Learning phase economics: why small budgets lose twice
Small campaigns pay an exploration tax and then never earn it back. A budget floor rule, and the trade-off hidden in the standard fixes.
A small campaign does not get a proportionally smaller version of the result a large campaign gets. It gets a worse one, and it pays more per unit for it. That is the claim, and it runs against the intuition most studios operate on, which is that a test budget is a scale model of a real budget. It is not. Automated bidding systems learn by spending, and the cost of learning is roughly fixed in conversions, not in currency. A campaign that cannot afford the conversions cannot afford the learning, and it pays for the attempt anyway.
The 2026 edition of the AppsFlyer State of Gaming for Marketers report, published last week with Unity and Newzoo, put the top tier of advertisers at between 2,400 and 2,600 creative variations a quarter, a rise of roughly a quarter to a third on the previous year. Every one of those variations is a small learning problem inside a large one. The scale players are not just outspending smaller studios; they are buying learning in bulk at a lower unit cost.
Where the first loss happens
Platform documentation is consistent about the mechanism even where the exact numbers differ. Meta's published guidance, for instance, describes an ad set as leaving the learning phase after roughly fifty optimisation events in a week. Other platforms describe similar thresholds without publishing them. During learning the system is exploring: showing the ad to segments and placements it is uncertain about, in order to discover where conversions are cheap. Exploration is expensive by design, because half its purpose is to find out where not to spend.
For a large campaign this tax is a small fraction of spend. For a campaign that clears fifty events in a week only if everything goes right, the tax is most of the budget. Worse, if the campaign never reaches the threshold, the system never moves into exploitation and the entire budget is spent in the expensive mode. The campaign reports a CPI, the team compares it to a benchmark, and the comparison is meaningless because the two numbers came from different phases of the algorithm.
Where the second loss happens
The second loss is subtler and rarely discussed. The impressions a learning campaign wins are not a random sample of the market. Exploration allocates budget to inventory where the model's uncertainty is highest, which correlates with inventory that better-informed bidders have already priced or rejected. A small campaign is, in effect, buying the impressions the confident bidders left behind, at a price set by the exploration logic rather than by the campaign's own economics.
The consequence is that the cohort a small test acquires is systematically different from the cohort the same game would acquire at scale. Retention and payer rates from a learning-phase test are not a preview of what a scaled campaign would deliver; they are a description of the leftovers. Studios that use a small test to decide whether a game can be scaled are reading the wrong data, and both false positives and false negatives are common.
The trade-off inside the standard fixes
The usual advice is to consolidate: fewer campaigns, fewer ad sets, broader targeting, so that conversions pool and the threshold is reached. That advice is correct as far as it goes, and it carries a cost that is rarely priced.
Consolidation trades learning speed for information. A single broad campaign learns faster but tells you almost nothing about which creative concept, which geography or which placement is doing the work. The platform knows; you do not. For a studio that intends to spend heavily later, that missing information is worth something. For a studio deciding whether to spend at all, it may be worth more than the efficiency gain.
The other common fix is to optimise for a shallower event. Installs are plentiful where purchases are scarce, so a campaign that optimises for installs exits learning in days rather than weeks. The trade-off here is what the algorithm is selecting for. An install-optimised campaign learns to find people who install, which includes a large population of people who install anything. The learning completes, the CPI looks good, and the cohort is weak. The exit from learning was purchased by changing the question.
A budget floor rule
A workable decision rule follows from the threshold logic. The weekly budget a campaign needs to learn is at least the number of events the platform requires multiplied by the expected cost per event, with a margin for the exploration premium.
An illustrative case: a platform threshold of fifty events a week, an expected cost per purchase of 40 in the target market, and an exploration premium of thirty percent during learning. The floor is fifty times 40 times 1.3, or 2,600 a week for that one campaign. A studio with 5,000 a week to spend across three markets and two platforms cannot run six purchase-optimised campaigns. It can run one, or it can run several install-optimised campaigns and accept the cohort risk, or it can pool markets and lose the geographic signal. Those are the actual choices, and the rule makes them visible before the money is spent.
- Below the floor: do not optimise for the deep event. Either shallow the event and check cohort quality by hand, or hold the budget until it can be concentrated.
- At the floor: run one campaign, keep it stable for at least two weeks after exiting learning, and do not touch bids or creative during that window.
- Above the floor: the surplus is what buys you segmentation. Spend it on splitting by the dimension whose answer changes your next decision, and no others.
Reading a small test honestly
If a small test is unavoidable, the practical adjustment is to stop reading it as a scale model. Compare learning-phase CPI only to other learning-phase CPI. Discount early retention for the leftover effect rather than projecting it forward. And record the number of optimisation events the campaign actually accumulated alongside every metric it produced, so that anyone reading the results later knows which mode the algorithm was in.
The hard truth for small studios is that a test budget buys a test of the algorithm's patience, not of the game. The studios that scale are often the ones that concentrated a modest budget into one honest learning run rather than spreading it across a dozen campaigns that each paid the tax and never collected.
