How to actually lower your CPI in 2026

Cost per install is rising and it isn't your imagination: CPIs on Meta and TikTok climbed 15 to 20% year over year, and the global averages now sit near $1.20 on Android and $3.60 on iOS, with gaming overall around $3.90 and finance apps paying upward of $8.70. Five levers reliably pull the number down. Ranked by effect size on the accounts we run: creative velocity, geo and platform mix, fraud filtering, bid-structure hygiene, and placement choice. This guide takes them in order, with the honest caveat that the first lever outweighs the other four combined.

Lever 1: creative velocity (the big one)

Ad auctions price attention, and platforms discount ads that earn it. Install rate, IPM in the trade, is therefore a direct CPI discount: double it and your effective cost roughly halves before a single bid changes. The catch is decay. Every concept fatigues, and the market has adjusted its metabolism accordingly: 78% of top-quartile campaigns now refresh creative at least weekly, nearly double the previous year's share. In practice that means a standing pipeline of concepts, a kill rule you obey, and format courage. Video carried 53.7% of gaming ad impressions in 2025, but playables, at 13.3% and climbing, convert impressions to installs at roughly 4.8 per thousand against video's 2.9. Our full method is published in the testing framework guide.

Lever 2: geo and platform mix

Averages hide arbitrage. US casual installs cost $3 to $8 while the same genre in tier-2 markets can cost a fifth of that, and Android undercuts iOS nearly everywhere. The lever is not "buy cheap geos", it's alignment: if your game monetizes globally through ads, cheap volume geos are rational; if revenue concentrates in high-ARPU markets, cheap installs are a vanity discount. Rebalance quarterly, cap unproven geos at a fixed share of spend, and read cohorts, not CPIs, when judging a market. The hybrid-casual storyincludes the geo that looked heroic for two weeks and then drowned the payback line.

Lever 3: fraud filtering

Roughly 22% of non-organic installs industry-wide carry fraud signals, up from about 17% in 2022, with exposure estimated above $5.4 billion a year. Fraud concentrates precisely where CPI looks best, because fake installs are cheap to fabricate. Turning on your MMP's fraud rules, maintaining network blocklists, and refusing inventory that won't survive a cohort read typically trims low-single-digit percentages of spend immediately, and more at scale. Note the newer twist: fraud has migrated toward fake organics, which now account for about half of fraudulent installs, so a suspiciously blooming organic curve deserves the same skepticism as a cheap network.

Lever 4: bid-structure hygiene

Nothing here is glamorous. Campaigns bidding for installs when the business needs payers; target CPAs set from hope rather than cohort data; budgets fragmented across so many campaigns that none exits learning; season-old bid caps nobody remembers setting. Each costs a few percent, and accounts accumulate all of them. A quarterly structural audit, one afternoon of unsentimental work, routinely returns 5 to 10% in blended CPI on accounts that haven't had one.

Lever 5: placement and format choice

Within one network, placements can price installs apart by multiples. Rewarded inventory delivers famously cheap installs whose users arrived for the reward, not the app, so its place in your mix depends entirely on how your product monetizes attention. Interstitials, banners-to-video, and store-native formats like Apple Ads' product-page placements each carry distinct intent profiles. The rule that generalizes: judge each placement by its cohort's day-7 behavior, never its CPI, and prune annually, because placements drift.

What we deliberately left off the list

Incentivized bursts, review-swap schemes, and anything promising installs at a tenth of market rate. They lower CPI beautifully and poison every number downstream of it. If a vendor's pitch only makes sense when you ignore retention, the pitch is the retention problem.