How do you decide how much relevance to trade for catalog coverage on a feed?

## Start by refusing the blended score The instinct is to write one objective: `score = relevance + lambda * diversity`, tune lambda, ship. It is a bad first move for a decision this political. A single blended number destroys the only information the discussion needs — how much of one thing you gave up for how much of the other — and the choice of lambda quietly encodes a business judgment nobody agreed to. The better frame is constrained optimisation with an explicit guardrail. State it in words: *maximise catalog coverage subject to the primary engagement metric not falling more than X% relative*. Now every candidate change reports two numbers, the tolerance X is a stated decision with an owner, and the argument moves to where it belongs — what is X, and why. ## Establish why coverage is worth anything at all You cannot price a trade whose benefit you have not named. Coverage is never valuable for its own sake; it is a proxy, and which proxy it is determines how much relevance it is worth. - **Supply-side retention.** In a two-sided marketplace, items with no exposure mean creators or sellers with no return. They leave, and the catalog you spent money acquiring shrinks. Here coverage is a direct input to future relevance, because tomorrow's good recommendations need tomorrow's supply. - **Insurance against narrowing.** A system that only ever exposes what it is already confident about generates training data that confirms its existing beliefs. Coverage is the cheapest observable that this has not happened. - **Cold-start throughput.** Coverage governs how fast new items acquire enough interaction to be rankable at all. A feed with low coverage has a slow catalog metabolism. - **User-facing variety.** Repetition is a real product defect, though this one is usually better measured on the list than on the catalog. Each of those has a different currency, and only some of them convert into engagement on the horizon your experiments run. ## Set the tolerance with the right horizon The honest difficulty is that the cost of the trade is immediate and measurable while the benefit is delayed and diffuse. Showing a less-certain item costs engagement today, in this session, visible within hours. Supplier retention, catalog metabolism and resistance to narrowing pay out over months. So a two-week test that reports "engagement down 0.6%, coverage up 9 points, no other movement" is not a verdict. It is the first half of one. What raises the quality of the decision: - **Long-horizon holdbacks.** Keep a population on each arm for a quarter or more and read retention, session breadth and supply-side metrics, not just click-through. - **Supply-side outcome metrics.** Creator or seller retention, share of suppliers receiving any exposure at all, time-to-first-hundred-impressions for new items. These make the benefit side of the trade a number rather than a belief. - **Explicit horizon in the guardrail.** "No more than X% relative engagement drop, measured at four weeks" is a decision. "Don't hurt engagement" is a veto that guarantees the status quo, because every diversification move loses something on day one. Be prepared to say that in some products the right answer is a very small X. A search-shaped surface where the user is looking for one specific thing has little use for coverage in the ranked result; a discovery-shaped feed with a large catalog and a supplier ecosystem has a lot. ## Guardrail, not target The last decision is what the coverage number is allowed to be organisationally. Made a team target, it is trivially gamed: pad the tail of every list with obscure items, and coverage rises while nobody engages with the additions. The relevance guardrail catches the crudest version, but a metric under target pressure drifts toward whatever satisfies it most cheaply. Two defences. Keep coverage as a **health guardrail** — a threshold that blocks a release rather than a number a team is bonused on. And measure it in a form that resists padding: rather than distinct items ever shown, use coverage weighted by meaningful exposure, or the share of the catalog that received engagement above a floor, so items dumped into slot ten with no interaction do not count. ## What a strong answer sounds like Name the constrained framing rather than a blend. Name the proxy — say which of supply retention, cold-start throughput or long-run breadth is the actual payoff in *this* business. Name the horizon mismatch and how you would fund a measurement long enough to see the benefit. Name the Goodhart risk and pick guardrail over target. And say plainly that the number X is a business decision you would put in front of the people who own supply and retention, with the evidence attached — not a hyperparameter you would tune alone.