The obvious theory was that variety drives retention: the markets with the deepest catalogs had the lowest churn, so more meals should mean more retention. Slicing by segment showed the real driver was liquidity, the right supply for each eater, and that distinction changed what was worth building.
Where the variety theory broke
At CookUnity, vegetarian supply looked sufficient in aggregate. Slicing supply by vegetarian users showed an under-supply problem the average had hidden, and when that gap was filled, retention for the segment improved well beyond the overall average. The lesson generalizes: an aggregate supply metric can be healthy while a segment quietly starves, and the starving segment is where the churn is.
The discovery ceiling
Even with sufficient SKUs, users weren't seeing them. In the deepest markets, eaters browsed only ~40-50% of the menu each week (smaller markets: ~60-78%, simply because there was less to scroll past), and discovery of newly launched meals capped out around ~30-50% depending on market. Past that point, adding supply was pushing inventory into a part of the menu nobody reached. The constraint had moved from supply to sorting and UX.
Habits beat novelty
Retention wasn't endless newness either. Customers who repeated meals they loved retained better, with exploration in balance rather than at the center. Spotify is the useful reference: the retention machine was never infinite catalog, it was playlists mixing favorites with new tracks.
The decomposition
Put together, marketplace retention decomposes into three levers, each measurable on its own:
The point of the tree is that when retention moves, you can say which lever moved it, which is the difference between a retention strategy and a retention anecdote.
Happy to walk through how we'd decompose it for your marketplace.
Common questions
Does this apply outside marketplaces?
The liquidity language is marketplace-specific, but the mechanism is universal: an aggregate coverage metric hiding a starving segment. In apparel it's the overall in-stock rate reading 92% and healthy while your best seller is missing its two best-selling sizes, which means you're churning your best customers specifically. In content subscriptions it's total library size growing while one taste segment has nothing new. The rule: whenever a coverage metric is an average across users, re-cut it per segment before trusting it, because the average is exactly where this failure hides.
What should I measure first?
The three branches of the tree, in order. Supply coverage per segment, where the segment definition matters: use behavior (users who buy vegetarian items) rather than surveys or profiles. Browse coverage per user, meaning the share of the catalog a user actually sees in a period, which you can build from view events you already collect. And repeat rate of loved items, from reorders or ratings. Each is buildable in a week from existing data, and any one of them can explain a retention move that the topline average can't.
Is adding more variety ever a bad idea?
Past the discovery ceiling, mostly yes. In our case eaters in the deepest markets browsed ~40-50% of the menu each week, so supply added beyond that landed in the half of the catalog nobody reached, and new items capped out at ~30-50% of users ever seeing them. Check browse coverage before funding more supply: if it's low, sorting and discovery work is cheaper than new inventory and moves the same retention number.
Retention questions like these are what our Retention & Lifecycle work is built around, cohort logic included.
See how it works →