The benchmark report is selling you something

Nobody publishes a fifty-page benchmark report out of public spirit. The MMP that tells you creative volume is up 25 percent also sells the SDK that counts creatives; the intelligence firm reporting that playables nearly doubled their share sells the panel that measures share. The ad network calling gamers affluent and receptive sells the inventory those gamers see. None of that makes the numbers wrong. It makes them selected, and a UA team that reads the chart without reading the incentive has handed a vendor its planning assumptions.

The thesis here is stronger than "be sceptical". A vendor benchmark's median flatters that vendor's customers by construction, which makes it a poor target for yours. If your CPI sits at the report's median, you aren't doing fine. You're doing exactly as well as the average client of a company that had every reason to publish an average its clients could feel good about.

Who is in the sample, and who paid to be there

Start with the mechanism. A publisher builds its benchmark from the data it can see, which is always a narrower thing than the market. An MMP sees its own clients, a market-intelligence firm sees its panel plus modelled estimates, an ad network sees its own auctions. Every sample has a shape.

AppsFlyer's State of Gaming for Marketers 2026, published in January with Unity and Newzoo, drew on roughly 9,600 gaming apps and 24.8 billion installs. Big sample. It is also, by definition, a sample of apps that pay for an MMP, which means apps already spending enough on paid acquisition to need one. The long tail of games that never scaled is missing by construction, and adding it back would drag the medians on CPI, on retention, on creative count, in a direction the report cannot show.

Sensor Tower's State of Gaming 2026, out in February, works differently: panel data plus estimation, covering console and PC as well as mobile. Its strengths are breadth and store-level revenue. Its weakness is that it cannot see inside your campaigns, so every UA metric in it comes from the outside, inferred rather than counted. When Sensor Tower says installs fell about 7 percent while IAP revenue rose 1 percent, that's a strong market-level observation. When the same firm's January State of Mobile report puts video at 53.7 percent of creative share, ask what the share counts: impressions the panel captured, not spend you'd recognise.

Then the third category, where the Axon and Kantar consumer study from March sits. A commissioned survey with a commercial purpose stated in its own press release, and we said as much in "AppLovin wants brands. The Kantar study is a pitch deck." The numbers may well be accurate. Whoever benefits from the answers chose the questions.

The second-order effect: benchmarks compress the market

The part most coverage misses is that benchmarks do not only describe behaviour. They change it, in the direction that helps the publisher.

When the AppsFlyer report says top advertisers ship 2,400 to 2,600 creative variations per quarter, three things happen. Studios below that figure feel behind and raise production. Vendors selling generative tooling quote the figure in their decks. And the platforms whose auctions consume those creatives get more supply of ad units, which improves both their model training and their fill. The number becomes the target, the target becomes the norm, and the party that gains from everyone chasing it wrote the norm.

The same compression happens with CPI. Publish a median CPI for puzzle in the United States and every puzzle buyer sitting above it gets a difficult conversation with finance. The response is to bid down, to take lower quality traffic, or to move to whichever channel the report implies is cheap. That's a market-wide correlated move, and correlated moves in an auction don't produce the median. They move it.

One practical consequence. A benchmark's medians hold up best for the quarter they describe and worst as a target for the quarter you're planning, because publishing them is itself an input into the prices everybody pays afterwards.

A decision rule for reading any report

The rule we use on the desk has four steps and takes about ten minutes per report.

• Name the product. Write down in one sentence what the publisher sells and to whom. If you can't, you don't yet understand the report.
• Locate your game in the sample. Are you the kind of client this company has? A mid-size hybrid-casual studio reading a report built from enterprise MMP accounts is looking at its competitors' bigger cousins.
• Find the number that doesn't help the publisher. Every honest report contains at least one. Sensor Tower's installs-down figure hurts its own volume narrative, and AppsFlyer's ATT and organic ratios don't flatter paid measurement. Those are the numbers most worth trusting.
• Convert medians to floors. When a sample skews toward well-resourced advertisers, its median is roughly what a competently run operation manages. Treat that as the minimum acceptable, then set your internal target from your own cohort economics.

A worked illustration

Take an illustrative mid-core studio reviewing its Q3 plan. A vendor report puts the genre's median D7 ROAS at 18 percent; the studio's own blended figure is 17. The naive reading is "on par, hold spend".

Apply the rule. That sample is the vendor's client base, which skews toward well-funded advertisers, so 18 percent is roughly the median of the well-funded. The studio's 17 percent sits a point below the floor rather than a point below par. So the question becomes which channels sit under 17, and whether anyone has refreshed the creative slate since the report's data window closed.

Check the date too. A report published in January describes Q4 of the prior year, when holiday CPMs distort every ratio going. The studio is planning a summer quarter, with a large console campaign due to enter the auction. Those two windows barely overlap.

Then the inconvenient number. If the same report notes paid install share rising while organic falls, the studio is holding 17 percent against a rising paid dependency, which means its blended ROAS flatters its paid ROAS. That matters more than the headline median.

What to do with the report anyway

None of this argues for ignoring vendor research. The publishers see more data than any single studio ever will, and the free reports are the closest thing this industry has to a shared vocabulary. What changes is the order you read them in: incentive first, then sample, then the numbers that embarrass the publisher, with the headline chart last and lightly.

One more habit worth adopting. Keep a short internal note per report recording what it claimed, the date it landed, and which sample it came from. Two or three cycles in, compare the vendor's medians against your own trajectory and see whether the report leads your numbers or lags them, or whether there's no relationship at all. The last of those is more common than the first, in our experience, and knowing it is the cheapest calibration a market-intelligence function can buy.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• AppLovin's Ad Tech Pivot: Selling Apps to Tripledot
• The hook is a promise; retention is the invoice