Acceptance rate: what it measures, and when it measures the platform

Acceptance rate is the share of offers a seller accepted out of the offers they were made, over some window. The numerator is unambiguous. The denominator is a choice the platform makes every time it decides who to offer a job to, which is why an uncorrected acceptance rate is partly a measurement of your own dispatch logic.

That is not a reason to abandon the metric. It is a reason to define the eligible offer before defining the rate, and to be precise about which decisions the number is then allowed to influence.

The definition, and the half of it the platform owns

An offer has to be delivered, visible, actionable and answerable before a decline means anything, and each of those is a system property rather than a seller property. Distribution shape matters too: under a broadcast to 20 sellers, 19 decline by design, so mixing broadcast and sequential offers into 1 rate averages 2 different metrics — the distinction in the four shapes a lead can be handed out in.

Four ways the denominator lies about a seller

The arithmetic is not marginal. Take a seller offered 40 jobs in a week who accepted 11: a raw rate of 27.5%. Remove 5 offers that were never delivered, 12 that required a 45-minute drive when their jobs are normally within 20, and 9 that arrived during hours they have never worked, and 14 eligible offers remain. The same 11 acceptances now read 79%. Nothing about the seller changed between those 2 numbers.

Every offer the platform should never have sent is a free reduction in someone's acceptance rate, and the seller has no way to appeal a denominator they cannot see.

Delivery is the class most often assumed away. If offers reach sellers by phone through a masked number pool, an exhausted or misrouted pool drops them silently and the deficit lands on the seller's record — which makes sizing and recycling a pool of masked numbers a metrics problem as well as a telephony one. A daily reconciliation of offers sent against offers acknowledged, alerting on any seller whose delivery share drops, is guarded automation of the kind we build under AI agents and automation.

The loop that turns the metric into a mirror

Feed acceptance rate back into dispatch and it starts changing its own denominator. A seller whose rate dips receives fewer offers, so the next rate is computed on a smaller sample and swings harder; a favoured seller receives more and stabilises. Neither movement reflects behaviour. Left alone, this is one of the mechanisms concentrating work in a few sellers, examined in when the same five sellers take every job and one argument for capping how much one seller can win.

The behavioural loop is worse. Once acceptance carries consequences, the cheapest way to protect it is to accept everything and cancel later, which moves the failure from a moment when the job could still be reoffered to a moment when the buyer is already waiting. Pair any penalty with a cancellation measure, or the metric improves while the service degrades.

Window, floor and shrinkage: making the number comparable

1. Fix a window that adapts to volume. The last 25 eligible offers or 30 days, whichever yields more, so a quiet seller is not judged on 3 data points.
2. Set a floor before display. Below roughly 20 eligible offers, show the sample size rather than a percentage: 100% from 2 offers outranks 88% from 60 in every naive sort.
3. Shrink small samples toward the market mean. With a prior weight of 10 offers at a 60% market average, a seller who accepted 3 of 4 reads 64% rather than 75%.
4. Segment before comparing. A rate built from long-distance low-value work is not comparable to one built from short local jobs.
5. Publish the definition to sellers: exclusion rules, window and floor, beside the number, so a disputed rate is a lookup rather than an argument.

What it may legitimately drive, and what it must never drive alone

The safe uses share a property: they change the odds of receiving work, and they are reversible within days. The unsafe ones are irreversible, or they hand a single noisy signal authority over the whole ranking.

• May drive a tie-break. Between otherwise close candidates, preferring the seller more likely to accept cuts buyer wait with no claim about quality.
• May drive batch size. If predicted acceptance is low, offer to 3 sellers instead of 1 — and a short accumulation window makes those offers better targeted anyway, as argued in why waiting 2 seconds beats nearest-first.
• May drive operations. A sudden collapse in one seller's acceptance is a useful alarm, and it is usually a broken device or a changed schedule rather than disengagement.
• Must not drive deactivation on its own. A metric whose denominator the platform controls cannot be the sole basis for removing someone's income.
• Must not become a hidden multiplier in the score. Policy levers belong in an inspectable layer, for the same reason sponsored results belong outside the organic ranker.

Acceptance rate earns its place as 1 signal among several, with a denominator you can defend and a published definition. Which layer it belongs to — measurement, policy or sort order — is worth stating explicitly, as separated in signal, score and sort order. It is part of the matching, ranking and dispatch work we do for marketplaces and two-sided platforms, and the same discipline applies to fleet dispatch, discussed in AI in logistics operations.

• dispatch
• matching
• marketplaces
• metrics