Your CPI benchmark is lying to you.

What a category CPI average is made of

Take every install a data provider can see in, say, health and fitness, and divide total spend by total installs. The resulting figure blends together things that differ from each other by an order of magnitude.

Variable	Typical spread within one category	Why it matters
Geography	Tier-1 markets cost many times Tier-3	A category average weighted towards cheap markets looks affordable and is unreachable in the US
Platform	iOS and Android differ substantially and inconsistently	Blending them describes neither
Channel	Search intent vs cold social vs incentivised	Intent is the largest single driver of both cost and quality
Brand vs non-brand	Branded installs cost a fraction of cold ones	A known brand drags the whole category average down
Attribution model	SKAN vs deterministic vs modelled	Different providers count different installs as installs
Seasonality	Q4 auctions reprice everything	An annual average describes no month you will actually buy in

Any one of those explains a two-to-threefold difference on its own. Stacked, they produce a range so wide that the midpoint is not a target – it is an artefact.

Why we are not publishing the table anyway

We could. It would rank, because the search volume is there and the bar for these pages is low. But a number with no methodology behind it is exactly the kind of thing we tell clients to stop optimising towards, and publishing one to catch traffic would make this page a worse version of the thing it is criticising.

What is worth reading, if you want direction rather than a target: Business of Apps maintains continuously updated user acquisition and Apple Search Ads cost pages, and AppsFlyer's Performance Index – the 2025 edition covered 16.2 billion installs across more than 39,000 apps – is the most rigorous public view of which networks deliver in which category. Both are useful for understanding market structure. Neither should set your bid.

The comparison that actually works

Hold everything constant except what you changed. Compare this month's cohort to last month's, within one geography, one platform and one channel. Product, price and audience stay fixed, so any movement in the number is attributable to something you did – which is the only property that makes a benchmark actionable.

Then, separately, work out your ceiling: the highest cost per install your funnel and your payback tolerance can support. That number is derived from your own economics and it does not care what anyone else pays.

What to do if you have no history

Work forwards, not backwards. Take your price, your store commission, your churn and your paywall conversion, decide how many months of payback you can fund, and solve for the maximum you can pay. Then start buying below it and find out what the market charges you.

This gives you something a category median never can: a figure you can defend to whoever controls the budget, derived from the business rather than from a blog post.