A look inside Google's advertising auctions

On Day 16, DOJ explored whether Google could illegally raise the prices of its ads.

Google’s advertising auctions are quite different from the type of auction depicted in this photo. (Photo by Spencer Platt/Getty Images).

Before this trial started, here’s what I thought of when I heard the word “auction”: 1) a seller who wanted to sell something could put that thing up for auction; 2) potential buyers would bid for it; and 3) the thing would be sold to the highest bidder at the price of their final bid.

I’d venture to guess that most people think of a process similar to what I just described when they think about auctions. But that’s not how Google’s auctions work.

Today, the DOJ continued its “deep dive” into Google’s advertising auctions with further examination of Dr. Adam Juda, a longtime Google employee who currently works as its Vice President of Product Management in Search Ads Quality Systems.

I’m going to try to summarize what we learned from Dr. Juda about the mechanics of Google’s ad auctions, but I’ll warn you beforehand, it’s pretty technical and involves a lot of acronyms (e.g, LTV, CPC, ECPM, RGSP, etc.). There are surely details I’m missing from how exactly this process works, but I am going to do my best to flag what I thought were the most important points.

As I hinted above, Google’s advertising auctions do not always sell an ad spot to the highest bidder. Instead, Google uses a complex formula to rank bids from advertisers based on the long-term value (“LTV”) offered by the ad.

Juda explained why Google does this even though it means Google frequently turns down short-term revenue gains. Basically, if Google always just sold to the highest bidder, they could end up with lots of low-quality ads that have little relation to the actual search query the ads are showing up for. Over time, offering these kinds of ads would degrade the value of Google ads in a number of ways: users would be less likely to click on the ads, advertisers would be less likely to bid for ads that users didn’t click on, the list goes.

So instead, Google ranks advertisements based on a combination of the bid price, the predicted click-through rate (“PCTR”) because advertisers only pay Google when their ad is clicked on, and a “Beta” factor, which seeks to estimate the long-term negative effects a specific advertisement might have for specific search terms. This is all captured in this simplified version of the formula where LTV = bid price x PCTR - Beta.

This means that advertisers with relatively lower-quality ads — which is determined by another complex formula — have to offer relatively higher bid prices to win the top spot in an auction. Any advertiser that offers a bid with an LTV of less than zero will be eliminated, leaving only the bids with positive LTVs.

Still with me?

Even after the bids have been ranked by their LTVs, there is yet another process for determining which bid will actually win the auction — and how much the winning bidder will have to pay. That’s at least in part because Google uses what are called randomized generalized second-price (“RGSP) auctions.

Google uses RGSP to randomize the final winner of the auction between the top bidders, and then the top bidder pays a price of the bid equal to the next-highest bid plus one cent. This is all obviously complicated and technical, but Dr. Juda explained some intuitive reasons for using RGSP.

One reason for the randomization is so that bids with LTVs that are only marginally lower than the bid with the highest LTV will still have a chance of getting placed. And one reason for using the second price is so that advertisers don't have to worry that they are significantly over-bidding for an ad, which could lead to them constantly adjusting their bids. Instead, the price advertisers pay is determined by the price of the next highest-ranked bid. Juda described this as an “advertiser-friendly” mechanism.

So far, mostly so good. DOJ presented an internal document showing Google employees communicating about how to explain the randomization to advertisers without creating a “perception problem”, but otherwise didn’t seem to raise challenges to much of what Juda described.

Instead, DOJ focused on two main lines of attack.

The first is that advertisers don’t get told the details of how the LTV formula works or how their bid is ranked with the formula. Instead, it’s a “black box” — a phrase that one of Google’s own internal documents used to describe the formula. Juda pointed out that advertisers do receive an ad quality score of 1-10 for their bids, which is informed by the same inputs that are used in the secret LTV formula. But as the DOJ pushed back, the ad quality score itself is not actually used in the LTV formula. This wasn’t fully spelled out in the questioning, but I think DOJ wants Judge Mehta to infer from this that advertisers are forced to place higher bids than they otherwise would if they had a better idea of how their bids are ranked.

Second (and probably most importantly), DOJ’s questioning focused on Google’s power to tinker with its auctions through the hundreds of “ad launches” it implements every year, many of which have an effect on the price of ads. Juda pushed back on these questions, but DOJ tried to show that Google had the power to raise its ad prices with little to no regard for its competition

In one exchange, DOJ presented Juda with an internal memo from 2019 in which Juda outlined various reasons that his team might not be able to hit Google’s annual revenue goal for its ad launch experiments. After walking Juda through the memo, the DOJ lawyer asked Juda if any part of the memo mentioned competitors as a reason Google might not be able to hit its revenue goal. Juda asked to review the rest of the memo before responding: “There’s implicit references here to competitors but I agree it’s not explicit.”

When Google cross-examined Juda, Google’s lawyer asked him specifically whether he agreed with this allegation from the DOJ’s complaint: “By suppressing competition, Google has more power to manipulate the quantity of ad inventory and auction dynamics in ways that allow it to charge advertisers more than it could in a competitive market.” Juda said he didn’t agree with the allegation.

But right after that response, Judge Mehta asked Juda a question that provided what may have been the best moment of the day for the DOJ. Judge Mehta asked Juda if there was any variable in Google’s ad auction algorithm that takes into consideration the price of ads on other digital platforms. I couldn’t hear exactly what Juda replied, but the answer wasn’t yes — I think he said something to the effect of no, but other variables do so indirectly. (I will try to update this article with the exact quote of what Juda said once I see the transcript.)

DOJ then followed up on this point in its redirect examination by asking Juda whether he was aware of anyone at Google performing analyses of Bing’s auction model or pricing. Juda said he wasn’t.

Other tidbits from today

Good for you if you made it this far. Today we also heard some testimony from Ryan Krueger, a Google product manager, as part of the State’s case against Google’s delayed integration of Microsoft features into its advertising platform called SA360. One interesting note is that he was only the second witness so far in the trial to be asked about his use of history-off chats after being placed on a litigation hold.

At the very end of the day, Judge Mehta announced he would be unsealing large portions of the closed-session testimony of DuckDuckGo CEO Gabriel Weinberg and Apple’s Senior Vice President of Machine Learning and AI Strategy John Giannandrea. Judge Mehta said he did a line-by-line review of the sealed testimony alongside the parties’ proposed redactions and applied the Hubbard factors in determining what to unseal. He also said he would eventually be unsealing portions of testimony from other witnesses like Apple executive Eddy Cue.

Judge Mehta’s unsealing of previously sealed testimony is just another example of the dramatic shift towards a much more open trial we have seen over the last couple of weeks. The trial has not had a single closed session this week as Judge Mehta has prodded the attorneys to do as much of their questioning in open court as possible.

I am still waiting to receive transcripts of the unsealed testimony, but I’ll try to share any important takeaways from the unsealed testimony once I’ve been able to review it. That’s all I have for today.