What would justify raising the share of a podcast feed's impressions reserved for items with no engagement history?

## What the budget buys and what it costs An exploration reserve takes impressions that the ranking stages would have filled with their best estimate and spends them on items whose value is unknown. The **cost** is visible within a day: the engagement gap between reserved slots and ranked slots, multiplied by the number of reserved slots. Anyone can compute it, and it will always be negative in the short run — that is what exploration is. The **return** is slower and lands in three different places, which is exactly why this is a judgment call rather than an optimisation: - **Rankable inventory.** An item that completes its quota has an estimate, so it can be ranked on merit forever after. The reserve converts dark inventory into inventory the funnel can use. - **Catalogue breadth.** A discovery surface whose value proposition is range cannot serve range from a few thousand proven items. - **Supply.** Podcast catalogues are two-sided. Creators who publish and receive no impressions stop publishing, and the inventory lost that way never appears in any engagement metric on the demand side. ## Readings that argue for raising it 1. **The graduation rate has fallen below the publishing rate.** With a 1% reserve of 20,000,000 daily impressions and a 200-impression quota, 1,000 items complete their quota per day. Against ~900 new episodes a day that is barely break-even; at 1,400 a day the cold queue grows without bound and no amount of tuning elsewhere absorbs it. 2. **p95 time-to-first-impression is drifting out** while p50 holds — the signature of a queue with a fixed drain and a growing arrival rate. The median item is fine and the tail is starving. 3. **Coverage is falling as the catalogue grows.** A reserve expressed as a fixed share of impressions does not scale with inventory; if the catalogue doubles and the share does not move, the fraction of items ever shown halves. 4. **Supply-side churn is measurable** among creators whose items never cleared the quota. This is the argument that most often moves a business, because it is about inventory disappearing, not about a slot's engagement. 5. **Graduated items survive on merit** at a healthy rate. If items that complete the quota then earn impressions on their own, the reserve is buying good inventory and more of it is worth buying. ## Readings that argue for lowering it - Graduated items almost never survive on merit — the reserve is buying data about items the audience does not want, and a cheaper pre-filter on eligibility belongs in front of it. - The engagement gap on reserved slots is much wider than the modelled value of the estimates being bought, and stays wide as the reserve grows. - The cold queue is short and the quota is being met with headroom to spare: the budget is not the binding constraint, so spending more buys nothing. ## The argument that does not carry "Reserved slots engage worse than ranked slots." They do, by construction — they are showing items the system cannot yet rank. Presenting the known price of a policy as evidence against it is the most common way this conversation goes wrong, and the counter is to quote the gap as a budget line that was already agreed, not as a discovery. ## Turning the argument into a number The reserve is easiest to defend when it is derived rather than chosen: - start from the inventory you intend to bootstrap per day — say the publishing rate plus a stated allowance for back catalogue; - multiply by the per-item quota to get the impressions required; - divide by daily impressions to get the share. At 900 new episodes a day plus 300 back-catalogue items, and a 200-impression quota, that is 1,200 × 200 = 240,000 impressions, or 1.2% of 20,000,000. Now the number has a reason attached, and every future argument is about one of the three inputs rather than about the percentage itself. The quota is a real lever here: halving it halves the required share at the cost of graduating items on weaker estimates. ## Who owns the number Because the cost lands today and the return lands later, nobody optimising a weekly engagement metric will ever propose raising it. That makes the reserve a stated business commitment with: - a named owner who accepts the short-term cost; - a stated horizon over which the return is expected; - a review cadence, since the publishing rate and catalogue size that justified the number both move; - configuration, not a constant compiled into the scoring stage, so the review can actually change it. ## The honest uncertainty The long-term return on exploration is genuinely hard to attribute: the benefit shows up as items ranking well months later, and you cannot easily separate it from everything else that changed. The defensible position is to say so, to keep the cost side rigorously measured, and to argue the size from the arithmetic of inventory and quota — which is checkable — rather than from a claimed long-term gain that is not.