,
19 tweets,
4 min read
Read on Twitter
First, the dataset itself is massively problematic. It's 3 years old (from May 2016), based on data from a single publisher, is data from a single week, and is referenced as "millions of transactions" repeatedly as if that's a good thing, when in reality that's a tiny dataset.
3 year old data is completely irrelevant to look at compared to today's digital ads ecosystem. Safari ITP didn't even exist until June 2017 and GDPR wasn't in effect. On the flip side, programmatic demand has increased by at least 50% in the interim.
The fact that it is data from a single publisher could be massively swayed by their programmatic stack. Did they run header bidding? (Given it was 3 years ago, odds are they didn't) What demand partners did they have? What flooring and blocking rules?
And for all we know, this publisher's stack was a total mess and completely inefficient.
Data from a single week could be skewed massively by seasonality, potential trafficking issues, configuration issues and any number of other things that can't be controlled for.
And millions of impressions? We've served tens of millions of impressions in the 20 minutes it took me to write these tweets. That's a tiny dataset. Was the data filtered or cleaned in any way? What was the actual technical source of the data? The ad server? Something else?
How were direct deals controlled for here? They're not even mentioned in this article, so maybe this publisher was programmatic only? But maybe not - there's no information.
If there were direct deals, they could have eaten a significant amount of the inventory (again, the research doesn't clearly say whether the dataset is all of the publisher's inventory or a filtered subset). Plus those direct deals could have acted as price floors for Google DA.
There is a lot of talk about the "presence of a cookie". Which cookie were they looking at? The presence of a cookie that the researchers were looking for would have no bearing on the presence of cookies that could be used by Google, Criteo, TTD or hundreds of other buyers.
Nor fingerprinting or a bunch of other ways that ad tech partners could know who a user is. Particularly in May 2016.
Assuming this publisher only had AdX at the time then they had a pretty basic (and poorly performant) ad stack, so are not representative of savvy publishers or the publisher of May 2019.
twitter.com/antoniogm Puts forward a pretty good theory around second-pricing being part of the problem here , and I agree it's part of it. But I think that only holds when there is a single behavioral bid.
If you get just two higher-paying behavioral bids, it would drive the price up dramatically. Plus in two months, second-price auctions are effectively dead, again speaking to the poor dataset from 3 years ago.
And from everything I have seen within our data as well as from many other publishers shows that "cookie rich" (Chrome) impressions perform way better than "cookie poor" (Safari) impressions on the same geo/device type/slot/domain/time/etc. At least 50% and usually more.
This isn't a commentary as to whether if behavioral ads totally went away that money would flow through other channels/targeting and make publishers more by removing ad tech. Maybe that's true, but this research certainly provides no useful data to support or deny that.
Also potential statistics issues and observations from our head of data science, Patrick McCann :
table 5 estimates p(cookie | iphone + mobile) = invlogit(2.344-0.439-.565)=0.79. A 79% iphone cookie rate must be talking about presence of first party cookie, which isn't all that relevant. Also 4% of 1.14 != 0.08 and table 4 says e($|cookie)=1.02 < 1.14, so thats odd as well.
And after all of that I realized I wrote advertisers instead of publishers in my first tweet. Whoops.
