Why Does It Cost So Much to Acquire App Users on Paid Social?

App acquisition costs have risen for four straight years, and most of the explanation is structural rather than competitive. Knowing which part is structural matters, because it determines which levers are worth pulling and which are theatre.

What does an app install actually cost now?

iOS cost per install reached $5.84 in Q1 2026, up 19% year over year. Android sat at $1.92, up 8%. That is a 2 to 3.5x gap depending on vertical.

Vertical	Rough CPI
Sports	~$26.81 (highest measured)
Finance	High end
Gaming (iOS / Android)	~$4.22 / ~$2.97
Casual gaming, music	Among the lowest

The spread across categories is enormous, running from under $1 to over $26, which makes the global average close to useless as a target. A finance app comparing itself to a blended figure will conclude it is failing when it is performing normally, and a casual game doing the same will conclude it is winning when it is not.

The only benchmark worth tracking is your own category on your own platform, measured against your own trend. Everything else is orientation.

Two things make the category averages more misleading than they look. Vertical mix is one: a portfolio containing both a finance app and a casual game has a blended CPI that describes neither. Geography is the other, and it moves the number as much as category does. A US install and an install in a lower-monetising market are different products at different prices, so any figure quoted without a region attached is close to meaningless.

Track your own CPI as a trend line rather than a point. The number that tells you something is the direction over eight weeks against a stable creative and audience setup. A single week’s CPI moves on auction noise, competitor launches and your own creative rotation, and reading it as a verdict produces a lot of expensive overcorrection.

Why is iOS so much more expensive than Android?

Two forces compound. Apple’s App Tracking Transparency framework raised iOS acquisition costs 20 to 30% since launch, and iOS users monetise better, so more advertisers compete for them.

The ATT mechanism is worth understanding properly, because it explains why account-level optimisation cannot fix it. Ad platforms optimise by learning which users convert and finding more people like them. ATT cut off much of that signal on iOS, so the platforms now work from a degraded and delayed picture. They find good users less efficiently, and the cost of each one rises. Nobody did anything wrong; the inputs got worse.

You cannot fix this from inside an ad account. You can only route around it, which is what the rest of this comes down to.

The competitive half of the gap deserves as much attention as the technical half. iOS users spend more per head across almost every consumer category, which means every advertiser wants them, which bids the price up independently of any measurement problem. Even with perfect attribution, iOS would cost more.

That matters for how you respond. If the gap were purely an ATT artefact, the answer would be better measurement. Because roughly half of it is genuine demand for a more valuable user, the answer is partly to accept the price and check whether the lifetime value justifies it. Plenty of apps discover that a $5.84 iOS install with materially better retention and conversion is a better purchase than a $1.92 Android install that never subscribes. The CPI comparison alone cannot tell you that, and optimising toward the cheaper platform on CPI is one of the more common ways to make an account look efficient while making the business worse.

Why is the install a bad conversion event?

Because it asks too much and happens too rarely, and both of those hurt you.

An install requires the user to leave the feed, load a store listing, wait for a download, open the app, and sign up. Every step sheds people, and none of them happen on a surface you control or can measure.

The sparsity problem is worse than the friction problem. Expensive events happen infrequently, and infrequent events are exactly what optimisation algorithms handle least well. Meta’s delivery system needs roughly 50 conversion events per ad set per week before its modelling becomes reliable. At a $5.84 CPI, clearing that bar on a single ad set costs about $290 a week in installs alone, before you have tested anything or learned anything.

Multiply that by the number of ad sets a real test structure requires and the arithmetic stops working for most brands below serious scale. You are paying for the privilege of generating enough data to optimise, rather than paying for customers.

This is why the practical answer is to stop optimising for installs. Route traffic to a web signup, optimise against that, and treat the install as a downstream step you measure but do not bid on. The signup is cheaper, far more frequent, and happens on a surface you control completely. It is a measurement decision rather than a budget one, and it is usually the single largest improvement available.

