← All posts

NavBoost, Glue and KScorer: what the Google trial proved — and what the leaked schemas only hint at

· 15 min read · By Mikhail Kuzmitskii

NavBoostGlueKScorerGoogle leakUS v. Googleuser signals

View as Markdown ↗

Straight to the spoiler: Google does count clicks, and a federal judge wrote it down — NavBoost runs on 13 months of click-and-query data. But “CTR is a ranking factor” and “click bots work” don’t follow from anything in the court record, and the schemas RESONEO published in October 2026, detailed as they are, show what Google logs, not what it weighs.

At the end of October RESONEO put out 137 internal Google schemas on NavBoost and Glue, then a second batch on a system called KScorer. I could guess the takes before reading one: “Google confirms CTR is a ranking factor”, round two. So I skipped the takes, opened three court documents from United States v. Google next to the two RESONEO pieces, and sorted every claim into piles: what the judge wrote, what witnesses said under oath, what only a leaked field name suggests. The piles turned out wildly uneven — and that’s the actual story!

Three levels of evidence

It sounds pedantic, I know. But without it you can’t tell a judge’s finding from a guess about a field name.

NavBoost, Glue and KScorer: what the Google trial proved — and what the leaked schemas only hint at
  1. Court findings — what Judge Amit Mehta wrote himself: the remedies opinion of 2 September 2025 and the Final Judgment text of 5 December 2025. Strongest level, with the judge’s own caveat that the court has “no expertise in the business of GSEs… or the engineering of GenAI technologies” (p. 2). Technical facts there are restated testimony.
  2. Testimony and exhibits quoted by a party — the plaintiffs’ proposed findings of fact (February 2024, May 2025). The quotes are real; the framing is the plaintiffs’ argument, not a ruling.
  3. Leaked schemas — field names. Authenticity unconfirmed, meaning inferred from names, no weights.

Below that: interpretations (RESONEO’s, a Russian-language course’s, mine) and patents — what Google could do, not what it does.

What the court actually wrote down

The sentence worth quoting, where the court repeats its own 2024 finding:

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. — Memorandum Opinion, US v. Google, ECF 1436, 2 September 2025, p. 89

One of the core models. Not “the” model. Hold on to that!

On Glue the opinion is just as plain:

Glue is essentially a “super query log” that collects a raft of data about a query and the user’s interaction with the response. … Navboost is a “memorization system” that aggregates click-and-query data about the web results delivered to the SERP. — Memorandum Opinion, 2 September 2025, p. 153

The court lists four groups of Glue data: the query (text, language, location, device), the ranking (ten blue links plus every triggered feature — images, maps, Knowledge Panel, People also ask), interaction with the page (clicks, hovers, time on the SERP), and interpretation data like spelling corrections and salient query terms. The 13 months are “queries and clicks” from all users, per Nayak; a Google exhibit labels the store “Glue Cache (13 months)” (p. 157). Glue data is “the bread and butter of Google’s scale advantage” (p. 157).

The Final Judgment (§ IV.B, p. 58) orders Google to share the user-side data used to “build, create, or operate the GLUE statistical model(s)” and “train, build, or operate the RankEmbed model(s)”. The data, not the recipe — the section “shall not require disclosure of intellectual property or trade secrets, such as algorithms, ranking signals, or post-trained LLMs”.

Two more findings puncture the “clicks decide everything” version. Quality signals are “constructed largely from sources other than user data”, with PageRank, built on links, as a key one (p. 143). And two exhibits suggest a popularity signal (P*) uses “Chrome visit data” and “the number of anchors”, but nobody testified about it, so the court “can say no more” (pp. 143–144).

So a court-confirmed click system exists, and it sits next to content and links, not above them. Less exciting than the takes? Probably.

What witnesses said under oath

The proposed findings hold the colourful quotes. Lehman, a Google engineer (remedies proposed findings, 29 May 2025, ¶632):

“Navboost is not a machine learning system. It’s just a big table. It says… for this search query, this document got two clicks.”

On Glue (¶622): “[Glue is] not a machine learning model. It’s a big table of data.” Pandu Nayak of Google, testifying: “Glue is just another name for NavBoost that includes all of the other features on the page.” The 2024 proposed findings (¶195) add that Glue “is a signal within Tetris/Tangram that triggers search features in the results alongside the web result”. And there’s a fast layer: Oard, the plaintiffs’ expert, testified that “Instant Glue is only looking at the last 24 hours of logs”, which lets updates reach search “in something on the order of 10 minutes” (¶1008).

