Scissors With Your Pizza: How One Supplier Broke Our Auction

A formally correct scoring formula still returned scissors instead of champagne - because the bidding field was never diverse. How we fixed catalogue monoculture structurally.

On August 6, 2026, our founder found a bug by doing what every user eventually does: he used our own demo. The question was simple — what drink goes with pizza? The answer was a pair of scissors. Not champagne. Not wine. Not a non-alcoholic pairing. Scissors. This is not an anecdote about a quirky one-off. It is the story of a structural failure at the heart of any auction-based match infrastructure — and why fixing it means more than closing one bug. Twelve auctions, one winner The measurement was unambiguous: twelve auctions in a row, all won by the same campaign — across completely unrelated queries like "standing desk", "headset", "wine", and "what drink goes with pizza". In the pizza auction, all ten participants in the bidding field belonged to that one campaign. Bid, quality and context were identical across all ten; the only thing deciding the outcome was similarity to the search query, and that was flat, between 0.43 and 0.54. Two obvious hypotheses were ruled out in the same session: product embeddings weren't missing (4,242 out of 4,242 present), and an expired end date wasn't even read on the auction path. The real cause was elsewhere. The auction only ever nominated products the organic product search had already found. One single supplier accounted for 3,931 of 4,180 biddable products — nearly 94 percent of the entire catalogue. The moment that supplier showed up in a search result, the bidding field collapsed onto them alone. A second supplier with 197 active products — including a champagne campaign — wasn't losing the auction. It never made it in. That's the difference between a fair contest a small supplier loses, and a contest it is never admitted to. The first fix wasn't enough The obvious fix: for queries with detectable pairing intent ("what drink goes with X?"), also admit campaigns whose declared target categories cover the complement — matched on word boundaries, so "Christmas decorations" doesn't accidentally open the champagne catalogue just because the words sound similar. The field grew from ten bidders to fifteen, from one campaign to three. The champagne campaign bid for the first time ever. And still lost: # Campaign Bid Similarity Quality Context Score 1 (winner) Pizza stone 1.00 0.46 0.70 1.00 0.78 11 Champagne 1.00 0.22 0.70 1.00 0.70 The reason lies in the nature of the question. A pairing query deliberately crosses a …

Author

Nexbid Engineering

Published

2026-08-11

Topics

Agentic, Agent Discovery, Verification