NavBoost and user interaction signals: what’s proven, what’s guesswork
For two decades, one question divided the SEO world more than any other: does Google use user behavior — clicks, pogo‑sticking, time on page — to rank results? Google’s public answer was no, repeatedly and with feeling. Then 2024 happened twice: an anonymous source leaked thousands of pages of Google’s internal API documentation, and the Department of Justice’s antitrust trial put Google’s own engineers and exhibits on the record. Today we can answer the question with unusual precision — and the precise answer is stranger than either camp expected. What exists is proven. What it weighs is not. And the entire click economy the system runs on is shrinking under our feet.
In short
- NavBoost is court‑documented: a click‑based re‑ranking system, running since around 2005, working on a 13‑month window of click data.
- The weight is unknown — the leak shows what data Google holds, not how the ranking formula uses it. Anyone claiming “NavBoost is the #1 factor” is guessing.
- Google’s rhetoric shifted: from “made up crap” (2019) to “aggregated and anonymized interaction data” (June 2024).
- The click pool is collapsing: with AI Overviews present, click‑through on the #1 organic result is down 58%; randomized experiments show double‑digit drops in outbound clicks.
- The signal migrates to brand demand: branded mentions correlate with AI visibility at 0.664, backlinks at 0.218.
- Don’t manipulate: Google explicitly says clicks are “easily manipulated” — the one thing proven to get you filtered.
What the court documents actually established
Start with what is now documented fact, not opinion.
In May 2024, an anonymous source shared thousands of leaked Google Search API documents with Rand Fishkin (SparkToro) and Mike King (iPullRank), who published an analysis. The leak covered 2,596 modules and 14,014 attributes of an internal API, and NavBoost — a click‑based system — appears in it 84 times across five modules. The leaked fields read like a taxonomy of user intent: clicks, goodClicks, badClicks, lastLongestClicks — defined as clicks that were both the last and the longest in related user queries — and a distinction between “squashed” and “unsquashed” click counts.
Then came the trial. Pandu Nayak, Google’s VP of Search, testified under oath that NavBoost has existed since around 2005, that its data window was 18 months until 2017 and is now 13 months, and that it works alongside Glue, a related system for ranking the non‑organic elements of the results page. A Justice Department exhibit (PXR0357) describes NavBoost as a measure of how frequently users click on a document for a query, segmented by location and device, over the most recent 13 months (DOJ exhibit index). Judge Amit Mehta’s August 2024 ruling cited material calling it “one of the most powerful ranking components historically.”
But the same documents contain the limitation that keeps this honest. Nayak’s notes state: “Google avoids simply ‘predicting clicks’ because clicks are easily manipulated and are a poor proxy for enhancing user experience.” And Mike King’s own write‑up carries the caveat every serious SEO should memorize: no scoring functions were in the leak — we do not know how features are weighted. The leak proves the data exists; it does not prove the recipe.
What Google says, before and after
The before‑and‑after is instructive. In a 2019 Reddit AMA, Google’s Gary Illyes dismissed the whole discussion: “Dwell time, CTR, whatever Fishkin’s new theory is, those are generally made up crap.” After the leak, in June 2024, Google’s Search Liaison Danny Sullivan acknowledged the company uses “a variety of different ranking signals including, but not solely, aggregated and anonymized interaction data” (coverage).
That is not a confession that CTR is a direct ranking switch. It is an admission that interaction data sits somewhere in the stack — consistent with everything the trial established. The honest position in 2026: user interaction signals are real and documented; their weight is unmeasured; and Google has a documented institutional bias against signals that are easy to fake.
The click economy is shrinking
Here is the part nobody argued about in 2005 and everybody has to argue about now: the pool of clicks the whole system feeds on is draining fast.
- Pew Research Center (900 US adults, ~69,000 searches): when an AI Overview appears, users click a conventional result in 8% of searches versus 15% without one (Pew, July 2025).
