How learning phase disruption happens
When a campaign is created or significantly edited, the bidding system starts without reliable information about which auctions are worth entering. It spends a period gathering that information, and results during it are unstable: cost per result swings and delivery is uneven. Once enough conversions have accumulated under the current settings, performance settles down.
A meaningful edit resets that process. Changing the bid strategy, the target, the conversion action, the audience, the creative or the budget by a large step tells the system that what it learned may no longer apply. Disruption is what happens when edits arrive faster than the campaign can finish learning: the account spends most of its life in the unstable state and never reaches the settled one.
Why it matters
The damage is quiet, because each individual edit looks reasonable. Costs rise, so the target is tightened. That resets learning, results get worse, so something else is changed. Every intervention is a response to instability the previous intervention caused, and the campaign never gets the run of steady days it needs to prove anything either way.
It is worse on smaller accounts. A campaign with few conversions a week takes longer to accumulate the volume the system wants, so the window in which an edit does damage is wider. That is the ordinary situation for advertisers in Nepal and for local service businesses anywhere.
Where it goes wrong
Daily optimisation is the biggest single cause. A habit of logging in every morning and adjusting something was rewarded under manual bidding and is punished under automated bidding, where the system needs consistency more than it needs attention.
Two other patterns cause it regularly. Large budget moves in a single step are read as a change of circumstances rather than a change of scale, so raise or cut in smaller increments and let each one settle. And restructuring during a downturn — splitting or merging campaigns to fix bad numbers — restarts learning across everything at once, at exactly the moment stability is most needed.
What to do about it
Batch your changes rather than trickling them. Decide what you want to alter, make it all in one session, then leave the campaign alone through the learning window and through a stretch of settled trading afterwards. Judging performance while a campaign is still learning is judging noise and calling it a result.
Agree an optimisation cadence and keep to it, with anything urgent limited to genuine faults such as a disapproved ad or broken tracking. When a change really needs testing, run it as a campaign experiment so the original keeps running while the variant is judged against it, instead of editing the live campaign and hoping the next fortnight explains itself.
