October 5, 2026

Paid acquisition for AI consumer apps: buy payers, not installs

Buy paying subscribers, not installs. Every active user of an AI app costs you inference, so an install that never pays still costs money, and a campaign can look profitable in the ad platform while it loses you money. Set bid caps from revenue after store fees minus inference, point Google, Meta and TikTok at trial starts or purchase value, and judge Apple Ads keywords on the trials they bring.

The revenue doesn't have to be in-app either. On the US App Store, apps can now link to web checkout without paying Apple a commission, at least until a court approves one.

Unit economics when every active user costs inference

Model inference from the spread of usage, as in our post on pricing an AI product, not the average.

Here's an illustration with made-up numbers. Say your app sells an $80 annual plan after a 7-day free trial. You're in Apple's App Store Small Business Program, so Apple keeps 15% and you net $68 a year. Inference costs $1.50 a month per subscriber, and 30% of subscribers renew once, so over two years the average subscriber brings in $88.40 and costs $23.40. Each trial costs $0.50 in inference, and each install that never starts a trial uses $0.20 of free usage. Campaign A optimizes for installs, campaign B for trial starts. Both spend $3,000 and convert 40% of trials to paid.

Illustration	A: optimized for installs	B: optimized for trials
Cost per install	$1.50	$3.00
Installs	2,000	1,000
Install to trial	5%	15%
Cost per trial	$30	$20
Paying subscribers	40	60
Cost per subscriber	$75	$50
Two-year revenue after store fee	$3,536	$5,304
Inference cost	$1,366	$1,649
Revenue minus inference	$2,170	$3,655
Return per $1 of ad spend	$0.72	$1.22

A has half the cost per install and still loses $830. Without inference it would look fine, with $3,536 back on $3,000. B makes $655 over two years, but its first year brings back $2,755 after inference against $3,000 of spend, so it pays back only when renewals come in.

Work backward from B's funnel and it stops making money at about $3.65 per install, $24 per trial or $61 per subscriber. Those are your caps, and they move with your model bill. Double every inference cost in the example and B's return falls from $1.22 to $0.67 per dollar.

No ad platform sees your inference bill, so count it as a variable cost in the contribution and payback math from ROAS, MER and POAS.

Which app campaigns should you run?

On every network, optimize for the deepest event that still gets enough volume, usually trial start, then purchase or value. Install optimization finds people likely to install, and that's all it promises.

Network	Deepest optimization
Apple Ads	Installs
Google App campaigns	In-app action (target CPA) or value (target ROAS)
Meta app campaigns	App event, such as START_TRIAL, or purchase value
TikTok app campaigns	In-app event (AEO) or value (VBO)

Apple Ads

Apple Search Ads became Apple Ads on April 14, 2025. Apple describes cost-per-tap pricing for ads on the App Store, while Apple Ads Basic charges per install. Maximize Conversions, added for search results in February 2026, bids toward a target cost per install. That's still installs.

So manage Apple Ads by keyword on downstream data in a Manage Bids campaign, because Maximize Conversions relies on Search Match and doesn't let you manage keyword bids. The AdServices token returns campaign, ad group and keyword IDs without identifying the user or device, but omits the keyword ID when Search Match is on. Join those IDs to trial starts and cut keywords whose cost per trial sits above your cap.

Google App campaigns

App campaigns cover installs, engagement and pre-registration. Bid with target CPA on an in-app action such as trial start, or with target ROAS, which requires the Google Analytics for Firebase SDK and events that pass a revenue value. Google suggests a daily budget of 50 times your target CPI or 10 times your target CPA, so a $20 trial target needs about $200 a day per campaign.

Meta app campaigns

Meta's ad set settings include goals for installs and for maximum purchase value, and app event optimization targets a specific in-app action. The events you can target include START_TRIAL and SUBSCRIBE. Meta has replaced the separate Advantage+ app campaign workflow with one Advantage+ setup for sales, app and leads. For iOS users who opt out of tracking, it measures app events with Aggregated Event Measurement through its iOS SDK.

TikTok app campaigns

TikTok's App Event Optimization and Value-Based Optimization go past installs to in-app events and purchase value. For VBO, it suggests a daily budget of 30 times your CPA and at least 7 days of learning. On iOS, it says AEO campaigns rely on SKAdNetwork and should each return at least 90 installs a day, or conversions may be withheld and reported as null. Smart+ App campaigns can optimize for installs or in-app events on iOS and Android, but value optimization and Cost Cap bidding are Android only.

How do you measure installs on iOS and Android?

iOS: ATT, SKAdNetwork and AdAttributionKit

On iOS 14.5 and later, you need permission through App Tracking Transparency to track a user or read their advertising identifier. For users who don't opt in, networks get Apple's privacy-preserving postbacks. SKAdNetwork is on version 4, which needs iOS 16.1 or later. AdAttributionKit arrived in iOS 17.4, and the two interoperate, so only one ad wins a conversion across both. Apple's SKAdNetwork page now tells developers to use AdAttributionKit for app ad campaigns.

Apple's update log shows June 2025 brought postbacks in multiple conversion windows, a country code in postbacks and attribution rules, which let you set view windows of 1 to 7 days and click windows of 1 to 30 days. Network support lags. Google's iOS guide, last updated August 19, 2026, says Google doesn't support AdAttributionKit for attribution yet and asks you to consolidate to 8 or fewer iOS App campaigns for install.

Conversion values for a trial app

AdAttributionKit sends postbacks for three windows: days 0 to 2, 3 to 7 and 8 to 35 after first launch. Only the first can carry a fine value, and only at the higher privacy tiers. The others carry a coarse value. As in SKAdNetwork 4, fine values run from 0 to 63 and coarse values are low, medium or high. The first postback arrives 24 to 48 hours after its window ends, the others 24 to 144 hours after theirs.

