A paid campaign was reviewed and found to be reaching somewhere between 148 and 175 million people, with essentially no interest or behavioral narrowing applied, at roughly a dollar per click. Nothing about that was intended — the account's own history showed a much smaller, deliberately narrowed audience had been set up at launch. Somewhere between then and the review, the platform's own broadening defaults had reasserted themselves, and nobody had a way to notice until the spend numbers stopped making sense.
Why a setting screen wasn't the fix
The obvious fix is to open the campaign and re-narrow the audience by hand. That was done — but it doesn't actually solve the problem, because the same setting that drifted once can drift again the same way, silently, the next time the platform nudges toward broader reach or the next time someone touches the campaign without knowing the narrowing was load-bearing. A manual fix corrects the state. It does nothing about the process that let the state go unnoticed.
What replaced it
Every ad set now gets built from a single machine-readable spec file rather than hand-set in a UI, with the platform's own "expand my audience automatically" option explicitly forced off rather than left at its default. A test now asserts that setting on every campaign definition before anything ships, so a campaign that would silently re-broaden fails a check instead of quietly spending against the wrong audience for however long it takes someone to notice the numbers look off.
The failure wasn't a bad targeting decision. It was that the correct decision had no way to stay true over time — it depended on nobody touching a setting that looked optional and wasn't.
The tradeoff
Building and maintaining a spec file plus an automated guard is slower than adjusting a targeting panel directly, and it adds a small amount of process to every future campaign, even the ones that would never have drifted. That cost is paid up front, on every campaign, whether or not this particular failure would have recurred. It's worth it because the failure mode is specifically the kind that doesn't announce itself — the campaign keeps running, the dashboard keeps showing clicks, and the only signal is a cost-efficiency number that erodes slowly enough to look like normal variance for a while.
What would change the answer
If a platform's targeting settings were sticky by default — if narrowing, once set, simply stayed set — none of this would be necessary, and the guard would be pure overhead. The decision only makes sense because the platform's actual behavior isn't sticky, and the cost of finding that out the hard way a second time is higher than the cost of the guard.