Campaign consolidation works until it does not
Fewer, bigger campaigns help network models learn faster. Past a point they hand the network your allocation decisions and hide the mix shifts that hurt.
Every network representative tells you to consolidate, and for the first half of its life the advice holds up. Merge your dozen thin campaigns into two fat ones, the model gets more conversions per day, exits its learning phase, and CPI falls. Teams that did this in 2024 and 2025 mostly saw the promised improvement. What the advice never comes with is a stopping condition, and the point where it stops helping stays invisible in the metrics the network puts in front of you.
The thesis: consolidation improves outcomes only while everything inside the campaign shares the same optimisation event and roughly the same value curve. Merge segments whose users are worth different amounts to you and you have stopped structuring campaigns. You have started handing your budget allocation to a model optimising for something adjacent to your profit rather than your profit.
What the model is actually doing with a bigger pool
A consolidated campaign hands the network's model one instruction: get as many of the optimisation event as possible at or under the target. Give it a bigger and more varied pool, and it does exactly what it should, shifting spend toward wherever that event comes cheapest. Merge a tier-one geo campaign with a tier-two one, and if installs cost less in tier two the mix drifts that way. Your CPI improves. Your reported ROAS may even hold, because the target is a blended number.
What changed is the composition of the cohort, and the campaign summary does not carry that. The network reports a campaign-level CPI and a campaign-level return. Underneath, the geo mix moved, the device tier mix moved and the placement mix moved with them; unless you export the breakdown every week, you find out a month later when day-30 revenue comes in light.
That is the mechanism. Consolidation converts a set of explicit allocation decisions you used to make (how much to each geo, each OS, each placement group) into implicit ones the model makes for you, using a proxy for value rather than value itself. The proxy is the optimisation event. If the event is a purchase in the first day, the model will happily buy users who make one small purchase and churn over users who make none on day one and a large one on day nine.
The second-order costs nobody puts on the slide
Two consequences follow that get much less attention than the CPI improvement.
The first: allocation confounds creative testing inside a consolidated campaign, because the model decides which creative gets which audience. A new concept that performs well in one geo and poorly in another has its budget routed to the geo where it converts, and the campaign report shows a healthy average. So you cannot tell whether the concept works or whether the model found the one place it works. Fragmented campaigns were clumsy. They also held the audience constant while the creative varied, which is what a test needs.
The second one costs more. Consolidation raises the price of correcting a mistake: when a single campaign carries most of a network's budget and the model has settled into a pattern you dislike, every intervention, whether you change the event or tighten the target or exclude a geo, sends the whole budget back into learning. Run five campaigns instead and you fix one while the other four keep going. Consolidation trades resilience for efficiency, and that trade only looks free while nothing goes wrong.
Three tests before you merge
Apply these to any two campaigns you are considering combining.
Same event. Do both already optimise toward the same in-app event, or would one of them have to change? If one changes, you are not consolidating but re-optimising, and the CPI improvement you see afterward comes partly from the new event rather than from the merge itself.
Same value curve. Do users from both segments reach comparable value by the day you judge them? As an illustrative example, if one segment's day-7 payer rate sits around four percent and the other's nearer one percent, a blended target lets the model buy the cheaper segment until the blend looks fine and the business does not.
Same lever. If you needed to cut spend in one segment tomorrow without touching the other, could you? Bid modifiers and geo exclusions are weaker than they look inside a learning-phase campaign.
A pair that passes everything can merge, and you get the learning benefit. A pair that fails on any one of them stays apart, and the slightly higher CPI is what you pay to keep the decision.
The decision rule in practice
A useful rule: consolidate up to the conversion volume the network needs to stay out of learning, and no further. Most networks publish a rough threshold, typically expressed as events per week per campaign. Once a campaign sits comfortably above that number, extra volume from a merge buys very little model quality and costs you a lever.
Consider an illustrative strategy game running six campaigns on one network, each below the learning threshold. Merging into two, split by value tier, moves both above threshold and CPI falls meaningfully. The team, encouraged, merges the two into one. CPI falls again, by a small amount, and over the next three weeks the geo mix moves twelve points toward the cheaper tier. Day-30 ROAS on the combined campaign eventually lands below either of the two previous campaigns. The second merge was the one that did not work. Nothing in the network dashboard flagged it, and the figures here are illustrative anyway; the shape is what recurs.
What to watch after a merge
Consolidation is not something you do once; it is a state you monitor. Export the segment breakdown weekly. Track what share of spend goes to each geo cluster, then do the same for OS and for placement type. Set a drift threshold, say a ten-point move in any dimension over a fortnight, and treat crossing it as a review trigger rather than a fact of life.
When drift crosses the line, the fix is rarely to un-merge everything. Usually you split out the one segment absorbing spend it should not, give it its own target that reflects its own value, and leave the rest consolidated. The network's advice to keep everything together is correct for the network's model. Your job is to work out where its objective and yours part ways, then put a campaign boundary exactly there.