A 7-day trial converts in the second or third window, so through postbacks the paid event only ever arrives coarse. Use the fine value for trial start, paywall view and early usage, mapped to predicted revenue, and save coarse values for paid conversion and renewal. Fewer, bigger campaigns help you clear Apple's privacy thresholds.

Android: install referrer and Privacy Sandbox

On Android, the Play Install Referrer API gives you the referrer URL and the click and install timestamps for each install. Privacy Sandbox won't replace it. On October 17, 2025, Google said it would retire the Attribution Reporting API, Topics and Protected Audience on Chrome and Android, plus SDK Runtime and Protected App Signals, citing low adoption. Its status page marks every Android technology as deprecate and remove.

What a mobile measurement partner does

Once you run two or more networks, use a mobile measurement partner (MMP). It collects the Apple Ads token, the Play referrer and postback copies, which Apple lets your app receive on your own domain, then decides which network gets credit, decodes conversion values and sends trial and purchase events back to each network. TikTok recommends an MMP for app marketers and ended its legacy MMP integration on March 31, 2025, in favor of a self-attributing network setup.

Web-to-app funnels and the US link-out rules

What changed in the US

April 30, 2025: the district court held Apple in civil contempt of its 2021 injunction and barred it from charging any commission on purchases made outside the app.
May 1, 2025: Apple updated its guidelines so US storefront apps can include buttons, external links and other calls to action.
December 11, 2025: the Ninth Circuit affirmed the contempt finding but said Apple may charge a commission for costs "genuinely and reasonably necessary" to coordinate external links, plus some pay for its intellectual property, once the district court approves a fee. Apple may also stop developer links from being more prominent than its own.
May 6, 2026: Justice Kagan denied Apple's request to stay the Ninth Circuit's mandate.
June 30, 2026: the Supreme Court agreed to hear Apple's appeal, limited to whether civil contempt can rest on an injunction's "spirit." Apple filed its brief on September 14, and Epic's is due November 13.
August 13, 2026: Justice Kagan refused to pause the district court proceedings, and Apple proposed commissions on linked purchases of 15% for standard apps, 10% for subscription renewals and three partner programs, and 5% for Small Business Program apps.

As of 4 October 2026, Apple's App Review Guidelines still exempt the US storefront from the link ban, and I found no order approving a fee. So US web sales through a link pay Apple nothing today, but model a commission too, since Apple proposed 5% to 15%.

Outside the US

In the EU, Apple's unified terms took effect October 1, 2026. Linked out-of-app offers pay a 15% store services commission, or 10% for Small Business Program members and subscriptions after their first year, on sales within 7 days of the link tap. Apple lists the same rates and window for Japan and Brazil. Other storefronts still bar links to other ways to pay, apart from specific entitlements.

On Android

Google's lower Play service fees, announced March 4, 2026, took effect in the US, EEA and UK on June 30, 2026. Subscriptions pay a 10% service fee, plus 5% for Play Billing. In the US, the external content links program lets enrolled apps link to web checkout, but Google charges the same 10% on subscription purchases within 24 hours of the tap, reported and paid from October 1, 2026.

How to build the funnel

Web-first funnels send ad traffic to a web paywall and then into the app to sign in. Link-out funnels send users from the in-app paywall to web checkout. Either way, guideline 3.1.3(b) lets people use in your app what they bought on your website only if those items are also available as in-app purchases, so keep in-app purchase live.

The case for web checkout is margin and measurement. On the $80 plan, Apple's 15% is $12, while Stripe's standard US card rate of 2.9% plus 30 cents is $2.62, or $3.58 with Stripe Billing's 0.7% and Stripe Tax's 0.5% added. Your pixel and server events also report each web buyer and amount to Meta, Google and TikTok directly, not through delayed, coarse postbacks. The cost is extra checkout steps, so test web checkout against in-app purchase before you switch.

Trials, paywalls and a test plan

RevenueCat's State of Subscription Apps 2026, covering more than 115,000 apps and over $16 billion in revenue, mostly from 2025, gives you baselines. Trials of 4 days or less converted to paid at a median 25.5%, against 42.5% for 17 to 32 day trials, and 55% of 3-day trial cancellations happen on day 0. Hard paywalls turned 10.7% of downloads into payers within 35 days at the median, against 2.1% for freemium. AI-powered apps earned 41% more revenue per payer but churned 30% faster.

In an AI app, a longer trial burns more compute per trial, and a free tier burns it on people who may never pay. Judge paywall tests on revenue per install after store fees and inference, not conversion rate alone.

Log the funnel: install, first AI result, paywall view, trial start, trial cancel, paid conversion, refund and renewal, plus inference cost per user by model.
Set caps from your unit economics: maximum cost per install, per trial and per subscriber, by platform and country.
Test one variable at a time, in this order: hard paywall against freemium, trial length (say 3 days against 7), plan and price, then web checkout against in-app purchase for US users.
Randomize users with your paywall tool or remote config, size the test before you start, and freeze campaign changes while it runs.
Run each test through the full trial plus a week. For a 7-day trial, read trial to paid on day 8, then check renewals.
Measure install to trial, trial to paid, day-35 download to paid, refunds, inference per trial and per subscriber, and revenue per install net of store fees and inference, with day-7 retention as a guardrail.
Ship the winner, switch each campaign's optimization event to match, and update your conversion value mapping.

If you want help

For AI startups, our growth audit reviews your ad accounts, tracking, creative and landing pages against margin, then sets breakeven and target numbers per channel in a prioritized plan. It costs $1,000 and takes two to three weeks. Book a 30-minute call.

Sources

Facts checked on 4 October 2026.