United States v. Google/Findings of Fact/Section 5F
- F. Text Ads Auctions (Also Greatly Simplified)[1]
238. Advertisers do not purchase ads on Google in the same way they do in traditional media, like newspapers (e.g., the cost of a half-page ad) or television (e.g., the cost of a 30-second ad during the Super Bowl). Instead, on Google, advertisers compete with one another through an auction to make an ad purchase. Id. at 463:14-16 (Varian). These auctions occur in a split second, between the time a user enters a query and when the SERP is displayed. Google designs the auction and controls underlying inputs that can affect the ultimate price generated by the auction. Id. at 1197:25–1198:4, 1205:12-18 (Dischler); UPX509 at 869 (“We also directly affect pricing through tunings of our auction mechanisms[.]”). Google runs billions of search ads auctions each day. Tr. at 1198:24–1199:5 (Dischler).
239. The auction determines the ads displayed and the order in which they appear on the SERP. Id. at 1198:5-17 (Dischler). An advertiser whose text ad appears on a SERP only pays Google if a user clicks on the ad. FOF ¶ 186. A text ad is priced on “cost per click” (CPC) basis. Id. The price of a text ad “is determined based on the results of the auction, and the maximum cost per click is specified by the advertiser.” Tr. at 1352:13-17 (Dischler). Google sets a “reserve price” for text ads, or a minimum price below which it will not sell the ad. Id. at 463:20-25 (Varian); id. at 1204:15–1205:3 (Dischler).
240. Google’s text ads auction is a classic second-price auction, with modifications. A second-price auction is one where multiple bidders enter the auction, and the winner, instead of paying the price of their highest bid, pays one cent above the first runner-up. Id. at 1200:2-21 (Dischler). This makes the “second price,” or the runner-up’s bid, very important. Id. at 1200:2225 (Dischler). Google runs a second-price auction because it views it as more advertiser-friendly. Id. at 4263:12-16 (Juda). It is also more efficient for Google, because when the final price is determined by something other than the top bid, advertisers will not “be constantly trying to move their bids up or down to see if they can get the same outcome for less money,” which is burdensome for both advertisers and Google’s advertising system (which is responsible for “consuming all these changing bids at all times and processing them”). Id. at 4264:1-14 (Juda).
241. An auction winner is not determined solely based its bid. The auction also relies on certain qualitative metrics, including the quality of the ad and the advertiser’s website. At a high level, the auction captures both the bid and the qualitative factors in the following formula:
LTV = bid x pCTR – β
UPX8 at 054.
242. In this formula, “bid” represents the advertiser’s chosen bid; “pCTR,” or predicted click-through rate, is a proxy for the ad quality; and “beta” refers to blindness, which tries to approximate future engagement with ads. Id.; UPX37 at 200, 202–03; UPX442 at 868. The pCTR is a score between 0 and 1: “[I]f a predicted click-through rate of 0.20 was used in a running shoes query, that would imply that the system thinks there’s a one-out-of-five chance that a user is going to click on the ad, or a 20 percent chance.” Tr. at 4281:1-4 (Juda). The formula’s result is an “LTV” score, which refers to the “long-term” value of the ad. UPX889 at 772–73. The higher the LTV score, the more likely the ad will win an auction. Id. at 772.
- 1. Pricing Knobs
243. Google can affect the final price paid for an ad through so-called “pricing knobs” or “pricing mechanisms.” Id. at 779, 783. Google has used three primary pricing knobs to influence prices: (1) squashing, (2) format pricing, and (3) randomized generalized second-price auction. Google has referred to these levers as “intentional pricing.” UPX509 at 869.
