment we made in 2020 to give independent French-language publishers a fair share of attention, and the everyday duty to show each listener something they will want to hear. My position is that these should not be blended into one ranking by adding boosts to a personalisation model. Each obligation should have its own labelled place on the screen, with a limit on how much of the screen it can take, and the listener should be able to tell which row belongs to whom.
This is a minority view among people who build recommendation systems, and I understand why. A blended ranking is more efficient. It can satisfy several constraints at once while losing very little relevance, and it spares designers the awkwardness of labels. I accept all of that. My objection is about accountability rather than efficiency.
Three obligations hidden in one ranking
Until 2024 our home screen had fourteen rows. Five were partly filled through visibility packages that our commercial team sold to publishers, which earned about €7.4 million in 2023. They carried no label. Our charter with the Réseau Tilleul des éditeurs indépendants committed us to give its members and other independent publishers at least 15% of home-screen impressions across each year. The remaining positions were personalised.
Late in 2023 our recommendation team introduced a new model that raised listening starts from the home screen by 9%. It also cut independent publishers' share of impressions from 15% to 8% within a quarter. Nobody had intended that. The model had learnt, accurately, that titles from our largest publishers arrived with heavier promotion elsewhere and so started more often. Tobias Wendling, who leads our recommendation science, proposed a technically elegant correction: add one weight for publisher independence and another for paid packages, and tune both until the charter figure and the commercial commitments were met.
Why we kept the obligations out of the model
I argued against the weights, and the argument lasted six weeks. My objection was never that weights fail. They work well. It was that a weighted ranking makes three different promises indistinguishable. When a title appears fourth in a row, nobody, including our own executives, can say whether it is there because the listener may like it, because a publisher paid or because we owed small presses a share of attention. When the weights are retuned each quarter by a data science team, the trade-offs between those promises are being made in a place that no commercial director, editor or board member ever looks.
The design we launched in June 2024 has twelve rows. No more than two carry sponsored placements, each marked as promoted by the publisher. One row is chosen each week by our editors, who report to our content director, and that row and a monthly feature carry most of our charter commitment. Every other row is personalised and says so, with a short line under each title giving the main reason it appears. The model is no longer asked to keep any promise except relevance.
The costs appeared within a quarter. Our commercial director, Arnaud Pellequer, had warned that labelled placements would sell for less, and they did: visibility revenue fell to €5.1 million in 2025. Home listening starts fell 2% in the first quarter, partly because the editors' row performed below the personalised rows it displaced. By the end of 2024 they were 3% above the old baseline, as the personalised rows, relieved of their boosts, improved. Independent publishers received 16% of home impressions in 2025 and 13% of listening starts. Some publishers who bought packages told us the label made their books look like advertisements. A few of them had a point, and we changed the wording twice.
The separation also changed conversations inside the company. Before, a publisher unhappy with its visibility would call the commercial team, who would call the data scientists, who would adjust something nobody else could see. Now the commercial team sells a defined, labelled product of a known size, and the recommendation team can decline a request to favour a title without the refusal becoming personal. Our editors, for their part, must defend their weekly choices in a quarterly report that anyone on the executive committee can read.
Explaining the main parameters in a listener's words
Separating the rows also made a legal duty easier to meet honestly. Because creators can upload podcasts to Lisandre, our lawyers concluded that part of the service is an online platform under the EU's Digital Services Act, which among other things expects platforms to set out in plain terms the main parameters of their recommender systems. A blended ranking makes that explanation either vague or unreadable. With separate rows we could write it in four sentences that a listener would understand, in French and in Spanish.
We also offered a home screen without personalisation. As I understand the Act, the duty to offer such an option falls on the very largest platforms, which we are not, but once the rows were separate it cost us little. About 4% of subscribers use it, and they are disproportionately over sixty. Several told our support team they simply prefer to browse a shop window that looks the same as their neighbour's.
There are honest objections to all this. Separate rows are cruder than a well-tuned blend, and a listener who scrolls past an editors' row has lost a position that a model might have filled better. Labels can stigmatise good books that happen to be promoted. Human editors have tastes and blind spots of their own, which are harder to measure than a model's. We answer the last point by reporting to our CSR committee each quarter what the editors' row featured, by publisher size and language, and by rotating editors through it.
I would put one question to any board overseeing a recommendation system that carries obligations, whether to advertisers, to rights holders or to a public mission. For each row or slot on the screen, who decided it should be there, and can the customer tell? If the answer is that the model balances several things, the board is not overseeing those obligations. It is trusting a tuning parameter to keep promises the company made in its own name, and nobody will notice when the parameter changes.