What the Machines Can See, in plain language

Read this first

This is the short, plain version of a long research paper. The full paper is called What the Machines Can See: The State of Mobile App Growth in 2026. It has the sources, the tables and the math. This page has the same findings in fewer and shorter words. The counts here come from the same build as the paper. The build also checks that the rest of the numbers here are in the paper, apart from years and small counts. Each part links to the chapter that backs it up.

The research runs to 27 September 2026. It looks at the US first, and uses world data where US data runs out.

What app growth is

Growth for an app is more than a new icon on a phone. A person has to find the app, get it, start to use it, keep using it, come back after a break, and pay. Ads do a lot of the first part. The ad machines set the bids, pick who sees the ad, choose where it runs and help make it.

They do not set the price, the first few screens, the pay screen or the store's cut. Those decide if an install ever turns into profit.

So this paper does not count installs as growth. It counts the extra money new users leave behind, after store fees, refunds, card costs and the like. Extra means money you would not have made without the ads. Growth is that money minus the full cost of winning the customer. Chapter 1 sets out the terms.

Where the money is

AppsFlyer is a firm that tracks app ads for its clients. From its client data, it puts world spend on app ads in 2025 at about $109 billion. Of that, $78 billion went to win new users and $31 billion went to win back old ones. The US holds about 42% of the spend on new users.

These are estimates from one firm's clients, not a count of the whole market. What people spend in the app stores is a larger, separate pool. Two data firms put its size $11 billion apart. Chapter 2 has the numbers.

What we found

Six findings drive the rest of the paper.

1. The machines learn from what they see, but in 2025 the money went where users were worth more

The ad machines at Google, Meta, TikTok and the app ad firms learn from signals: an install, a sign-up, a sale. The more signals they get, and the sooner, the better they bid. Google's app ads need at least 10 of the chosen actions a day before they can bid for a target return.

Apple lets ad firms see far less on iPhones. In one firm's data, only 35% to 38% of iPhone users who are asked agree to be tracked. For each install it credits to an ad, Apple sends the ad firm up to three short reports. The reports carry no user ID, and each comes days after its time window shuts. Chapter 3 shows how this works.

So you might think the money would move to Android, where the machines see more. It did not. In 2025, spend to win new users grew 35% on iOS and fell 1% on Android. It grew 18% in apps that are not games, and 3% in games. Other things explain the shift at least as well as signal does. One is what users are worth. Another is the mix of app types. A third is a rush of ad money from China-based shopping apps. These are world figures from one firm's panel, not US figures. Chapter 5 has the detail.

2. Win-back ads grew faster than ads for new users, and we found no independent test of them in apps

Some ads chase people who already have your app, to bring them back. The trade calls this remarketing. That spend rose 37% in 2025, to $31 billion. Spend to win new users rose 13%. Win-back ads made up 29% of app ad spend in AppsFlyer's panel.

A fair test splits people at random. One group sees the ads and one does not, and you compare what each group does. Fair tests of win-back ads on the web find that they work, but not by much. For apps, we did not find a public, independent fair test that backs up the returns they get credit for. That does not mean they fail. It means we did not find the proof. Chapter 2 has the spend and chapter 4 has the tests.

3. The app ad firm with the best margin in our table owns an auction, but that is not the whole story

AppLovin took in $5.5 billion of net revenue in 2025. It reported 82% of that as adjusted EBITDA, a measure of profit before some costs. That is the highest margin in our table. Its machine learns from the results of its own auction, where apps sell their ad space.

Owning an auction does not explain the lead on its own. Liftoff owns no such auction and still reports a 55% margin, counted the same way. Unity's comeback followed a new machine more than a new asset. Google and Apple own more of the chain than any of these firms, and they say less about it. Chapter 2 follows the money.

4. Credit is not cause

Every ad firm reports which installs it drove. That is credit. It is not the same as cause, because some of those users would have got the app anyway.

