One account I inherited had 23 search campaigns: split by province, split by device, even split by "brand + generic" combinations. The rationale was familiar: "This way we control everything separately." Control in appearance; in reality, two or three daily conversions divided into 23 slices, each condemned to data starvation. In the smart-bidding era, account structure has a single economy: conversion data is the fuel of learning, and every split divides the fuel.
Where the old habit comes from
The over-splitting reflex is a legacy of the manual-bidding era, when splitting genuinely meant control: separate structures were needed to bid separately by device, hour or geography. Smart bidding inverted that equation — the system already evaluates device/hour/geo/audience signals per auction; what you split by province, it can use as a signal inside one campaign. The "control" you gain by splitting is, in most cases, a manual and worse version of what the system already does. What you lose is concrete: 23 separate learning processes, none reaching statistical significance.
Legitimate reasons to split — a short list
The only solid question for a split decision: "Do these two structures differ in goal or economics?" If yes, split: brand vs generic search (one is defence, the other acquisition — with entirely different target ROAS), different countries (different currency, margin, seasonality — the story I told in the PMax article), different business lines/budget owners, and anything regulation requires managed apart. One operational exception: a strategic product group whose budget must be protected can be isolated so the system can't neglect it. Nearly every split outside this list — device, province, match type, "so it shows separately in the report" — works against smart bidding. Splitting for reporting is the most needless of all: segmentation already exists at the reporting layer; the structure doesn't need to be the report's slave.
The test I run on inherited accounts
On an over-split account my sequence is this. First, mapping: each campaign's 30–90-day conversion volume is listed; campaigns running below what a bidding strategy needs to learn healthily get flagged (in practice, ~30 conversions per month is my alarm band). Then grouping: flagged campaigns with identical goals and economics are consolidated by theme — province campaigns fold into one national campaign with geography as a signal; device splits fold into one structure. Consolidation is gradual: never everything on the same day, because a large structural change resets learning; waves of two-three weeks, each allowed to settle. The outcome is almost always the same pattern: campaign count drops to a third, total conversions rise on the same budget, and the account becomes manageable for the first time.
Consolidation has a limit too
The other side of the coin: "throw everything into one campaign" is also an ideology, and applied without context it also does damage. Merging a 15%-margin product with a 70%-margin product under one target ROAS can end with the system scaling the low-margin one and melting profit — conversion value rules, or a separate campaign, become necessary where the economic gap is wide. So does merging themes so unrelated they pollute the signal. The equilibrium is a principle: as consolidated as learning requires, as separated as economics demands. What that sentence means inside an account differs per business — and that is exactly where the craft lives.
In short: account structure isn't an org chart; it's a data-flow design. The question isn't "how many campaigns should there be?" but "how many healthy learning processes can my conversion volume feed, and how many economically distinct goals do I have?" The first answer sets the ceiling of your structure; the second sets its floor.