I guess the claim was that they manipulated where bids were placed based on who had the highest bid. So that if a non Google advertiser was going to win the auction, they removed the second highest bid if it was a Google advertiser s.t. the price paid for the slot dropped (second price auction). So in effect, they maintained a higher payout for pubs when Google programmatic bidders won by dropping the second price floor on non Google-programmatic winners. Thus making more money AND converting people towards programmatic.
I believe their counter argument is that the data they had actually just made programmatic better from the start by predicting these outcomes, no fraud needed
[1] https://www.wsj.com/articles/googles-secret-project-bernanke...
This is a second price auction that we speak of. Every ad company is polled to send no more than two bids. Naturally those would be the two highest bids from thousands that they handle. So let's say that Google first and second bids are G1 and G2, Facebook F1, and F2, some other company C1 and C2 and so on. So if those bids are sorted in descending order it may look like, for example,
A1 > B1 > A2 > C1 > B2 > C2 > ...
In this example, the ad belonging to A1 gets served, the price is the value of bid B1.That's why it is called "second price" auction. If it was a first price auction, the A1 advertiser would pay A1 bid price, and could save some money by bidding slightly lower, but still more than B1. Neither of these companies know up-front what the other bid will be, but with a second price auction there is no need for it, the A1 advertiser pays the amount of B1 bid.
Same goes for bids G1 and G2, they don't know what other bids will be. It could end up like this:
G1 > G2 > C1 > ...
From time to time Google overhauls their AI. Let's say for simplicity that the two highest bids from Google are the same ads, the new bid for G1 is G'1 and the new bid for G2 is G'2. Though it could change who wins the auction, for instance if it ends up like this G'2 > G'1 > C1 > ...
or
C1 > G'1 > G'2 > ...Or it could change only what the price is, for example if it ends up like this, the price will lower from G2 to C1
G'1 > C1 > G'2 > ...
Or if there is some minimum bid value Z required by publisher and it it would be G'1 > C1 > Z > G'2, it would be only G'1 that is passed to the auction. "Google allegedly dropped the second-highest bids from publishers' auctions".The "inflate only the bids belonging advertisers who used the company's Google Ads" is a mixture of G'1 > G'2 > G2 > ... and G'1 > C1 > G1 > ... and that sort of stuff
There is no accumulating money in a pool, keeping it to spending later on something; there is a need for balancing the two effects above, so that the overhaul won't go overboard either way.
The case of C1 > G'1 > G1 > ... is missing from the Reuters snippet. If it is in the lawsuit, it would end up as "secretly forcing the advertisers who used competing ad network to pay more", I guess
- WSJ: "Texas alleges that Google used its access to data from publishers’ ad servers—where more than 90% of large publishers use Google to sell their digital ad space—to guide advertisers toward the price they would have to bid to secure an ad placement."
- This is bid optimization & bid shading no? Both helping buyers place a high enough winning bid. And then also offering bid shading probable. With a second price auction the winning bid clear price is 2nd price + .01 or whatever. All DSPs offer this and use the open market data?
- Next WSJ graph seems like the crux of the insider trading is not really accurate? If all major bidding platforms (dsp's) offer this feature. e.g. it doesn't " unfairly compete against rival ad-buying tools and pay publishers less on its winning bids for ad inventory." ?
- "Google allegedly dropped the second-highest bids"
- So publishers still got paid the highest bid?
- Or maybe see above about bid shading. Maybe this could lower the clearing amount paid to the publisher? E.g. $10, $7, $5 bids. Drop $7, winning price is now $5.01 ?
- "accumulated money into a pool and then spent that money to inflate only the bids belonging advertisers who used the company's Google Ads"
- I don't understand this. The WSJ article also doesn't mention what money is pooled.
> - "Google allegedly dropped the second-highest bids" >> - So publishers still got paid the highest bid?
During the relevant period and until end of 2021, Google is running a second price auction. The highest bid wins and then pays the second highest bid.
If that's the flat out only thing literally happening, I can't think of a reasonable reason that would be in the best interest of the publishers paying google to sell supply side.
it's all jumbled when you control all sides of the auction and infrastructure which is the real problem IMHO.
Maybe I can see a scenario where bid shading is conflated here a bit or at least complicates the scenario.
For instance 'drop' the 2nd bid maybe means placing a lower bid instead of 'deleting' one.
all DSPs have a good sense of what will win the auction and offer bid shading to buyers.
Just trying to think through that line of thought. For instance I buy political ads. Google knows that video inventory X usually sells for $30 CPM, but my settings have $50 CPM max bids because during the election it's hard to get scale sometimes. my DSP would bid shade lower to say $30.01 on that individual impression, and if I happen to be the 2nd price maybe that is 'drop' they are refering too?
IDK i don't think traditional news media reporting here will ever give enough information / they understand enough
You need to realize just one more thing: to ensure that there is no peeking into bids of other networks, the actual code of the auction could be executed on a remote server. But only a naive implementation would ask every ad network client to send all their ad bids to the server; on the other hand having ad networks send only the highest bid would not be enough for a second-price auction...
adx ssp sends 1 bid of $10, even though their next highest internally is $9 mopub ssp or whoever sends 2 bids of $5 and $7
So if nytimes.com is doing a header bid on their own (or even through google's ad serving product - again imho the biggest problem) they don't see the $9 and the clearing price becomes $7.01
I guess it would be a question of contracts. This is beyond what I do so I have no idea what the standard number of bids each SSP sends to a header bidding auction, obviously don't want to send google scale 1 billion lol...
- google is basically offering bid shading on both sides of the transaction. (RPO on pub side according to this tweet) - Both dynamically increasing floor to buy NyTimes.com based on the current likely bid - And for buyers increasing/decreasing bid to pay the best 2nd price auction clearing price
which imho is the main problem that should be legislated; you can't own all sides of a transaction AND the pipes. instead of trying to convince a jury of some complicated insider trading monopoly confusing mess.
"ECPC works by automatically adjusting your manual bids for clicks that seem more or less likely to lead to a sale or conversion on your website"
Of course, it is based on historical data
But you end up having a quite complicated code calculating the auction bid.
I doubt that you could code it easily based on a cursory reading of the About Enhanced CPC page.
Even more futile is guessing what specific details of that code function that project Bernanke optimized, if your only source is the lawsuit doc and newspapers relating it
https://ia601404.us.archive.org/8/items/gov.uscourts.nysd.56...
There may be more details in the 3rd amended complaint that was just filed. Has anyone found a copy online.
It would be surprising if anyone could ELI5 the mechanics of Project Bernanke in a HN thread. If they could, the state AGs would probably like to hear from them. :)
Can someone rephrase this and possibly explain like I'm five? I'm not following the mechanics...