244. Squashing premiered in a launch that Google code-named “Butternut Squash.” See generally UPX442. Squashing artificially raises the pCTR of the runner-up, thereby inflating its overall LTV score. UPX889 at 784. This increases the likelihood that the runner-up takes the top spot (even if its bid is not the highest). See id. at 784–86; Tr. at 1221:17–1222:10 (Dischler) (squashing tries “to prevent runaway winners and to create a chance for smaller advertisers to participate in the auction”). But squashing also “[e]ffectively simulates auction pressure” by making the runner-up more competitive, thereby creating upward pricing pressure on the top-rated bidder. That top bidder must pay more to win the auction so as to offset the runner-up’s artificially increased LTV score. UPX889 at 784; Tr. at 1386:6-9, 1383:19-21 (Dischler); id. at 4281:17–4283:2 (Juda). As a result, on average, the winner of an auction subject to squashing pays more than they would have absent squashing. See Tr. at 1222:3-10 (Dischler); id. at 8857:2-13 (Israel).
245. Format pricing is Google’s practice of charging advertisers for “formats,” or additional text and links that appear on general search text ads. Id. at 4254:3-8 (Juda) (discussing DXD11 at 5, 8). A formatted text ad is illustrated below.
A non-free image has been removed from this page.
DXD11 at 5 (links entitled “Find cars near you,” “How it works,” and “Getting started”). Formats allow an advertiser to create a customized and complex ad copy that provides the consumer with more information than an ordinary text ad. When first implemented, formats came at no extra cost to advertisers. See UPX430 at 580. But in 2017, Google adjusted the auction to impose price increases for formatted ads, after it determined that “strongly increased format prices” resulted in long-term revenue gains. UPX729 at 979; FOF ¶ 250 (discussing the Gamma Yellow experiment).
246. In 2019, Google developed a randomized generalized second-price auction, or rGSP, another ad launch that affected pricing. Tr. at 1222:11-17 (Dischler). Put simply, rGSP occasionally randomly switches the LTV scores of the two top auction entrants, thereby allowing the runner-up to win the auction despite its originally lower LTV score. Id. at 1222:18–1223:7 (Dischler); UPX1045 at 422; UPX512 at .009–.010. Much like squashing, rGSP artificially enhances the runner-up’s score, creating more competitive auctions and driving up final prices. UPX45 at 840 (“Ads pay a higher price to win with certainty, which increases revenue.”); Tr. at 4177:20-25 (Juda) (one way that advertisers can avoid being swapped is to increase their bid to counteract the other LTV score impacts). rGSP replaced format pricing because it was even more effective at driving revenue. See UPX512 at .002. Advertisers cannot opt out of rGSP. Tr. at 4302:9–4305:5 (Juda).
- 2. Increasing Text Ads Prices
247. Many of Google’s ad innovations seek to deliver additional value to advertisers and users. See UPX430 at 577; UPX45 at 838–39. “[A]nother important objective is the revenue that the platform (Google) makes.” UPX45 at 839.
248. Google strategically has used pricing knobs to raise text ads prices. Google’s “intentional pricing launches,” or “intentional exploration,” arose from the concern that it was not capturing in its pricing the full value of the ad to the advertiser. In other words, Google believed that it could increase ad prices because its pricing was below what advertisers would be willing to pay for an ad.
249. That intention is perhaps best captured in a January 2018 strategy document titled “How should AQ think about Pricing,” which drew on lessons from past pricing experiments and outlined possible future pricing strategies. UPX509 at 869. The record observed, “[w]e know there is still significant upside left in the different auction pricing knobs . . . but we’ve only dared capture[] a small fraction.” Id. It then asked: “Should we stop working on pricing exploration despite our belief we’re leaving money on the table?” Id.; see also UPX737 at 462 (“[T]he value created . . . was left underpriced,” meaning “that the cost of incremental clicks did not rise along with volume following the original click cost curve.”); UPX430 at 578 (“[T]here is a lot of opportunity to increase prices for search ads.”).
250. Google had learned from earlier ad experiments that small but substantial price
increases would generate sustained long-term profits. For example, a study conducted in 2017 termed “Gamma Yellow” sought to evaluate the long-term effects of increased format prices. See UPX729 at 979. The experiment exposed 15% of advertisers to “strongly increased format prices” for six weeks. Id. Google found that “50% of the initial revenue gains stuck” and “found no evidence of notable format opt-out behaviour.” Id.
