Why findings of fact are different from everything else in the field
Most of what is written about how Google ranks pages comes from three places: Google's public statements, the documentation leaked in 2024, and practitioners' experiments. Each has a weakness a litigator will recognize. Public statements are a party's description of itself. The leaked documentation is a field dictionary with no weights attached. Experiments are rarely reproducible and almost never peer reviewed.
The record in United States v. Google LLC, No. 1:20-cv-03010-APM (D.D.C.), is different in kind. It consists of findings of fact entered after a contested bench trial, grounded in sworn testimony from Google's own engineers and in Google's own internal documents admitted as exhibits. The liability opinion issued on August 5, 2024 and is reported at 747 F. Supp. 3d 1. The remedies opinion followed on September 2, 2025, and the Final Judgment was entered on December 5, 2025. Appeals are pending in the D.C. Circuit, so check their status before relying on any finding in a filing.
For a search expert, this is the most reliable public account of how Google's ranking systems used user data at the time of trial. It is also routinely quoted for more than it says, which is the reason for this guide.
What the court found about click data
The court defined click data precisely. It includes the results a user clicks, whether the user returns to the results page and how quickly, how long the user hovers over results, and how the user scrolls (FOF ¶ 88). That definition is the court's, drawn from Google's own exhibits, and it is the one to quote.
The court then described Navboost, a ranking signal that pairs queries with documents by memorizing user click data. Since 2017, Google has trained Navboost on 13 months of user data; before 2017 it used 18 months (FOF ¶ 96). In the same paragraph the court found that 13 months of Google's user data is equivalent to over 17 years of data on Bing. A related signal, Query-based Salient Terms, which identifies words that should appear prominently on relevant pages, is trained on about 13 months of user data.
The broader finding is that user data improves quality at every stage of search: crawling, indexing, retrieval, and ranking. That finding is what gave the scale of Google's data its significance in a monopolization case. For a search expert it settles a question that public guidance long left ambiguous: aggregated user interaction is an input to how Google ranks pages.
Glue, the thirteen-month query log
The remedies opinion added detail. It described Glue as essentially a super query log that collects data about a query and the user's interaction with the response, including the query text, language, location and device, the ranking information shown, and the user's clicks, hovers, and time on the result. The court found that Navboost data is an important component of Glue, and it cited an internal exhibit labeled "Glue Cache (13 months)" for the proposition that training on 13 months of data means the queries and clicks of all users worldwide.
The same opinion described RankEmbed and RankEmbedBERT, newer deep-learning ranking models, as trained on a portion of 70 days of search logs together with scores from human quality raters. And it recorded that PageRank, the link-based measure of a page's authority, remains a key quality signal and an input to Google's quality score.
Set beside each other, those windows are instructive. Google's memory of user behavior for ranking runs about 13 months. A site owner's own record of how its pages performed in Google, in Search Console, runs 16 months. Both roll forward continuously, and neither preserves itself.
Google's own description: three pillars of ranking
Two internal Google presentations admitted at trial describe ranking in Google's own words. The first, a deck titled "Life of a Click" and admitted as UPX0004, sets out what it calls the three pillars of ranking: the body of the document, meaning what the document says about itself; anchors, meaning what the web says about the document; and user interactions, meaning what users say about the document. It lists clicks, attention on a result, swipes on carousels, and entering a new query as user interactions.
The second, a ranking presentation admitted as UPX0203, states that Google does not understand documents directly and instead watches how people react to them and memorizes those responses. The court quoted the companion line: a document that gets a positive reaction is treated as good, and one that gets a negative reaction is treated as worse.
Both exhibits are public only in redacted, abridged form, and both are undated in the public versions. Quote them as exhibits, with their exhibit numbers, and do not present them as a current description of Google's systems.
Newer systems did not displace the older ones
A common argument in current disputes is that generative AI has made the older ranking machinery irrelevant. The trial record does not support that argument for the period it covers. The court found that newer large language model signals did not replace Navboost and Query-based Salient Terms, and it quoted a Google document describing Navboost as "one of the most power ranking components historically" (FOF ¶ 102, quoted as printed). It also found that generative AI had not, or at least not yet, eliminated or materially reduced the need for user data to deliver quality search results (FOF ¶ 115).
The remedies opinion adds a qualification on the other side: Google does not use click-and-query data to pre-train its base Gemini models, although it post-trains the model that adapts Gemini for AI Overviews on search data of an unspecified type. The record dates largely from 2023. It describes Google's systems at the time of trial, not necessarily today.
What the record will not support
Here the findings are routinely overstated, and an expert who overstates them hands opposing counsel a Rule 702(d) argument.
- It gives no weights. The court found that Navboost is trained on 13 months of data. It did not find how much Navboost counts in final rankings, for any query or any site.
