The account where 41% of conversions were not people

The problem

A sweepstakes advertiser came to us with a Tier-3 pop account that looked fine from the buying side. Roughly $14,000 a month across Brazil, the Philippines, and Bangladesh, a $0.81 CPA against a $0.90 goal, volume steady. The complaint came from the other direction: the network paying for the leads had started rejecting them in bulk, and the advertiser's effective payout had quietly fallen by a third over two months.

The desk-review version of the diagnosis took one afternoon. Single-opt-in sweeps leads convert on a click and a form fill, which is exactly the kind of conversion cheap bot traffic fakes well. The tracker said the campaign converted at 1.9%. The advertiser's back end said barely more than half of those leads ever opened a confirmation email.

The approach

We rebuilt the account's reporting around postback-verified events before touching a single bid. Sub-IDs were remapped so every conversion carried its zone, and the advertiser agreed to pass back a quality flag, a simple binary for "showed any human signal within 72 hours." That flag is the whole method. Without it you are optimizing toward form fills; with it you are optimizing toward people.

Then the scoring ran. Of 312 zones with spend in the trailing 90 days, 41% of conversions failed the quality flag. The failures were not evenly spread: 19 zones produced 78% of the junk, with time-to-conversion distributions that clustered under four seconds, an Android share of 99.2% in a market where StatCounter puts it near 87%, and rebill rates of exactly zero. Each got the kill rule: three times payout spent with no quality conversion, gone that day. Week three added a second pass at the network level, moving spend from run-of-network to a whitelist of 44 surviving zones and letting the bid ladder climb $0.05 CPM at a time inside it.

The result

Six weeks in, quality-flag failures sat at 9% of conversions, spend was back to $13,000 a month, and the raw CPA had risen from $0.81 to $0.84, which is the honest cost of refusing junk. The advertiser's rebill-adjusted payout, the number that actually pays invoices, came out 2.3 times higher than the pre-purge baseline. The rejected-lead conversations stopped.

One thing did not work, and the file says so: we tried a fingerprint-based bot filter in front of the prelander for two weeks, and it cost 11% of real traffic while catching junk the zone scoring had already flagged. It came out. The zone-level kill rule did the work; the gadget did not.