251. In 2017, Google began testing a launch called Momiji. See generally UPX36. Momiji sought to determine how much Google could raise prices through format pricing. See UPX456 at 274–75; UPX36 at 063, 065–67. Google admitted that it had “no way to say what formats should cost,” but it knew that format pricing was the “best knob to engender large price increases.” UPX507 at .026. Because it had “no principle to say what the cost should be,” Google decided to “follow [its] long term revenue focus.” UPX506 at .005 (“So, we follow our long term revenue focus. We put a reasonable price to Top-1 extra clicks and see if advertisers are willing to pay it (if it sticks in an AE). Try to bring the Top-1 headroom down closer to the other position headroom.”); see also UPX456 at 274 (“We are making this tuning in order to better share in the value that AdWords and formats create, and to raise text ad prices on Google.com.”). Acknowledging that it “shouldn’t launch” if it thought it would “see large scale format opt out,” UPX506 at .008, Google nevertheless pushed significant format price increases because its experiments had revealed that advertisers would not drop out in significant numbers, Tr. at 1274:21–1275:3 (Dischler) (Momiji led to an increase in search ads revenue); see also UPX36 at 064, 069 (describing Momiji format pricing increases: “We’ve launched things at 15% and heard nothing” and “[w]e don’t see mass opt-out of anything”).
252. Similar studies showed that Google could raise prices using squashing without losing advertisers. In a 2017 study code-named “Kabocha,” Google determined that squashing was “long term revenue positive[.]” UPX745 at 085. The study showed that the “stickage factor” after price increases “was also [] roughly 50%,” meaning Google “expected 50% of gains to stick post advertiser response to the changes introduced[.]” UPX737 at 462.
253. Still, as reflected in the January 2018 strategy document, Google understood that “at any given point in time[,] there is some price or ROI ceiling above which” advertisers may abandon advertising on Google. UPX509 at 869–70. To ensure profits while remaining under the “ceiling,” Google outlined four paths, two of which involved no “pricing exploration” (thereby leaving “money on the table”) and two of which would continue “price exploration.” Id. at 871–72. Google appears to have selected “Path 3,” which it termed “Control the walk.” Id. at 872. This is the “scenario under which [Google] believe[d] the ceilings are still high and [it] want[ed] to maximize [long-term] revenue.” Id. (emphasis added). “This sharing of value implies getting closer to these ceilings without passing them, which we need to do in a controlled pricing environment.” Id. In other words, Google believed that it could raise prices using pricing knobs without losing advertisers—since “ceilings are still high”—thereby growing its revenues. Google proposed that price changes could be made through “[i]ncidental launches throughout the year,” and “[p]rice adjustments to the new state of the world would be done once or twice a year through dedicated pricing exploration using existing . . . and [h]olistic . . . tools.” Id.
254. Later launches and studies show that is precisely what Google did. UPX745 at 085 (AION six-month advertiser experiment, from early 2018, demonstrating that Google “can confidently increase format prices” because “there is still large headroom in format pricing”); UPX 737 at 462 (stating that AION’s “[s]pend response trends to the 15% change have stabilized at roughly half the initial gains, confirming our belief that there is still room for price tuning”); id. at 461–64, 476 (Potiron study, from June 2018, showing that “[f]ine grained squashing” showed the same 50% “stickage” in the long term). Increasing prices through format pricing plainly was a success. When it was replaced by rGSP, format pricing had risen to make up about 20% of Google’s text ads revenue, measured per thousand queries (also known as revenue per mille, or RPM). See UPX512 at .002 (format pricing comprised about 20% of Google’s RPM).
255. The launch of rGSP in 2019 was equally successful. Google’s pre-launch experiments indicated that rGSP would increase CPCs for top slot ads on non-navigational queries by 5.91% on PCs and tablets and 4.85% on mobile phones with a long-term “stickage factor” of 40–50%. UPX457 at 258–60. Experiments showed that a 5.74% revenue gain persisted two months after launch. UPX45 at 838; see also, e.g., UPX745 at 085–86 (new launch known as “Stateful Pricing” demonstrated “over $6 billion in short term incremental annual revenue in headroom”).
