Risk in online advertising

I realized this morning that Google, once strictly a cost-per-click ad network, is now making use of cost-per-impression (CPM), cost-per-click (CPC), and cost-per-action (CPA) advertising models, depending on who’s being asked to take the risk. As Google should be. A few months ago, I was convinced that online advertising was moving inexorably from CPM through CPC to a CPA model, in line with increasing amounts of measurability and advertiser-demanded accountability. Recently I’ve come to believe the most appropriate pricing model depends entirely on who’s doing what to whom – he who takes the risk gets to make the decisions.

In cost-per-impression advertising, since the publisher gets paid based on ad views, the advertiser assumes all the risk. It’s possible that not a single person will take the action desired by the advertiser, resulting in a 100% loss. Therefore this model is most appropriate for situations where the publishers have no control over the advertisements appearing on their sites – the advertisers, who are risking all of their money, make the decision where to appear. Traditional banner ad sales works like this; so does Google’s Onsite Advertiser Sign-Up program.

Cost-per-action advertising is entirely the opposite – there’s no risk for the advertiser, since they only make a payment when an action occurs that’s worth more to them than the payment itself. Here the publisher assumes all the risk – if not a single person sees the ad and takes the action desired by the advertiser, the publisher gets nothing, and loses ad inventory he could have put to use elsewhere. This model is therefore most appropriate for situations where the publishers decide what advertisements appear on their sites, and the advertisers have very little to do with it. Traditional affiliate ad networks work like this; so does Google’s AdSense Referrals program.

Finally, cost-per-click advertising divides the risk between advertiser and publisher. The clicks advertisers get aren’t guaranteed to convert into more desirable actions, and could possibly be fraudulent; the ads publishers show aren’t guaranteed to be clicked on. This is Google’s bread-and-butter; they attempt to minimize the risk to both sides by ensuring advertisers and publishers are appropriately matched, showing ads only where they’ll work.

Google has a good understanding of risk in online advertising; in each situation it uses the model best suited to the circumstance. Of course, the CPC model is the hardest to manage, since you’ve got to balance the risks of two parties against each other. Google’s done a pretty good job – arguably the best job – minimizing the risk to both sides. The billion-dollar-question – why does Google do this the best? The standard assumption – oh, it’s Google, and they’re so brilliant – might not be true; I’ve been suspecting more and more that Google’s advantage is primarily due to the size of its network. When you’ve got thousands upon thousands of advertisers and publishers, matching them is relatively easy; depth and variety can compensate for any flaws in your matching algorithm. A new ad network that can match advertisers with publishers ten times better than Google is still going to perform miserably in comparison if they’ve got a limited publisher base – those Viagra ads have to go somewhere, even if none of your sites is a good fit. Would-be competitors to Google therefore need to accumulate more partners, not out-algorithm. Which is still a daunting problem, but it’s a business problem rather than an engineering one.

To sum up:

• When advertisers pick publishers, publisher risk needs to be mitigated, and therefore the CPM model works.
• When publishers pick advertisers, advertiser risk needs to be mitigated, and therefore the CPA model works.
• When advertisers and publishers are matched by a third party, both sides assume risks, so a CPC model is a decent compromise.
• When advertisers and publishers are matched by a third party, I suspect the size of the network matters at least as much as the matching algorithm.
• Increasing the size of one’s network of partners is primarily a business problem, not an engineering one.