Google Leak: Why User Signals Matter More Than Thought

The SEO world is upside down. I admit it: hardly any other document has given me sleepless nights like this one recently. The court documents from the antitrust case against Google give us insight that goes deeper than anything we knew before. For years we discussed the importance of clicks, dwell time and user satisfaction, often based on correlations and tests. Now we have it in black and white.

What these documents contain is no longer a vague “user experience is important”. It is the technical blueprint for how Google measures user interactions and turns them into a ranking basis. If you thought SEO revolved mainly around keywords and backlinks, buckle up.

For this version I looked up every single page reference. 230 pages, eleven quotes. Three page numbers were wrong. One passage I had misread the first time around. Both are corrected below and verifiable in a table.

The central finding is that user data forms the foundation of the entire Google system. The documents unequivocally describe it as a critical input for search quality. A quote from internal page 152 gets to the heart of it:

“At every stage of the search process, user data is a critical input that directly improves [search] quality.”

So user data is a “critical input” that “directly improves” search quality. One could hardly state the importance more strongly. The sentence also says where this applies: at every stage of the search process, from crawling through indexing to retrieval and final ranking.

On internal page 92 the mechanism behind it appears:

“Every [user] interaction gives us another example, another bit of training data: for this query, a human believed that result would be most relevant.”

Every single user interaction therefore gives Google another training example showing that a human considered this result the most relevant one. Every click is a vote for the relevance of your content.

It has long been speculated how Google captures click data. The documents name two central systems: Glue and NavBoost.

1. Glue: the super query log

The system named “Glue” is described on internal page 153 as essential infrastructure:

“Glue is essentially a ‘super query log’ that collects a raft of data about a query and the user’s interaction with the response.”

Glue therefore works as a “super log” for search queries. It collects not only the query itself but a wealth of data about the user’s interaction with the displayed results. The document lists four groups: the query including language and location and device type · the ranking information of the results page · the interactions · the interpretation of the query. Among the interactions it explicitly names clicks, hovers, and duration on the SERP. So clicks, mouse movements over the results and time spent on the search results page.

2. NavBoost: the memory for user behavior

NavBoost builds on this data and is defined on internal page 153 as a memorization system:

“Navboost is a ‘memorization system’ that aggregates click-and-query data about the web results delivered to the SERP.”

NavBoost therefore remembers the aggregated click and query data. It is the huge database recording which results worked best for which queries in the past. Engineer Eric Lehman describes both systems in his testimony simply as “just a giant table”. Pandu Nayak adds: “Glue is just another name for [N]avboost that includes all of the other features on the page.”

The power of this system becomes clear through the sheer volume of data (internal page 89):

“NavBoost, one of Google’s core ranking models, runs on 13 months of Google click-and-query data. That is the equivalent of over 17.5 years of Bing data.”

To put that in perspective: Google collects as much click data in 13 months as Microsoft would need 17.5 years to accumulate. Phew, that is quite a statement. How these systems work together in the ranking cascade is covered in detail in my post on Google’s AI ranking with RankBrain, DeepRank and NavBoost.

The answer is RankEmbedBERT. This deep learning system combines two main data sources. On internal page 154 of the decision it says:

“RankEmbed and its later iteration RankEmbedBERT are ranking models that rely on two main sources of data: [share redacted]% of 70 days of search logs plus scores generated by human raters”

RankEmbedBERT therefore rests on two pillars: the search logs containing user click data and the scores from human quality raters. The exact share is redacted in the public version. Nayak himself places the effect like this: “RankEmbedBERT was again one of those very strong impact things, and it particularly helped with long-tail queries where language understanding is that much more important.”

The implications go beyond direct ranking. First, on crawling (internal page 93):

“Google has continuously deployed user data to, among other things, determine which websites to crawl, in what order, and at what frequency”

Google therefore uses user data continuously for the decision about which websites to crawl and about their order and frequency. Popular pages with heavy user interaction receive more crawl resources. The full chain from finding a URL to the final placement is covered in the overview of the Google algorithm.

This also gives Google an advantage on niche search queries, as internal page 91 explains:

“Google is better equipped to handle these types of queries in part because it simply sees more of them.”

Put simply: Google can answer niche queries better because it simply sees more of them. Through the sheer volume of data Google recognizes patterns in areas where other search engines have to guess. The 80/20 problem on internal page 139 belongs to this picture.

This question is central for SEOs, and this is exactly where reading the original pays off. On internal page 143 sits the sentence quoted in many SEO articles:

“Do you understand that most of Google’s quality signal is derived from the webpage itself?”

