Anatomy of a Q4 CPM spike

The Q4 CPM spike that UA teams brace for every autumn carries the wrong name. Most of what shows up as a higher CPM in your November dashboards isn't a higher price for the impressions you were buying in September; it's a different set of impressions, won at a different rate, against a different set of bidders. Treating it as inflation leads to the standard response, which is to pause or cut, and that response is usually wrong for exactly the campaigns that receive it.

The argument here is that a Q4 CPM move should break down into three components before anyone touches a budget: the price of the same inventory, the shift in which inventory you win, the shift in how often you win it. Teams that skip the decomposition end up pausing campaigns whose underlying unit economics barely moved, while keeping campaigns whose cohort quality quietly collapsed behind a stable-looking CPM.

Who actually arrives in the auction

The in-app auction in Q4 doesn't simply get more crowded. It gets crowded with a specific kind of bidder. Retail and e-commerce advertisers bid against a near-term purchase value that peaks in the six weeks before Christmas, and they bid with that value in mind, whereas a game advertiser bidding to a D7 ROAS target has a value per impression that barely changes between September and November. The retailer's does.

This year the composition shift has two extra ingredients, both flagged earlier in "GTA 6's autumn campaign will distort mobile CPMs. Plan for it now". Take-Two has confirmed the marketing push for the November 19 release begins this summer, and AppLovin's Axon Ads Manager has been open to self-serve e-commerce advertisers globally since late June. Neither is a game bidding against your ROAS model. Both are demand for the same screens.

The mechanism that matters is how a value-based bidder loses, and it doesn't lose evenly. When a higher-value bidder enters, your bid stays flat because your model hasn't changed, so you lose the most contested impressions first. Those are, on average, the impressions that were attracting the most bidders because they sat in front of the most valuable users. Your win rate on the best inventory falls. The mix of impressions you do win drifts towards the pockets nobody else wanted.

That drift is why CPI often rises faster than CPM in Q4. The CPM on the impressions you still win may have moved modestly, but the install rate on that residual inventory is worse, because it is worse inventory.

The three-part decomposition

Take one campaign on one network in one geo and compare a stable September fortnight to the November fortnight in question. Then split the CPM change into three parts.

Price on like-for-like inventory comes first: hold the placement and publisher mix constant and compare clearing prices. This is the genuine inflation component. Mix shift is second, meaning how much of the CPM change comes from winning a different blend of placements and publishers at different hours of day. Win rate shift is the third, meaning how the share of auctions you entered and won has moved, and on which inventory.

Most networks won't hand you a clean decomposition, but the pieces exist. Publisher-level reports give you mix. Bid distribution or win rate reports, where offered, give you the third component. Where a network offers neither, the mix component alone is often enough to change the decision.

An illustrative example, with invented figures for the arithmetic only. A puzzle campaign moves from a 12 unit CPM in September to 18 in the second week of November, a 50 percent increase. Holding the September publisher mix constant, the like-for-like CPM is 14, so roughly a third of the rise is price and two thirds is mix and win rate. The CPI has moved from 2.0 to 3.4. If the D7 ROAS of the November cohort has held, the campaign is paying a higher entry price for the same quality and the decision is about payback tolerance. If D7 ROAS has fallen by a similar proportion to the install rate, the campaign is buying the leftover inventory and the correct response is to change targeting or creative rather than to cut the bid.

Two different problems, one dashboard number.

The response that makes January worse

Here is the second-order effect most coverage skips. Because nearly every game buyer reads the spike the same way and reacts on the same calendar, the reaction itself becomes a market event: budgets come off in the third week of November, competition among game bidders thins, and the residual inventory that game advertisers are left with gets cheaper for the few who stay.

Then the retail demand vanishes on December 26. The same buyers who paused come back in mid January all at once and rebuild their learning phases together, so the January trough that everyone celebrates is shorter and shallower than a staggered re-entry would produce.

A team that has decomposed its spike can behave differently. If the like-for-like price component is small, it can hold spend through late November on the campaigns whose cohort quality held, accept a longer payback for six weeks, and be fully warm on December 26 when the retailers leave. The cost of doing this isn't the higher CPM. It's the working capital tied up in slower payback, plus the risk that quality degrades in the weeks you aren't watching closely.

A decision rule for the review meeting

Write the rule down before the spike arrives, because judgement under a red dashboard is poor.

• If like-for-like price explains most of the CPM rise and D7 quality holds, keep spending and extend payback tolerance by a fixed, pre-agreed number of days.
• If mix and win rate explain most of the rise and D7 quality has fallen, don't cut the bid. Change what you're bidding on: new creative angles that convert on the residual inventory, or a targeting change that moves you back into pockets you can win.

There is a third case. If both components are large and quality has fallen, that campaign has lost its place in the auction for the season; pause it and redeploy to channels where the retail overlap is smaller, typically the placements with less e-commerce demand, such as rewarded formats inside other games. That's the only case where the standard pause is correct, and in most portfolios it applies to fewer campaigns than get paused.

One more trade-off worth stating plainly. The decomposition takes analyst hours during the busiest weeks of the year, and the win rate data is patchy across networks. Teams that can't do the full version should at least do the mix component on their top three campaigns by spend. Even that partial view tends to show the same thing: the campaign everyone blames for the CPM rise isn't the one whose economics actually broke.

Related archive reading

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

• The auction, not the audience, sets your CPI
• What Cyber Week Cohort Quality Analysis Actually Shows