256. In February 2020, Google reported that the rGSP “tuning point,” or increased bid, was about 3.7. UPX466 at 939. This meant that in order for the top bidder to keep its position, it would need to bid 370% more than the runner-up to account for the swapped LTV score. Id.; Tr. at 4178:8-14 (Juda); see id. at 4177:20-25 (Juda) (one way an advertiser may avoid swapping is by increasing its bid). If that bidder was successful, it would ultimately pay significantly more than it otherwise would have for the same ad placement.
257. Google’s records make clear that growing its revenue was a principal goal in launching these price tunings. See, e.g., UPX51 at 228 (“Main goal: Long-term Revenue”); UPX442 at 868 (Google will use its launch “to recover lost revenue from launches which create value for our users and advertisers, but reduce revenue for Google”) (squashing); id. (Google “wants to continue launching such advertiser value creating launches, but needs a mechanism to help Google share in the value that [the] launches create”); UPX507 at .004 (“Prices could be higher, and we think we would keep the money,” because “[r]evenue gain from higher prices > revenue loss from response” by advertisers) (format pricing); id. at .010 (describing philosophy as to “[g]et the highest RPM point possible”); id. at .027 (ranking format pricing, squashing, and reserves by “effectiveness,” measured as increased RPM); UPX430 at 577 (Google adjusts “the parameters of the auction function in order to improve Long Term Revenue. . . . This work has resulted in products which add several billions of dollars in incremental revenue annually.”); UPX45 at 837 (rGSP solves the “difficult problem” and “major priority” of “increasing revenue in auctions with low competition[.]”).
258. In fact, Google used ad launches to meet revenue goals or make up for perceived deficits in its ad revenue growth. See, e.g., UPX745 at 085 (projecting “+4% RPM from standalone pricing launches” and expecting additional billions in “incremental annual revenue” from format pricing and squashing); UPX456 at 298 (predicting at +1.3% revenue increase). As Dr. Adam Juda, Google’s Vice President of Project Management, testified, a positive 20% increase in revenue “was an annual objective that we would try to get to over the course of an entire year.” Tr. at 4140:1-20 (Juda); see id. at 7549:6-9 (Raghavan) (discussing UPX342 at 824) (same).
259. And Google met that objective year after year. As the below chart shows, Google has enjoyed unusually consistent revenue growth from 2010 to 2018 that hovered at or above the 20% expectation.
UPX342 at 824.
260. If Google grew concerned about meeting its revenue targets, it called for a “Code Yellow effort,” where its “top priority” would be to “deliver [] revenue launches” through intentional pricing. UPX738 at 406; see UPX733 at 203–04 (describing the Sugarshack format pricing launch, which was used to meet Google’s revenue targets in response to a Code Yellow); UPX514 at 386 (describing ad launches implemented to meet Code Yellow revenue goals).
261. Google’s pricing decisions also reflected an understanding that increasing its revenue in the ways discussed might occasionally come at a cost (or no improvement) to advertisers. See UPX734 at 509 (“cleverer . . . auction pricing” comes “at a cost to advertisers”); UPX507 at .015 (“Sales struggles to explain these [price increases] in terms of user/advertiser value[.]”); UPX889 at 780 (auction pricing mechanisms are “[n]ot designed to increase clicks”); UPX36 at 065 (“[C]urrent system has issues. We’re acknowledging the current CPM space is giving them different prices at the same value.”).
262. For instance, Google claimed that the primary motivation for implementing squashing was to help smaller advertisers, but that is not borne out by the record. Tr. at 1386:1019 (Dischler) (“The primary reason that we implemented squashing was to prevent certain winner-takes-all dynamics in the auction. What we were finding is that there were a few large advertisers that were kind of winning every auction in a particular category, and we weren’t sure actually whether that was a good user experience. It was becoming much harder for the runnerup to break through and show up in the top position.”). In fact, after squashing, Google displayed the same ads on about 95% of queries measured by impressions and clicks, generating 88% of its revenue from queries returning the same ads in the top placement. UPX442 at 872. In other words, the overwhelming majority of revenues resulted from the same placements before and after squashing. Moreover, Google measured success not based on improved ranking for smaller advertisers, but by whether a “squashed” auction produced positive revenues for Google. In one record, Google described squashing as “desirable” when CPCs increased, and “undesirable” when they did not due to “reranking.” UPX737 at 464. Because squashing produced desirable results 60% of the time, Google believed that “coarse squashing provide[d] overall positive metrics” but was “suboptimal due to these mixed effects.” Id. Google proposed to further refine squashing to optimize revenues. Id. at 464–65.