Anyone treating that sentence as a finding is reading it wrong. It is a question from the plaintiffs’ lawyers during cross-examination of Google’s expert witness Dr. James Allan. The court cites it to explain why Google does not have to hand over its quality signal. On the same page sits the court’s own assessment:

“‘quality measures’ are constructed largely from sources other than user data. For instance, a key quality signal is PageRank, which captures a web page’s quality and authoritativeness based on the frequency and importance of the links connecting to it.”

PageRank is therefore not being downgraded there, it is named a key quality signal. The frequently quoted line “PageRank […] is a single signal relating to distance from a known good source” sits in the same paragraph and continues with “and it is used as an input to the Quality score”. It describes what PageRank measures rather than diminishing it.

The honest reading is therefore: quality signals and user signals are two separate things in the ruling, not competitors in a hierarchy. User signals run through Navboost and Glue, quality through signals such as PageRank. Turning that into “content beats links” reads the passage against its own statement.

The documents point to a popularity signal labelled P* that uses Chrome data (internal page 144):

popularity signal (P*) “uses Chrome data”

At least one popularity signal therefore uses data from the Chrome browser. Google potentially sees how users interact with your website even when they did not arrive through Google search.

Two limitations belong with this. They were missing from the earlier version of this article. First, according to the same exhibits the popularity signal rests not only on Chrome visits but also on “the number of anchors”, meaning the number of links between pages. Second, the court applies its own brake:

“…but the court can say no more, as Plaintiffs offered no testimony on the matter.”

The record gives nothing beyond that hint. Deriving an instruction from it goes past the source. What remains is a plausible direction: brand loyalty and returning visitors certainly do no harm, but their ranking effect is not proven by this passage.

The document quoted here is the remedies decision of 2 September 2025. It therefore dates well after the 2024 leak cycle. It also contains an order that puts the whole dispute in perspective. On internal page 157 the court explains why the Glue dataset in particular has to be shared:

“The sharing of the dataset underlying the Glue statistical models, on the other hand, presents a stronger case for inclusion in the final judgment.”

Google has to make this data available to qualified competitors at marginal cost. The models themselves stay explicitly out of scope. What gets shared is the raw material and not the algorithm on top of it. For the question of whether user signals matter, that is a stronger argument than any correlation study: a court orders the release of exactly this data and grounds the order in its competitive importance.

The case is not closed with that. Appeals against the first-instance decision remain possible, and what will ultimately be implemented is still open. The quoted findings about how the systems work are unaffected by this.

Anyone wanting to go through the record can find the full procedural history in the public docket at CourtListener. Google’s official counter-presentation sits in its own documentation on how Search works.

Okay, enough theory. What do you do with this information now? Running away is not an option. Here are your concrete next steps:

• Optimize beyond the click: It is no longer just about click-through rate. Analyse user behavior on your site rigorously.
• Reduce pogo-sticking: Does the user jump straight back to search from your page? That is a negative signal. Make sure the first visible area satisfies the search intent immediately.
• Increase dwell time: Offer real value, useful internal links and content that holds someone. Keep in mind: what Google measures is not the number in your analytics account.
• Encourage interactions: Make sure users do something on the page, watch videos, leave comments, click through galleries.
• Master the long tail: Win on niche queries with comprehensive content that solves very specific problems.
• Answer real questions: Research “People also ask” questions and answer them in detail. How Google parses the query beforehand is covered in the post on query processing.
• Semantic depth: Cover not only the main keyword but all relevant subtopics and entities.
• Prioritize E-E-A-T and content quality: My comprehensive guide to E-E-A-T shows you how to get there.
• Build trust: Invest in transparency, author profiles and genuine user reviews.
• Encourage returning visitors: Build a brand that users trust. Whether Chrome data feeds into ranking is open according to this record, but brand loyalty does no harm either way.

And one item that does not belong on the list: buying clicks. Why such campaigns stand out is covered under detecting CTR manipulation.

The court documents mark a turning point. They confirm what many SEOs have been saying for years: long-term success is based on genuine user satisfaction and not on signal manipulation. Google has a system that measures whether your content offers added value. Every shortcut that ignores the user shows up eventually through systems like Navboost and Glue.

What never appears once in 230 pages of the record is a number of seconds or a threshold value. What gets described is a system with a memory spanning months for the question of which result people considered the fitting answer. So create content that users not only click on but are satisfied with. Because Google is watching, more closely than we thought.