What is CPI?

A campaign spends A$20,000 in a month and the attribution provider credits it with 10,000 installs. CPI is 20,000 ÷ 10,000 = A$2.00. If 30% of those installs open the app twice and 4% ever pay, the cost per paying user is 20,000 ÷ 400 = A$50. The A$2.00 figure is what the campaign report shows; the A$50 figure is the one that has to be compared against what a paying user is worth.

Cost per install divides what an app campaign spent by the number of installs credited to it. It is the headline metric of app advertising, the easiest number in the category to move, and one of the least informative about whether the money worked.

The first complication is who did the counting. An ad platform credits installs under its own attribution rules, and a mobile measurement partner credits them under a different set. The two will not agree. The disagreement is structural rather than a bug: different windows, different treatment of view-through, different deduplication when several networks all claim the same install. Both figures can be correct within their own definitions and still differ by a third. An account that has not chosen one of them as the system of record will spend its reporting time reconciling rather than deciding.

The second complication is that the install is an intermediate. Nobody has revenue because an app was installed. The install is the first of a chain — open, register, activate, retain, pay — and the correlation between the first link and the last is weaker than the prominence of the metric suggests.

The decomposition is the familiar one, with an extra step that is not an advertising setting at all:

Cost per install = cost per click ÷ store conversion rate

The store page sits between the click and the install, and it converts at a rate the ad account does not control. Icon, screenshots, the first two lines of the description, the rating, the download size and the number of permissions requested all move it. Two campaigns with identical click prices can have cost-per-install figures a long way apart because one of them sends traffic to a better listing.

This is worth stating plainly because it is routinely missed: a meaningful share of the movement in cost per install has nothing to do with the campaign. Store-listing work is frequently the cheapest available improvement, and it improves every campaign at once.

Any market that pays for an event will produce a supply of that event at the lowest possible quality, and app installs are the clearest case of it in advertising.

Rewarded and incentivised placements pay a person something — in-game currency, a life, a hint — for installing. The install is real, the device is real, and the person has no interest in the app whatsoever. Prices on this inventory are a fraction of what non-incentivised traffic costs, and an account managed on cost per install alone will drift toward it automatically.

Even without incentives, the spread across networks is wide, and most of it is inventory quality rather than auction efficiency. Installs sourced from low-value apps, from audiences selected for a propensity to install anything, or from placements where the ad is hard to distinguish from the content all come cheap.

The defence is not to police networks one at a time. It is to stop optimising on the install.

The single largest improvement available in most app accounts is moving the optimisation event past the install.

A registration predicts value better than an install. A completed tutorial predicts better than a registration. A first purchase predicts best of all. Every step deeper reduces the volume of events the bidder learns from, which is the constraint: below the platform's own threshold the model has too little to work with and delivery becomes erratic.

The practical rule is to optimise on the deepest event that clears that volume, and to revisit it as the account grows. An account producing four hundred registrations a week should not be buying installs.

Value-based optimisation, where the account sends back the revenue rather than the event, goes one step further and is worth the engineering when purchase behaviour is skewed — which in apps it almost always is, because a small share of users produce most of the revenue.

A cost-per-install figure read without retention beside it cannot distinguish the two outcomes the account cares most about.

Day-one retention says whether the app was opened again at all. Day-seven says whether it survived the first week of competing for attention. The share still present at day thirty is, in most categories, the closest thing to an early read on lifetime value.

Two campaigns at the same cost per install with day-one retention of 40% and 12% are not comparable in any sense. The second one is buying a number. Splitting retention by acquisition source is the report that exposes it, and it is the report that justifies paying more per install for better traffic — which is usually the correct decision and always the uncomfortable one.

Organic uplift. Paid installs move store rankings, and better rankings produce installs nobody paid for. That effect is invisible in the campaign's own figure, which both understates the value of good paid activity and overstates the value of cheap volume that never ranks.

Cannibalisation. Some of the installs a campaign is credited with would have happened anyway, particularly on branded search and retargeting. A holdout or geo test is the only reliable way to size either this or the uplift above.

Fraud. Install fraud is a mature industry. Click injection, click spamming and device farms all produce installs that cost money and deliver nothing, and they are attracted to exactly the campaigns that reward raw install volume.

Work the store listing, because it sits between every click and every install and it is usually under-tested.

Then move the optimisation event as deep as the volume allows, and send the platform the post-install signal it needs to learn on.

Then split by source, by placement and by creative, and judge each on retention and revenue per install rather than on the install price. Expect the ranking to change when you do.

Then set the ceiling from the business: what a paying user contributes, multiplied by the share of installs that become paying users. That number is the most an install can cost, and it is the only benchmark that belongs on the page.

• Optimise on a post-install event as soon as there is volume for it. An install is the cheapest event in the funnel to buy and the least correlated with revenue; a registration, a tutorial completion or a first purchase all predict value better.
• Read cost per install beside day-one and day-seven retention. A campaign that buys installs which never open again is buying a number, and the number is the only thing it delivers.
• Split by source before drawing a conclusion. Incentivised inventory, rewarded video and some ad networks produce installs at prices no other channel can match and quality no other channel would accept.
• Watch the gap between platform-reported and attribution-provider installs. They will not match, and the size and direction of the gap is itself a diagnostic on the measurement rather than on the campaign.
• Test the store listing with the same energy as the creative. Icon, screenshots, the first two lines of the description and the rating all sit between the click and the install, and none of them is an advertising setting.