263. When it made pricing changes, Google took care to avoid blowback from advertisers. For instance, records show that Google had concerns about the impact of transparency on their efforts to increase prices. See UPX507 at .015 (“Worry that if we tell advertisers they will be impacted, they will attempt to game us and convince us to abandon the experiment. . . . But, if we don’t tell them, they will react more naturally (how they’d react if they believed they couldn’t influence our decision at all).”); UPX519 at .003 (“A sudden step function might create adverse reaction.”).
264. Google therefore endeavored to raise prices incrementally, so that advertisers would view price increases as within the ordinary price fluctuations, or “noise,” generated by the auctions. See, e.g., UPX507 at .023 (describing a 10% CPC increase as “safe” because it is “within usual WoW noise”); UPX519 at .003 (acknowledging that advertisers would notice a 15% price increase, but “this change is to [be] put in perspective with CPC noise,” that is, “50% of advertisers seeing 10%+ WoW CPC changes”); id. (comment stating that 15% is “probably an acceptable level of change (from a perception point of view) because these are magnitudes of fluctuations they are used to see[ing]”).
265. With respect to format pricing, one Google document states: “A progressive ramp up leaves time to internalize prices and adjust bids appropriately[.]” UPX519 at .003; UPX509 at 870 (stating that “[i]ncremental launches and monitoring should help us manage” the risk that price increases would lead advertisers to “lower[] their bids or modify[] other settings . . . to get back to a given ROI, leading to less revenue for Google than the initial impact hinted to”). Similarly, in 2020, Google raised prices on navigational queries using multiple knobs and recognized that it was “[o]bviously a very large change that we don’t intend to roll out at once,” instead planning a “[s]low 18 months rollout” to “[l]eave[] time for advertiser[s] to respond rationally[.]” UPX503 at 034; id. at 038 (“A slow roll ensures we don’t shock the system, gives time for advertisers to respond and us to monitor changes and stop early if needed.”); see also, e.g., UPX505 at 312 (prior to implementing squashing, concluding that “[a]dvertisers should perceive AdWords as a consistent system, and not be subject to constant large impacts due to Google changes,” in part to “improve[] advertiser stickiness”); UPX506 at .018 (Momiji slide deck: “Unlikely that advertisers will notice by themselves and respond. However, a bad press cycle could put us in jeopardy.”).
266. Google’s incremental pricing approach was successful. In 2018 and 2019, Google conducted ROI Perception Interviews, which raised no red flags about advertisers’ attitudes as to ad spending on Google. See generally DX187; DX119. While advertisers could tell that prices were increasing, they did not understand those changes to be Google’s fault. Google’s studies revealed that advertisers facing CPC changes “dominantly attribute[d] these shifts to themselves, competition[,] and seasonality (85%)—not Google.” UPX1054 at 061; see also UPX737 at 464 (“They often attribute these changes to things in the world or what they’ve done, not just things happening on the backend[.]”).
267. When it made these pricing changes, Google did not consider its rivals’ text ads pricing. See UPX509 at 959 (Dr. Raghavan querying why “all of the discussion on advertisers’ reactions to [Google’s] pricing changes seem to presume that this is a 2-person game between the advertiser and [G]oogle,” even though it is “really 3 players—the advertisers, [Google], and [its] competitors”); id. (noting that “the discussion seems insensitive to where else the advertiser could obtain traffic of similar quality and price”).
- 3. Limiting Advertiser Control
268. Google also depreciated the quality of its text ads product in two primary ways: by reducing the information available to advertisers in Search Query Reports and by loosening keyword matches to create more crowded and higher price-generating auctions.
