Know what changes the account on its own.
We left the paused campaigns paused and intact. Deleting them would have thrown away the only record of what had and hadn’t worked.
The change history showed the overnight bid change on our brand campaign came from Google’s recommendations engine. No person had clicked anything. So in the first week we switched auto-apply off. Some of it could only be turned off by hand in the interface. One setting stayed on deliberately: a measurement setting that doesn’t touch bids, keywords or ads.
Rolling the bid change back wasn’t the answer either. We worked out what it took to show at the top for brand searches and set the bidding ourselves. Then we put an alert on the account that flagged any auto-apply setting being switched back on. Later we found Google-generated ad extensions still serving, in the wrong language and pointing to the wrong pages. We removed them, and removed them again when new ones turned up.
If a setting can change your bidding overnight, you can't tell your own decisions from Google's afterwards.
Before building the new non-brand campaign we read the old search history. It showed where money had leaked before, like drawing and CAD programs, so the new campaign went live with those excluded from day one.
We built it around four intent clusters, from searches for one specific task up to the broad construction-software category. Each cluster got its own ads and, at launch, its own bid limit. When searches meant for one cluster landed in another, a routing exclusion sent them to the right group and its matching ad.
Brand searches were excluded from every non-brand campaign through one shared list, and we added spelling variants as they slipped through. Inside the brand campaign, people searching to try the product got their own ad group, with an ad and a destination built for that step.
Each group answers one question the searcher types. Match type is a setting inside the group.
We mapped every loose keyword to the searches it actually triggered. Where one keyword fed a whole family of off-topic searches, we paused it instead of piling on exclusions, and kept its exact version live. On loose matching, under automated bidding with no cap on the click price, one category keyword pulled in the whole construction-software market, from structural calculation to CAD and planning tools. Same fix there: the exact version stayed, the loose one was paused.
We judged leaks on the last few days, because older data was full of searches we had already blocked. We also traced the search behind the non-brand conversions. It sat outside the product’s category, so we excluded it.
On one campaign, every keyword had to carry a construction word. We extended that to every keyword outside brand and competitor searches: bare software terms were paused and replaced with construction versions, and trade words that only builders use counted as their own qualifier.
Exclusions treat the symptom. When one loose keyword keeps pulling in the wrong market, the match type is the lever.
Decisions with a name behind them, and a structure you could follow.
A new bid strategy needed a person's approval, and an alert watched for auto-apply being switched back on. The old history stayed in the account. At launch, every non-brand keyword sat in a cluster built around one intent, with a first headline that matched it.
Before you build, find out what is changing your account without asking.
Every engagement starts with a 30-minute audit — free, no slides, just a screen-share through your account.
Book the audit →