How important is NavBoost? Here the piles disagree. An internal document (UPX0191) calls it “by far the most important component in search”; another (UPX0197) says “I’m pretty sure that NavBoost alone was/is more positive on clicks (and likely even on precision/utility metrics) by itself than the rest of ranking.” The plaintiffs wrote “one of the most, if not the most, impactful systems” (¶183), while the witness behind it, John Giannandrea, said “very important”. Nayak, asked by the judge, declined to rank it: in a sense, he said, the words on the page matter most, and NavBoost is one of the important signals (as the Dzhilavdarov course reads the transcript). An “I’m pretty sure” in an internal doc versus a witness under oath — I’d weight the oath.

Nayak also placed NavBoost in the pipeline: retrieval yields tens of thousands of candidates; topicality, PageRank, localization and NavBoost cut them to a few hundred; RankBrain looks only at the top 20–30; Tangram assembles the page. Does NavBoost alone get it to 200–300? No, he said — many documents have no clicks. NavBoost comes in only after a document is retrieved. Clicks don’t pull a page into the candidate set. Pin that above your desk the next time someone sells you “behavioral factors”!

And why Google never discussed any of this (proposed findings 2024, ¶207):

Do not discuss the use of clicks in search, except on a need to know basis with people who understand not to talk about this topic externally. Google has a public position. It is debatable. But please don’t craft your own.

Ha! That paragraph explains a decade of SEO arguments better than the arguments did.

What only the leaked schemas show

RESONEO says it obtained 137 schemas with 789 fields in 25 proto files — NavBoost, Glue, Instant Glue, Intent Brain, First Result Out — and that “Of the 88 Glue structures in these schemas, the leak documented one”, meaning the 2024 Content Warehouse leak.

Reliability first. The source isn’t disclosed, authenticity is unconfirmed. None of the 789 fields carries a comment (only one type does), so every meaning is inferred from a name. That one type, VoterTokenBitmapMessage, matches the 2024 leak word for word, Google’s comment included — a weak plus, since a comment can be copied. Some claims in the article (LlmGlueSignalData, KE_NAVBOOST, the example about Nice) aren’t in the published schemas at all.

With that said, what the names suggest in RESONEO’s reading:

  • Visibility. view_time_on_per_pixel divides time on screen by block height, so a two-line link and a carousel compare fairly; clicks_below marks a result seen and skipped.
  • Expected CTR. Models predict click and abandonment from a block’s appearance; the gap between observed and expected is what counts. Google also hides blocks on purpose to measure their effect.
  • The whole journey. Leaving without a click after an answer isn’t failure (abandoned_for_webanswers); reformulating after it means it wasn’t enough; if a refinement ends well, the previous page gets partial credit (downstream_mrr_goodness).
  • The date in the snippet. “For each broad intent, satisfaction is summed by the age of the displayed date: under a day, three days, a month, three months, a year.”
  • AI Overviews as one more Glue block. Counters for magi: expanded, collapsed, refused, scroll-backs, clicks opening the sources. “If searchers reformulate as often after an AI Overview as without one, the answer did not answer.” The name is real: the court’s findings describe Project Magi as the effort “which eventually became AI Overviews” (FOF ¶6).
  • The old 2024 click fields. CrapsSignals with good, bad and last-longest clicks is one branch of 34 per result. Squashed and unsquashed clicks, RESONEO now says, are one counter at two scales — correcting its own 2024 claim that unsquashed ones were the “reliable” clicks.

RESONEO sums it up as “A click is never counted alone. It is read against what the searcher saw, what they skipped and what they did afterwards.” Great sentence! It’s RESONEO’s, though, not Google’s. And Andrey Dzhilavdarov’s course checked: of the 12 fields RESONEO calls the main ones, only lastLongestClicks existed in the 2024 leak.

KScorer: the strongest part of the leak, still not the court

The KScorer article is better sourced: 15 schemas, 287 fields, and 145 of them carry Google engineers’ comments with formulas. Provenance is still undisclosed.

KScorer scores each interpretation of a query (“jaguar speed” — animal or car?), gives the dominant one rank 0 and drops the weak ones. The winner decides what appears around the results: a direct answer, a knowledge panel, a local block. That takes 56 signals: 11 of them vote on the query alone (PRE_RESULTS_VOTING), 41 once documents arrive (PACKER_VOTING), and 6 decide whether to show a card.