- a. Search Query Reports
269. Google began offering Search Query Reports (SQRs) in 2007 to help advertisers determine whether to add new affirmative or negative keywords to their lists. UPX526 at 538; Tr. at 1481:16-20 (Dischler) (“They use it in order to measure their advertiser effectiveness, or they could use it in order to improve the range of keywords that they use in order to be able to target users that are looking for their products or services.”). Google was aware that SQRs were “widely used by advertisers of all segments.” UPX526 at 539, 556.
270. Prior to 2020, SQRs included all queries that resulted in an ad click, even if there was only a single click (i.e., the “one-click threshold”). See generally id. Ostensibly out of privacy concerns, Google removed the one-click threshold. Id. at 543. It did so notwithstanding “substantial” projected data loss for advertisers and knowing that specific major advertisers, like Expedia and Booking.com, had stated they would be harmed. Id. at 545, 549.
271. Google’s own records show that the privacy rationale was suspect. See id. at 525 (email from Dr. Juda questioning whether the proposed trimming of the SQR report “could or should be turned into a [privacy-focused thing] without a lot of thought”); id. at 531 (“While a query can contain sensitive information, I have the ability to type anybody’s SSN into my search box. Therefore, queries are not PII, even if I am the only person ever to search for your SSN.”); id. (opining that “even when we do share keywords which are identical to the query and contain sensitive information, I would argue our documentation is accurate”); id. at 541 (unnamed commentor stating “queries aren’t PII”). Some advertisers, as well as U.S. Plaintiffs’ expert Dr. Kinshuk Jerath, also view Google’s privacy-related justifications with skepticism. Tr. at 3850:5-7 (Lowcock) (“[I]t would be reasonable to continue to share that sort of information with us without breaching privacy regulations.”); id. at 5473:13-25 (Jerath) (“[T]his is not a valid reason because the search query reports were never using user level data.”). Still, Google decided in the fall of 2020 that all queries must receive 50 cookied impressions daily to appear on an SQR. See UPX532 at 566 (“This decision is rooted in Google’s treating search query data as personal data for this use-case, even though Google has reasonable arguments such data [(i.e., queries)] may not be personal data in many instances.”).
272. The less fulsome SQRs negatively impacted advertisers, who already have limited insight into how Google’s auctions work. See, e.g., UPX519 at .016 (advertisers “would like to see . . . more transparency in the definition of quality”); Tr. at 3850:16-18 (Lowcock) (“[W]e know what price we paid. We have no true visibility in the way that the price is determined and how the auction is conducted.”); Alberts Dep Tr. at 213:21–214:6 (“[I]t does limit some of the visibility in some of the terms that are triggering keywords that we would not like to match to.”); Tr. at 5174:16-20 (Booth) (same); see also id. at 5468:6-21 (Jerath) (additional examples). For instance, JPMorgan Chase estimated that prior to the change, about 5% of the keywords were not visible on SQRs, but afterwards the number rose to 20%. Tr. at 4866:13–4868:10 (Lim) (“It just gave my team less information to work with.”).
273. Google did not inform advertisers how the threshold had changed. UPX532 at 568 (internal informational Q&A for press inquiries advised not to reveal the threshold for making the SQR “in keeping with our privacy and security policies”); Tr. at 5222:2-19 (Booth); Alberts Dep. Tr. at 166:17-25. And because advertisers no longer received a report of every query that involved an ad click, advertisers purchased ads on certain queries generating fewer than 50 cookied impressions. See Tr. at 5469:18–5471:12 (Jerath) (“They were buying certain queries but they were not being told . . . which queries they’re buying,” as if you purchased “a product in a supermarket but they don’t tell you what you actually bought.”); id. at 5471:10-12 (Jerath) (“This is data that you’re actually buying. This is indeed where your spend is going. You should be entitled to know that at least this is where I spent my money.”).
