Understanding conversion optimization in digital advertising

One of the primary innovations in digital advertising over the past few years — especially as a contributor to commercial value for advertisers — is “event-based bidding,” or conversion optimization. I wrote extensively about this practice when Meta (then, Facebook) introduced the App Event Optimization (AEO) and Value Optimization (VO) campaign strategies in 2016 and 2017, respectively, but the conversion optimization certainly isn’t unique to Meta; most scaled ad platforms and ad networks offer it as a campaign optimization strategy. Background on this topic can be found in these pieces:

I view conversion optimization as the practice of allowing an advertiser to bid on a specific, in-product outcome — for instance, a purchase. Marketers often speak of a conversion “funnel”; this funnel concept is really just a sorted list of completed actions by users in the product across some time frame, with the most frequently-occurring action representing the “mouth” of the funnel and the rarest action representing the far, pointed end. The diagram below is a hypothetical user acquisition funnel for a consumer app: “Impressions” is a count of every user that saw an ad for the product, and “Purchase” is a count of the number of purchases made by that group of users. Tautologically, every user who saw an ad for the product “converted” on the basis of impressions; only a small proportion of users who saw an ad for the product made a purchase.

Conversion optimization allows an advertiser to bid on impressions against “down-funnel” objectives. It’s common for advertisers to optimize against purchases, since those represent transactions and thus revenue, but an advertiser might optimize against any instrumentable event (meaning any countable event that can be observed by the product), using that event as a proxy for future transactions or ultimate user value. Ad platforms usually encourage advertisers to optimize against events that the platform defines in a standard library — meaning events that are nearly ubiquitous across apps and websites, such as “purchase,” “add to cart,” or “registration” — but it’s often possible for advertisers to define custom events against which campaigns are optimized, too. Using custom events can deprive the platform of the volume of conversion data needed to optimize a campaign. Rare events can take longer to optimize for simply given a paucity of data; the purpose of a standard events library is to share optimization insights across advertisers’ similar products.

When an ad platform or network allows conversion-based optimization, it effectively tells the advertiser that it will only charge them for the successful delivery of those events (this isn’t strictly true, especially early in a campaign’s lifecycle when impressions have been served but no conversion events have been registered, but it’s true enough to be thought of as a rough guideline). This is obviously attractive to the advertiser, which can determine how much any event is worth to them, bid up to that amount, and only pay the ad platform when those events transpire. This is essentially risk-free money if the advertiser feels confident that its bid pricing model is statistically reliable and rigorous: each time the advertising platform delivers some event, the advertiser pays that platform less than what that event is worth to them.

Of course, risk can’t be eliminated, only transferred or accepted. If the advertiser transfers conversion risk to the platform, then the platform’s commercial prospects are determined by its ability to price and absorb that risk. Advertisers buy ad impressions through an auction process; an exploration of advertising auctions is beyond the scope of this piece, but this podcast provides a helpful overview. Given a finite number of impressions and multiple conversion events that can be bid against, a platform must use conversion probabilities to derive the expected values of competing ads to determine the winner of any auction.

The expected value of an advertiser’s ad being filled into an impression is not equivalent to that advertiser’s bid: it’s the advertiser’s bid adjusted by the probability that the ad will lead to a conversion since the advertiser will pay nothing to the platform if it doesn’t. An ad platform’s ability to manage and scale conversion-optimized campaigns is predicated on its capabilities related to conversion probability determination. Each impression presents an opportunity cost risk: if a platform fills an impression with an ad that doesn’t convert, it loses the potential revenue from any ad in the auction’s bid stack that would have (albeit for less revenue).

If an advertiser’s ad delivery — meaning, the number of impressions its campaign fills — is unsatisfactory, assuming sufficient scale on the platform for the audience being pursued, it’s because the expected value of the advertiser’s ads is too low to win auctions. The advertiser faces two paths to delivery improvement: increase its bid, or increase the conversion rate of its ads. Bids should (theoretically) be constrained by the monetization of the product, which is difficult to alter significantly in the short term. Ad creative experimentation and exploration can improve conversion rates, which is partly why the advertising use cases of Generative AI pose such immense promise.