Recently on the Internet Oldtimers List, someone posted a link to a video mashup where someone had taken a clip from the movie "A Few Good Men" and replaced the famous "You can't handle the truth!" dialogue between Nicholson and Cruise with a farcical semi-humorous debate about demand-side platforms (DSPs). What was interesting about this clip was that its central argument was that DSPs lower the CPM of premium publishers' impressions (with Cruise arguing for the premium publisher and Nicholson arguing for the DSP).
The video is cute -- pretty well done, and worth a view if you're someone on the inside of this particular space online. But what really surprised me about it was that very few people seem to really understand what's happening with DSPs in general -- and there's obviously misinformation going around. This particular debate about DSPs lowering the yield of publisher impressions was one I hadn't heard articulated before.
So let's get started digging into this by discussing what a demand-side platform really is. These advertiser/agency facing systems let buyers do self-service media buying from publishers; publisher aggregators (sometimes now being called sell-side platforms, or SSPs) like PubMatic, AdMeld, Rubicon, and others; and ad exchanges. The most important part of these mechanisms is that they enable real-time bidding against inventory on these sites. This is really important because in real-time bidding, the DSP can let the buyer specify business rules describing the value of impressions based on their audience attributes. That means the buyer can assign monetary value against specific audiences, and the DSP can bid on every impression in real time based on its actual value to the advertiser.
One reason real-time bidding is so valuable is that advertisers can bring multiple data sources to bear on the valuation problem. This would include the targeting attributes that the publisher lists about its own impressions, data attributes from third-party data providers like BlueKai and others, and most importantly, proprietary data that the advertiser owns about its own set of customers. Based on all these different targeting attributes, the buyer can assign various business rules that align the campaign goals against potential impressions, and the bids can be set against all the various providers of inventory.
The DSP then will begin bidding across the sell-side platforms, exchanges, and any publishers that directly support real-time bidding, and will automatically optimize the bids based on success and results. The result can be as simple as reaching 100,000 people that fit some specific criteria -- or it could optimize across CPC or CPA. Real-time bidding is vastly superior to other mechanisms when it comes to ensuring that the advertiser gets the best ROI. But there are some issues.
I've heard from many of the DSPs that they are running out of real-time biddable inventory, meaning that their CPMs are rising because their supply is constrained. This might sound funny to those who fondly quote that there is unlimited supply of display inventory -- but consider that there are short- and long-term factors driving this imbalance. In the short term, the sources for this type of inventory are still somewhat limited; even with the explosive growth we're seeing in this category, there are not enough impressions available to satisfy demand. DSPs can still participate in non-real-time auctions in order to supplement impressions, but they lose the extra value they bring to the table when they can examine the impression before bidding.
Long term, there will be lots of impressions being made available. (In fact, I predict that most impressions will ultimately be made available in real-time.) But this real-time bidding world is all based on audience targeting -- and the same users that Whole Foods wants to reach are also highly valuable to Best Buy and The Home Depot. This means that those impressions driven by highly desirable audiences will be a small percentage of the total number. But note: Although from a percentage perspective we're talking small numbers, from a volume perspective that could still represent massive amounts of high bid-density inventory. Paid search impressions are a tiny fraction of display impressions today, yet drive half the revenue in online advertising. This could change significantly if we can drive enough bid density on a small fraction of display inventory that represents valuable audiences.
I have heard some premium publisher folks state concerns that there could be issues with real-time bidding on display inventory due to asymmetric bidding and low bid density. Consider the following example that illustrates how low bid density (leading to asymmetric bids) could be a problem in the future as more impressions become available for real-time bidding. I'll make it unrealistically simple to illustrate the issue:
An impression shows up for bid. It has the following attributes:
- 34 years old
- Greater than $150,000 income
- Chicago DMA
- New parent
- Auto shopper
- Jewelry shopper
- Health club member
- Impression is 300x250 pixels
- Site category is entertainment
Four advertisers participate in the auction:
Advertiser 1: Pampers -- knows nothing extra
Advertiser 2: Ford -- knows user owns a BMW and has been shopping for Land Rovers through proprietary data deals
Advertiser 3: Zales -- has existing customer data that shows this is an inactive customer, a high spender in past who bought an engagement ring three years ago
Advertiser 4: An independent Chicago diaper service -- knows nothing extra
The bidding follows like this:
Pampers bids $1 CPM.
Ford bids $5 CPM -- it knows it has a low likelihood of converting this profile, so it doesn't bid very high.
Zales bids $40 CPM -- it knows that this customer bought his engagement ring at Zales three years ago, and given the new parent status, he is likely to be open to buying an expensive Mother's Day present.
The Chicago diaper service bids $10 CPM based on simple CPA optimization.
Because this is a second price auction, Zales will win, but only pay $10 CPM for the impression. In this simple example, that might not seem too bad. But in reality, it should be possible for the publisher to predict that this impression, based on past bids on similar impressions, would sell for much higher than $5 CPM. So the publisher has not gotten the maximum yield it could have gotten based on the auction it had in play.
In the future, I predict that publishers will make use of yield optimization technology to fix this problem. The publisher should be setting a floor price on a per-impression basis based on its prediction of value to the advertisers in the marketplace. The publisher probably could have comfortably set a floor price that would have given it a higher yield (e.g., set the price at $12 or even $20 CPM based on historical trends for this type of impression and the current bidders in the auction). But this is a very hard technology problem to solve.
In paid search, we've seen high bid density drive very high CPMs on highly desirable keywords within the auction. And where the bid density is lower, we've frequently seen lower CPMs. Essentially, bid density refers to how many participants within an auction are bidding over the same item. In paid search, overall this hasn't been a problem -- mostly because there are "single digit" millions of commercially viable keywords, and about half a million advertisers competing over them. This leads to pretty good distribution, with some keywords getting lots of competition, and some getting very little -- and overall the average yield being very high for the search engine. It's a supply and demand problem for the most part.
But in online display advertising, there are trillions of display impressions a month with fewer than 10,000 advertisers (at least, in the world we live in today), with most dollars being spent in the U.S. coming from fewer than 3,000 advertisers. Further, the role of agencies could significantly change under this new set of mechanisms. There's no reason that an agency using a DSP couldn't withhold bids from its stable of advertisers so that only the top bid available for any advertiser for each impression would be placed. From a bid density perspective, this could be damaging without the kind of yield optimization I mentioned above and the creation of competition between multiple advertisers that normally wouldn't have competed in the past. But there are still things that could drive lower bid density and lower publisher yield.
For instance: In an extreme world, each agency holding company could have its own DSP, and each of these would offer only one bid per impression as it reviewed the available targeting parameters and determined -- based on each advertiser's business rules -- which of their campaigns would have the highest bid. In other words, each DSP could run an internal auction prior to placing a bid in the publisher-facing system. That would reduce the density of the auction on the publisher side significantly, causing the publisher to reduce yield. But it does require significant process change from how things are done today.
In the end, I think publishers would be foolish to worry too much here. It's likely that their highest value impressions are going to go way up in yield, even if they see a drop on the rest of their impressions. And at the least, those two things should make up for each other. At the best, this could drive average yield higher in online display than we've ever seen before.