274. Advertisers not only identify the keywords that may trigger participation in an auction, they also can identify so-called “negative keywords,” which are keywords that an advertiser selects so as to avoid entry into an auction. Alberts Dep. Tr. at 214:10-21; Tr. at 400:37 (Varian) (“[I]t’s the advertiser that provides the keywords. Google is seeing if those keywords match the query, and then it’s determining that. So it’s really the advertisers’ choice of keywords that are determining whether it serves an ad.”). Without the single-click information, Google thus not only constrained advertisers’ ability to withdraw keywords but also to identify negative keywords to remove themselves from undesirable ad auctions. See Tr. at 5472:11-24 (Jerath).
- b. Keyword Matching
275. Google also reduced advertisers’ ability to remove themselves from certain ad auctions by expanding its “keyword matching” functionality. “[T]he typical way that advertisers interact with search advertising is using keywords, which is literally the advertiser [] guessing what the users might be querying, which is very complex. And so doing that for millions of products is sort of an undue burden on advertisers so [Google] came up with an automated system where [it] do[es] more of the matching.” Id. at 1353:21–1354:2 (Dischler).
276. One way Google does this is through “semantic matching,” which tries to “understand[] the meaning of [key]words and replac[e] those with analogous words so that things that mean the same thing in a particular language are treated the same way.” Id. at 1363:12-16 (Dischler). The chart below depicts how semantic matching works for the keyword “kids clothing.”
| New matches for keyword*: +kids +clothing | ||
| kids → children | kids clothing → kidswear | clothing → apparel/outfit |
| clothing for young children | nikolai kidswear | creative apparel for kids |
| children's clothing in singapore | tj maxx kidswear | kids outfits |
| kids clothing canada | kids winter wear for girls | kids apparel in citywalk |
| best children's clothing brands | sean jean kids wear | |
| childrens beach clothes | kids wear online | |
| newborn children's clothing | kidswear outlet | |
| Note: Table is a sample of matches, not exhaustive. * Includes both S&R, SNE and SemPhrase & SemBMM matches (all are new). | ||
DX18 at 721. Another example is correcting misspellings. See Tr. at 1365:15-22 (Dischler); see also id. at 3848:17-20 (Lowcock) (describing “products like keyword matching and broad match modifier, which means the algorithm of a machine that the search engine is running can look for synonyms or understand what might be associated”).
277. Google has changed its keyword matching over time, beginning in 2012. Id. at 4283:13–4284:15 (Juda). The narrowest category, “expanded match,” initially included only the keyword itself or grammatical variations (e.g., plurals) but today includes misspellings. UPX8055 at .001–.002; Tr. at 5477:15–5478:1 (Jerath) (discussing UPXD103 at 40). When Google began including misspellings as part of “expanded match,” about 25% of advertisers (by ad revenue) opted out of the new feature, including many of Google’s largest advertisers, like Amazon. UPX518 at 573. Nevertheless, Google removed the opt-out option in 2014, UPX8049 at .003; Tr. at 1478:12-14 (Dischler); id. at 4298:6-16 (Juda), despite recognizing that this move would “[r]emove[] control from advertisers,” UPX518 at 572. Thereafter, Google continued to expand the keyword match types. See UPX31 at 471. There are presently three types: broad match, phrase match, and exact match. UPX8023 at .001.
A non-free image has been removed from this page.
UPX31 at 471.
278. Because broader matching enters more advertisers into an auction, it leads to thicker auctions (i.e., more auction participants), which creates upward pricing pressure. Tr. at 1477:1824 (Dischler); id. at 4298:22–4299:1 (Juda). As advertisers cannot opt out of matching, the only way to ensure that a certain query does not trigger an ad is to provide a negative keyword. Id. at 4297:23–4298:3 (Juda). But identifying negative keywords is a far more cumbersome way for advertisers to avoid undesirable auctions, a challenge made even more difficult with less information from SQR reports. See id. at 5472:11-24 (Jerath).
- ↑ At trial, Plaintiffs repeatedly confronted Google’s ad executives with company records containing their own statements, as well as the statements of their colleagues, regarding Google’s text ads auctions. In many instances, the witness professed to lack an understanding of the record or sought to contextualize it in highly technical ways. In making these Findings of Fact, the court gives greater weight to the contemporaneous statements contained in the company’s internal records, than later trial testimony in which Google employees declined to ratify those statements.