How does Navboost work?
The evidence describes a lookup table rather than a learned model. Notes of a 31 January 2025 call with Nayak (plaintiffs' remedies exhibit PXR0357) describe it as "a QD table, a query-to-document lookup table" holding counts of user activity by document, subset by location and device type. Nayak testified that Navboost slices data by locale and by mobile versus desktop; asked whether Google has a specific Navboost for mobile, he said "It's one of the slices, yes."
Its place in the pipeline is early. Nayak explained to the court that retrieval culls the index to tens of thousands of documents, then signals including Navboost, topicality, PageRank and localisation reduce that to a few hundred before deep learning systems such as RankBrain re-rank the top results. He resisted the idea that Navboost alone does this, noting "there might be lots of documents that don't have clicks."
The May 2024 leak adds field names. The QualityNavboostCrapsCrapsData model lists goodClicks, badClicks, impressions, lastLongestClicks, and country, language and device slices, plus squashed and unsquashed variants.
- Training window: 18 months of user data before 2017, 13 months since, per the liability opinion.
- Scope: web results only. Nayak said Glue covers "all of the other features on the page" (transcript).
- Position: a core, traditional signal that feeds scoring before the machine learning re-rankers, per Nayak's testimony.
I wouldn't quite say navboost gets us to 2- to 300, because there might be lots of documents that don't have clicks.
Pandu Nayak, Google, Transcript of bench trial, Day 24 afternoon session, 18 October 2023, United States v. Google LLC
What evidence is there that Google uses Navboost?
The label here is Trial or leak evidence, from three separate sources. First, sworn testimony in United States v. Google by Nayak, Lehman and John Giannandrea, summarised in Judge Mehta's 2024 opinion, which found that Navboost ran on 13 months of data, "equivalent to over 17 years of data on Bing". Second, remedies phase exhibits: the HJ Kim call notes say Kim "has many patents related to Navboost". Third, the 2024 Content Warehouse leak, in which iPullRank counted 84 mentions of Navboost and five modules with it in the title.
What the evidence does not show is weighting. An exhibit quoted in the opinion calls Navboost "one of the most power ranking components historically", but no document gives its weight today. Google's response to the leak, reported by Search Engine Land on 29 May 2024, cautioned against "out-of-context, outdated, or incomplete information."
Navboost, a measure of how frequently users (subset by location and device type) click on a particular document for a particular query is a traditional signal. Uses most recent 13 months of data;
Notes of call with Pandu Nayak, PXR0357, notes of 31 January 2025 call with Google engineer Pandu Nayak, US Department of Justice
What does Navboost mean for your site?
Treat the search result as part of the page. If a page ranks but its title, snippet or lead fails the searcher, the evidence suggests the click record for that query and document can work against it over a 13 month window. The same notes record Nayak saying Google avoids simply "predicting clicks" because clicks are "easily manipulated", which is a reason not to chase click volume with bait.
The testable proposition: pages whose click-through rate sits well below the site's own norm for their position, over a sustained period, should on average lose position more often than pages that meet it. That can be falsified with your own Search Console history. Buying or faking clicks is not a strategy the evidence supports: the leaked models include squashed clicks and IP priors, which point to click spam filtering.
- Write titles that promise exactly what the page delivers.
- Answer the query in the first screen so the click is a long one.
- Monitor CTR against position, not raw CTR.
What are the common myths about Navboost?
Myth one: Navboost is the ranking algorithm. Nayak repeatedly told the court it is "certainly a factor, but it's not the only factor by any means." Myth two: the leak proves dwell time is a ranking factor. The leak names clicks and last longest clicks; it does not contain a field called dwell time, and SparkToro's own write-up states the documentation does not "prove which elements are used in the ranking systems." Myth three: Navboost is new. Testimony places it around 2005.
What are the key dates for Navboost?
- 2005: Approximate origin of Navboost, per Nayak's testimony ("somewhere in that range") (source)
- 2017: Navboost training window reduced from 18 to 13 months of user data (source)
- 2023-10-18: Pandu Nayak testifies about Navboost and Glue (source)
- 2024-05-28: Content Warehouse leak analyses published by SparkToro and iPullRank (source)
- 2024-08-05: Judge Mehta's liability opinion describes Navboost's 13 month training data (source)
- 2025-01-31: Call notes with Nayak (PXR0357) describe Navboost as a query-to-document table (source)