A fair test with a holdout group is the best way to find cause. One large study looked at 663 such tests on Facebook. The researchers then ran two stats methods on the same ads, with the holdout groups left out. Those methods made the lift look three to thirteen times as big as the tests found, at the median. The size of the gap turned on the method and on which step it measured. These were stats fixes, not the credit rules ad firms use.

Two older studies come closer to how credit reports work. At eBay, a standard model put the return on search ads above 4,000%. A 60-day test then turned off some of those ads in about 30% of US regions. It found a return of minus 63%.

The gap does not always run one way. On iPhones, Apple's privacy rules also hide some installs that ads did cause.

The big platforms do offer lift tests. The ones we read about run through an account team. And for ads on streaming TV, we found no independent test of their effect on app installs. Chapter 4 goes through the tests and chapter 8 covers TV.

5. The store's cut is now a lever

Apple and Google take a cut when users pay inside an app. Take a $9.99 monthly plan. At Apple's standard first-year rate of 30%, the app keeps $6.99. At 15%, it keeps $8.49. On a web checkout, at Stripe's posted card rate, it keeps about $9.40. That is about 34% more than at the 30% rate, and about 11% more than at 15%. Both figures come before the cost of running your own billing.

The US rules on paying outside the store changed in 2025. Apps may now send users to pay on the web. How much Apple may charge on those sales was still being set in court at the cutoff.

Who stays counts as much as the cut, and most users do not stay. In a panel of 16,000 games, the median game keeps about 22% of its users to day 1. By day 30 it keeps under 1%. In one panel of apps that sell plans, the median app keeps only 28% of yearly subscribers past the first year. The first year is paid. Most renewals never come. Forecasts of what a user will be worth fill the gap. But public proof that those forecasts hold up for apps is thin. Chapter 7 does the math.

6. "Apps DSP" is a checklist, and AI agents move the controls

"Apps DSP" is a term from the author's own playbook. It means one tool that does all the jobs: win new users, win back old ones, bid on what users are worth over time, and reach TV and shopping. Is it a real category? We made our own list of seven things such a tool should do well. It should bid for value, win back users, run lift tests and give the buyer control, among other things. Just one of the eleven products we scored shows all seven in its public docs. That is Moloco. Five more show at least five of the seven. But showing a feature is not proof that it works better. And no one has shown in public that the full bundle wins. Chapter 11 tests the idea.

The next shift is AI agents. AI already sets the bids. Now agents can plan and change campaigns too. Google, Meta and TikTok have each put out servers that let AI agents work with their ad systems. Google's can only read. Meta's and TikTok's can create and change campaigns. Meta's makes new campaigns paused, so they do not run until they are switched on. Neither shows a step where a person approves edits to a live campaign. So the buyer has to set the limits, the logs and the data rights. Chapter 11 has the detail.

The firms, in bands

The paper scores 11 products that buy app ads on eleven skills. It scores 5 big channels, such as Meta and Google, the same way as a benchmark. The skills cover how a product bids and what it can measure on iPhones. They cover lift tests and fees. They cover what you can see and what you can control. A score counts only what a firm shows in public. It says nothing about how well its ads work. Two separate AI passes scored each product, and fixed, written rules settled the cases where they differed.

A few things stood out:

• Bidding for value is the best-shown skill. 9 of 11 products show it.
• Fees are the worst. Only Kayzen puts out a fee schedule.
• No product shows, in its tech docs, that it works with both of Apple's report tools.
• None of the specialist firms shows a standard lift test with its method in full. Meta, Google and TikTok show lift tests, but they run through account teams.

The scores then roll up into one index, the App Quotient, or AQ. It sorts the products into four bands by how much they back up in public. The paper calls them high, documented, partly documented and thin.

The top band has Moloco and Mintegral. The next band has Remerge, Unity, Verve Dataseat ◆, AppLovin, Jampp, Liftoff and Kayzen. The third has Digital Turbine and Smadex. No product fell in the thin band. Mintegral clears the top line by less than one point. Its Reach score rests on 1 of 3 fields. The ◆ marks Verve Dataseat, which is owned by the author's former employer, Verve Group.