Why does CPI drift upward over a campaign’s life?

Creative fatigue, and it is a running cost rather than a launch cost.

CPI does not stay flat. It rises as your audience sees the same assets repeatedly, and the only real countermeasure is new creative. Budget $1,000 to $5,000 a month for production on top of media.

Teams that treat creative as a one-time launch expense watch CPI climb month over month and misattribute it to competition, seasonality, or the platform. The account did not get worse. The assets did, and nothing replaced them.

The related trap is running your best-performing asset until it dies. By the time fatigue is visible in the CPI, it has been eroding for weeks, and you have no replacement ready because the winner was working. A standing production cadence solves this; a reactive one does not.

Frequency is the mechanism worth watching, because it moves before CPI does. As the same audience sees an asset repeatedly, response rates fall, the platform has to buy more impressions to produce a conversion, and your cost rises. Frequency climbing while CPI is still flat is the earliest reliable warning you have, and it typically arrives two to three weeks ahead of the cost increase.

There is also a subtler version that catches teams out. Fatigue is not uniform across your audience. Your best-responding segment saturates first, so the composition of who is still converting shifts underneath a stable-looking CPI. By the time the blended number moves, the segment that was carrying the account has already stopped responding, and the replacement creative has to win back an audience that has now seen you a lot.

Which levers actually move the number?

In rough order of leverage:

Change the optimisation event. Moving from install to web signup is usually the biggest single improvement available, and it costs nothing but a decision.

Fix the pixel before you need it. Tracking has to be live before traffic arrives. Visitors who land before the pixel exists are permanently untaggable. There is no retroactive fix, and the earliest traffic is often the most responsive, so this is a deadline rather than a task.

Supply more creative. Volume is the most reliable lever on cost, and one new asset per $3,000 of monthly spend is a defensible floor.

Match creative to platform. TikTok CPI varies 10 to 50% against Meta almost entirely on how native the creative is. Cross-posting a produced Meta asset to TikTok is the most common self-inflicted cost increase in app marketing.

There is a sequencing point buried in that order. The first two levers, the optimisation event and the pixel, are one-time structural decisions with permanent consequences. The second two, creative volume and platform fit, are ongoing operational commitments. Teams routinely attempt the ongoing ones while the structural ones are still broken, which is why so much creative effort produces so little movement: the account cannot act on better assets if it is optimising toward an event it cannot generate enough of.

Fix in order. An account optimising for web signups with a working pixel will get more out of ten assets a month than an account optimising for installs will get out of forty.

What does not move it?

Bidding tactics, audience micro-segmentation and dayparting get far more attention than they deserve at small and mid budgets.

All three are optimisations layered on top of a signal that ATT already degraded. Tuning bid strategy when the platform’s underlying model is working from partial data is precision applied to the wrong variable. It is not that these never matter; it is that they matter at scales where the structural levers are already handled.

The structural levers, meaning what event you optimise for, whether your tracking exists at all, and how much creative you can supply, are worth more than every bidding adjustment combined. They are also less interesting to discuss in a weekly meeting, which is part of why they get skipped.

The reason these persist is that they are legible. A bidding change is a decision you can make in an afternoon and report on by Friday, and it feels like control. Rebuilding your conversion event or committing to a monthly creative budget is slower, less visible, and harder to attribute in a weekly meeting even when it is the thing that worked.

A useful test before spending time on any account-level tactic: if this works perfectly, how much does the number move? For most bidding and segmentation adjustments at mid budgets the honest answer is a few percent, against 20 to 30% for the structural changes. That ratio should determine where the week goes.

The short version

iOS installs cost $5.84 because the targeting signal got worse and the conversion is hard, not because your account is badly run. Optimise for something cheaper and more frequent than an install, get tracking live before the first dollar of traffic, and keep creative supply continuous. The rest is noise.