- Ahrefs (300,000 keywords): CTR of the #1 organic result when an AI Overview is present fell 34.5% by April 2025 and 58% by December (Ahrefs).
- A randomized field experiment on 1,065 real users (SSRN, 2026) measured a 39.8% drop in outbound organic clicks when AI Overviews appeared, with zero‑click share rising from 54% to about 73% (study overview).
- A preregistered RCT from Penn/Northeastern (N=1,100) found forced AI Mode reduced external clicks by 18.8 percentage points — and pushed 11.2 points of users toward Bing and DuckDuckGo (overview).
One nuance survives all the doom: Seer Interactive’s data shows brands cited inside AI Overviews received 35% more organic clicks than expected. Being the cited source still pays. Being merely listed does not.
Fewer clicks, more brand: where the signal went
If NavBoost runs on clicks and clicks are vanishing, what happens to the signal? The best current answer comes from Kevin Indig’s clickstream analysis of 846,000 Google sessions (February–March 2026): in AI Mode, 64% of users click nothing at all, and behavior splits into two modes — “AI Mode is autoplay. AI Overviews is the Netflix browse” (Growth Memo). Most users accept the AI’s shortlist without browsing.
Meanwhile the measurable currency of visibility moved. Ahrefs’ study of 75,000 brands found branded web mentions correlate with AI answer visibility at 0.664, while backlinks manage 0.218 (Ahrefs). Lily Ray’s ongoing tracking adds a warning for shortcut‑takers: Google’s AI Overviews now cite self‑promotional “best of” listicles 38% less, and omit the brands behind them 83% of the time (Lily Ray). AI systems trust consensus, not self‑description — the same pattern as the click signals, one level up: you cannot force the signal, you can only earn the underlying behavior.
What to actually do
The practitioner consensus that survived the leak is almost disappointingly old‑fashioned:
- Keep the promise of the snippet. The documented failure signal is the bad click — a user who clicks, sees clickbait, and returns. Write titles the page can cash in on. This maps directly onto the click types Marie Haynes organizes her guidance around: the click, the long click, the last click (Marie Haynes on NavBoost).
- Build brand demand. The documented correlates — mentions, branded search, presence in the places AI systems read — all point the same way. Brand is the input you can actually invest in when click volume can’t be relied on.
- Measure the right things. Google’s own Search Console AI reports show impressions, not clicks, for AI surfaces — a hint about what Google itself considers trackable. Track branded demand and cited‑source status, not just sessions.
- Treat every click you still get as scarce. One good click from a satisfied user is worth more than ten bounces; the documentation says Google literally counts the difference.
What not to do
Click manipulation. Not primarily for moral reasons — for documented, practical ones. Google’s trial exhibit says clicks are “easily manipulated”; the leak contains the machinery of filtering; and circulating claims that a few dozen purchased clicks a day can push rankings remain unproven speculation, contradicted by the very documents they cite. If someone sells you CTR bots, they are selling the one lever Google has publicly said it defends against. We cover the wider picture — including this shift from links to mentions — in Is SEO dead in the AI era?.
Summary
The behavioral question of SEO is no longer “if” — it is “how much,” and “how much” is genuinely unknown. What is known: NavBoost exists and runs on clicks; the click pool is shrinking by double digits a year; the surviving leverage is satisfaction and brand demand. Optimize for the user who clicked and stayed, invest in being mentioned, and ignore anyone selling clicks by the dozen.
If you want a strategy built for this reality — SEO and GEO that targets satisfaction, brand demand and AI citations — that is our SEO & GEO service, grounded in the technical SEO basics that still decide whether any of these signals can reach you at all. For the file layer those crawlers depend on, see our guide to robots.txt, sitemaps and llms.txt.
Yuri Osipov
Founder, SEO & analytics
Runs Take the Space. Specialises in technical SEO, measurable analytics setups and AI search visibility.