Read the band, not the order. We changed the weights in four sensible ways. Moloco stayed in the top band each time, but other products moved by up to 7 places. The AQ still favours the firms that publish the most. To choose a partner, use the full table, not the index. It shows what each product backs up, links the source, and marks what to ask for. Chapter 9 has the scores and the index.

If you buy app ads

Work through five checks, in order. Each one has a pass mark.

1. Get your events right. Send the events that show value from your own server, count each one once, and take out refunds.
2. Compare like with like. Judge a new partner in the same tracking tool you use now, with the same time windows.
3. Run a holdout. Keep the ads from a random group of people, or a set of regions, on at least one campaign. Write the plan down before launch. Make sure the test is big enough to see the effect you expect.
4. Count what you keep. Measure payback on net revenue after fees. Show what you have seen apart from what you forecast.
5. Test the next step up. When you raise the budget, judge the extra spend by what it brings back. Do not judge it by the average return.

On results alone, control over your ads pays for itself only if you can use it. That takes enough spend, clean data and the means to run a holdout. Without a holdout, you are picking between two credit reports, not two platforms. Chapter 10 has the full advice.

Before you let an AI agent run your ads, ask three things:

• Who sets the goal? The agent should chase your own profit measure, not a platform default such as installs.
• Who approves? Changes past set limits should need a person, and each one should be logged where you can see it.
• What leaves with you? Your event data, test results and logs should be yours to take. Your own test history is the learning you can move. Chapter 11 covers agents.

If you sell app ads in the US

The demand is there. In AppsFlyer's panel, the US holds about 42% of world spend on new app users. US spend to win new users on iOS grew 25% in 2025.

For a challenger, what holds it back is more often proof and signal than the sales pitch. The big platforms run their own lift tests. A firm that cannot show cause through a third party has to compete on credit reports. Those favour the big firms, which have more data. Apple's privacy rules also give richer reports to bigger campaigns. So a new firm with little spend per campaign learns more slowly. Chapter 10 has the case.

How this was made

This research was done with AI help, with the author in charge and the method in public. AI agents worked in ten streams. They drew on 621 web addresses and logged 272 findings. Five check passes went back to the sources and tried to prove the findings wrong. Two passes scored the products apart from each other. One pass played a hostile critic and went after the case. Another checked the numbers and the sources again.

ChatGPT, from OpenAI, then reviewed three builds of the paper. Each point of fact it raised was checked against the sources. That help is not an audit, and OpenAI does not endorse the paper. Every change, 119 in all, is in a public log. The research has not been peer reviewed or audited. Chapter 12 tells how it was done.

The author, Evgeny Popov, has worked at Samba TV since May 2025. Samba TV sells data on what people watch on TV, and ways to measure TV ads. It shows up in chapter 8 as the data source in one method that links TV ads to app installs. From 2022 to 2024 he worked at Verve Group, whose Verve Dataseat is scored. He co-leads a work group in the Ad Context Protocol, a standard for AI ad agents. He founded No Fluff Advisory, which puts out this paper and sells an app growth playbook. The paper tests that playbook and in places corrects it. Before that, he held senior roles at Hearts & Science and Lotame. No firm in the paper saw a draft or had a say in what it says.

What this does not claim

• Web proof, used for apps. Most proof of cause here comes from web, search, shop and audio ads, not app installs. Applying it to apps is our inference.
• Vendors' own clients. Spend and retention numbers come from vendors' own clients. They describe those clients, not the whole market, and most are world figures, not US ones.
• Not found is not none. When we say we found no test, we mean we did not find one. A private study may exist.
• Scores track what is published. A firm that says less can look weaker than it is.
• The rules are moving. Court cases, store fees and privacy screens were still in motion at the cutoff. Check the dates before you act.
• The author's position. The author sells advice to firms in this market and works for a TV measurement firm. The rules were written to limit that bias. You should still weigh it.

Read the full paper

The full paper has the sources, the data, the models and each step of the math. Read What the Machines Can See in full. Its data files are in the last chapter. The full paper scores 57.6 on the Flesch reading scale. This page was written to score 85 or more, and its score is measured at the top.