  • Clicks predict the raters. user_interaction_score: “This is an estimated IS score using user interaction data.” IS is the quality raters’ satisfaction scale; raters see few queries, clicks stand in for the rest.
  • Trust in clicks grows with impressions: “UIS confidence = total_impressions / (20 + total_impressions).”
  • The top 10 votes, weighted 1/position, so the first result weighs ten times the tenth. Entity confidence is a harmonic mean: entities at 0.9 and 0.2 give 0.33, not 0.55. One shaky entity drags the interpretation down.
Confidence in the click-based score vs impressions, N/(20+N)
20 impressions 50%
100 impressions 83%
1,000 impressions 98%

source: engineers' comment in the KScorer schemas published by RESONEO, Oct 2026 (authenticity unconfirmed)

Now the caveats. RESONEO says the top-10 vote has seven sources; its own schema marks six of them deprecated = true. The “secret sauce behind Googles ranking algorithm” line comes from ChatGPT retelling internal training guides that now return 410 — and even there, the course notes, it’s about choosing meanings, not ranking pages. The course’s sharper point: only the first vote precedes ranking. The second already knows web positions, so the chosen meaning drives answers and blocks; whether it moves web rankings, the schemas don’t show. In the 2024 leak, KScorer outputs sit in an Assistant ranker message (used_by_pfr_only: true).

What does the court add? Google’s proposed list of things it wouldn’t have to share (p. 58) names “information satisfaction score” and “query interpretation information”, so such objects exist. The name KScorer doesn’t appear in the court documents I checked. And per Google’s Elizabeth Reid, click-and-query data doesn’t play a “particularly meaningful” role in building the Knowledge Graph (pp. 149–150): clicks may help decide which card shows, but they don’t write the facts in it.

The evidence table

ClaimEvidence levelWhereDoes NOT prove
NavBoost counts clicks per query–document pair, 13 monthsCourt findingRemedies opinion, pp. 89, 153, 157That raw CTR moves rankings
NavBoost is “one of” the core ranking modelsCourt findingp. 89That it’s the main one
Glue logs the whole SERP, feeds TangramCourt finding + testimonyp. 153; PFF 2024 ¶195Weights, or AI Overview triggering
Instant Glue: 24 hours, ~10-minute updatesPlaintiffs’ expertPFF 2024 ¶1008That fake click bursts stick
Quality built largely from non-user dataCourt findingp. 143That clicks don’t matter
Clicks read against visibility and expected CTRLeaked schemas, no commentsRESONEO, Oct 2026Any weight or threshold
KScorer: clicks predict IS, 1/position voteSchemas with engineer commentsRESONEO, Oct 2026That meaning moves web rankings

The myths

“CTR is a ranking factor.” Not the CTR you see in Search Console! Kim, the Google engineer behind NavBoost, described a click historically as time on the page until the return to the results — in the visible part of the call notes (PXR0356); about two lines after it are redacted. Nayak’s call notes (PXR0357) say an early impressions-versus-clicks signal was biased by position, so an engineer built a position-corrected version. The schemas add expected CTR. A 2024 leak module that aggregates clicks per document keeps totals “in the decayed manner” and says to “divide by chances” for a rate. Asked whether CTR is a direct factor, the course gives a flat no: aggregated, normalized signals.

“Buying clicks works.” Nothing in the base shows it. What is shown: an IP-based prior of bad clicks in the 2024 leak, used in a file called craps-penalty.cc (algorithmic; nothing about manual actions); a count of distinct users per entry, which the leak describes as “used for privacy-related filtering”; Nayak’s call notes saying “clicks are easily manipulated and are a poor proxy for enhancing user experience”; NavBoost acting only on retrieved pages over 13 months. RESONEO in 2024: the effect is short-lived, then the 13 months dilute it. So does it never work? Nobody can say — thresholds are closed. There’s just no evidence it pays.

“NavBoost uses Chrome and Analytics.” RESONEO’s 2024 series: “the misconception that Navboost uses Chrome data, which it absolutely does not”. Chrome shows up in P*, weight unknown. On Analytics, the course cites Gary Illyes (2017) and John Mueller (2022): not used in the algorithm.

Same rule for “TopicAuthority”, “TrustScore” and other composite scores, ours included: tools’ heuristics, not Google signals.

What to do, without the conspiracy

You don’t need a single leaked weight for any of this:

  • Keep the snippet’s promise on the first screen. RESONEO’s advice: find Google landing pages with the least engaged visits — relative to the typical session for that page type — and fix the first screen.
  • Show a real date where freshness matters. Without a visible date a page probably drops out of the date-age counters (RESONEO’s reading). Same discipline as refreshing decaying content.
  • Answer the next question — details, options, the next step. A good refinement after your page credits you back.
  • Check which meaning the SERP chose before writing. The free rank checker shows who holds the top 10 and whether an AI Overview sits above. The course’s rule of thumb — under 40% of pages of your type in the top 10 means pick a query that names your meaning — is the author’s heuristic, to verify at T+7 and T+30.
  • Make entities unambiguous, named the way the Knowledge Graph knows them — the entity SEO basics. RESONEO’s advice, not schema content.
  • Judge CTR against your own position norm: average CTR of query–page pairs by rounded position, brand queries removed, deviation = CTR / norm − 1. Leave good abandonment (calculators, facts, rates) alone. Why clicks are scarcer anyway is in the autumn 2026 trends roundup.
  • Test snippet changes with a control group for 2–4 weeks — a title updates in a day or two, behaviour seems to shift over weeks (per an Illyes talk, via the course).
  • Keep bounce rate and time on site as your own diagnostics, not “Google factors”.
  • Don’t buy click manipulation; audit contractors selling “behavioral factors”.

In short

  • Court-confirmed: NavBoost counts clicks over 13 months and is “one of” the core ranking models; Glue logs the whole SERP; quality comes largely from non-user data.
  • Leak-only: visibility, expected CTR, the journey, date-age and magi counters, KScorer. Unconfirmed, zero weights.
  • Doesn’t follow: “CTR is a factor”, “click bots work”, “a field equals a weight”.

The upshot is boring in the best way: make the search end on your page. I suspect that’s the one recipe all three piles agree on.

Sources

  • U.S. District Court for D.C., United States v. Google LLC, No. 1:20-cv-03010-APM, Memorandum Opinion on remedies, ECF 1436, 2 September 2025 — pp. 2, 89, 143–144, 149–150, 153, 157; FOF ¶6 (PDF)
  • Same case, Memorandum Opinion with Final Judgment text, 5 December 2025 — § IV.B, p. 58
  • Plaintiffs’ Proposed Findings of Fact, liability phase, 23 February 2024 — ¶¶183, 195, 207, 1008
  • Plaintiffs’ Proposed Findings of Fact, remedies phase, 29 May 2025 — ¶¶622, 632, 639
  • Exhibits PXR0356 and PXR0357 — notes of calls with Google’s Hyung-Jin Kim and Pandu Nayak, remedies phase
  • RESONEO (Olivier de Segonzac), NavBoost & Glue schemas, October 2026 — think.resoneo.com/navboost-glue
  • RESONEO, “KScorer: how Google decides what a search means”, October 2026 — think.resoneo.com/kscorer; Intent Lab guides — think.resoneo.com/kscorer/intent-lab-guides
  • RESONEO, “Google Leak” part 5 (click data, NavBoost, Glue), 2024 — resoneo.com
  • Andrey Dzhilavdarov, course “SEO for Google from primary sources”, lessons 02–04 and 09, November 2026
  • Our own field checks against the 2024 Content Warehouse leak and RESONEO’s published schemas, 28 October and 14 November 2026

FAQ

Is CTR a Google ranking factor?

Not in the raw form Search Console shows. The court confirmed that NavBoost counts clicks for query–document pairs over 13 months, but notes of a call with Google's Pandu Nayak (exhibit PXR0357) describe a position-corrected version of the early impressions-versus-clicks signal, and the leaked schemas show clicks read against visibility and expected CTR. No source in the record gives a weight or threshold for CTR.

What is NavBoost?

A memorization system that aggregates click-and-query data about the web results Google shows. Google engineer Lehman testified it is 'just a big table'. The September 2025 remedies opinion calls it 'one of Google's core ranking models' that runs on 13 months of data. According to Nayak's testimony, it works only after a page has been retrieved as a candidate.

How is Glue different from NavBoost?

Nayak described Glue as 'just another name for NavBoost that includes all of the other features on the page'. The court calls it a 'super query log': the query, the whole results page with every triggered feature, clicks, hovers and time on the results page. It feeds Tangram, which assembles the results page. A fast layer, Instant Glue, looks at the last 24 hours with updates in about 10 minutes, per expert testimony.

What is KScorer and is it confirmed?

KScorer appears only in schemas RESONEO published in October 2026: it scores possible interpretations of a query with 56 signals, and click data predicts what quality raters would say. The schemas' authenticity is unconfirmed and KScorer is not named in the court documents we checked. Whether the chosen interpretation changes web rankings, rather than answers and SERP blocks, the schemas do not show.

Does buying clicks work on Google?

Nothing in the court record or the leaks shows that it does. The 2024 leak contains an IP-based prior of bad clicks applied in a file named craps-penalty.cc, Nayak's call notes say 'clicks are easily manipulated', and NavBoost only acts on pages already retrieved. Thresholds are closed, so 'it never works' is not proven either — but there is no evidence it pays.

Was this helpful?
Launch SEO/GEO with seo·matrix → or a request without Telegram →

Comments

  • …

Comments are moderated.