- A training window is not an effect duration. Thirteen months describes the data a system learns from. It does not establish that a change in user behavior affects a particular site's rankings for thirteen months.
- It does not attribute any site's decline. The findings describe Google's systems in general. Whether a specific site lost rankings because of user signals, links, content, or an update is a question the record cannot answer.
- The market shares are defined. Google's 89.2% share of U.S. general search queries in 2020, and 94.9% on mobile, are shares of general search services as the court defined them. They are not shares of all searching.
The defensible form is narrow: the court found that Google's ranking incorporates aggregated user interaction data collected over a period measured in months, and that finding is consistent with the expert's analysis. The indefensible form is the one that moves from the finding to a conclusion about a particular site without a method in between.
Where it helps in a damages case
Used within those limits, the record is useful in three recurring places.
First, channel substitution. A defendant may argue that traffic lost from Google could have been recovered elsewhere. The court's findings on Google's share of general search, and on the volume of queries Google receives relative to its rivals, are the most authoritative public evidence on how far other search engines can substitute for Google visibility.
Second, recovery lag. Where a site's rankings recover slowly after a migration error is corrected, or after a penalty is lifted, the finding that ranking incorporates months of accumulated user interaction is consistent with a slow recovery. It does not prove the cause of any particular lag, and the opinion should say so, but it answers the suggestion that a correctly repaired site should have recovered overnight.
Third, AI Overviews. The remedies court found that Google Search queries in the United States increased 1.5% to 2% after AI Overviews were introduced, and that some evidence suggests features like AI Overviews have reduced user interaction with organic web results. Both findings are hedged and neither measures any site's traffic, but they are findings of a federal court, and they frame the question better than most published statistics.
Citing the record accurately
The record is long and much of what circulates about it is secondhand. Cite findings by paragraph number, and pin cite the page of the opinion. Distinguish what the court found from testimony the court quoted and from exhibits the court cited; they carry different weight and opposing counsel will notice the difference. Quote exactly: the court states the Bing comparison as "over 17 years" in FOF ¶ 96 and as "over 17.5 years" in its conclusions, so quote whichever passage you cite.
The Final Judgment requires Google to make certain search index and user-interaction data available to qualified competitors on stated terms. It does not make that data available to litigants or the public, and it should not be described as a source of discovery in private search disputes.
Finally, keep the record in its place relative to the 2024 documentation leak. Where the two agree, as on Navboost, the court's findings corroborate the leak on that point. They do not corroborate the leaked material as a whole, and the findings are the stronger authority wherever they speak.
Frequently Asked Questions
Did the court in United States v. Google find that clicks affect rankings?
Yes, in a defined sense. The court found that Google's Navboost signal pairs queries with documents by memorizing user click data and has been trained on 13 months of user data since 2017, and that user data improves quality at every stage of search, including ranking. The court did not find how much weight click data carries in final rankings, and the findings do not establish why any particular site gained or lost rankings.
What is Navboost?
Navboost is a Google ranking signal described in the trial record as pairing queries and documents by memorizing user click data. The court found it was trained on 18 months of user data before 2017 and 13 months since, and quoted a Google document describing it as one of the most powerful ranking components historically, printed in the source as "most power." The remedies opinion describes Navboost data as an important component of Glue, Google's query and interaction log.
How is the antitrust record different from the 2024 Google documentation leak?
The antitrust record consists of findings of fact entered by a federal court after a contested trial, based on sworn testimony and admitted exhibits. The 2024 leak is a set of internal API documentation describing data fields without weights or context, which arrived outside any legal process. Where they overlap, as on Navboost, the court's findings corroborate the leak on that point. The findings are the stronger authority, and they should not be used to vouch for the leaked material as a whole.
Does the record show how much weight Google gives to click data?
No. The findings establish that click data is collected, how the court defined it, which systems were trained on it, and over what period. They do not quantify the contribution of any signal to final rankings. An expert who cites the record for a specific weight, or for the proposition that a change in click-through rate caused a specific ranking change, is attributing to the court a finding it did not make.
Did the court find that AI Overviews reduced clicks to websites?
The remedies opinion found that some evidence suggests features like AI Overviews have reduced user interactions with organic web results, and separately that Google Search queries in the United States rose 1.5% to 2% after AI Overviews were introduced. The first finding is expressly hedged, and neither measures any individual site's traffic. They are useful context in an AI Overview damages claim, not a measure of the claimant's loss.
Is the Google search antitrust case still on appeal?
As of this page's last review, yes. The liability opinion issued on August 5, 2024, the remedies opinion on September 2, 2025, and the Final Judgment on December 5, 2025, and appeals followed in the D.C. Circuit. Findings of fact remain citable as what the trial court found, but anyone relying on them in a filing should confirm the current status of the appeal first.
Published