(Basics)

Navboost

Definition

Navboost is an internal Google system that memorizes searchers’ click history and uses it to reorder search results. Because it adjusts the ranking of results that have already been retrieved, it’s classified as a re-ranking system. Pandu Nayak, Google’s VP of Search, acknowledged in court during the 2023 antitrust trial that Navboost is one of the important signals Google has. Trial transcript

It’s quite different from the AI models people usually picture. Google engineer Eric Lehman testified that Navboost is not machine learning but closer to a giant ledger recording how often which documents were clicked for which queries. Search Engine Land

An internal Google document quoted in the court’s opinion sums up the idea like this: if a document gets a positive reaction, we figure it’s good; if it gets a negative reaction, it’s probably bad. Grossly simplified, that is the source of Google’s magic. Court opinion

How does it work?

  • According to the court’s opinion, Navboost is a signal that memorizes user click data to match queries with documents. Court opinion
  • Nayak confirmed that its memory window is the past 13 months, and that before 2017 it was 18 months. Trial transcript
  • The opinion noted that these 13 months of data are equivalent to more than 17 years of data collected by Microsoft’s search engine, Bing. Court opinion
  • Navboost can only boost a document’s ranking after the document has been retrieved in the first pass and made it into the candidate set. Trial transcript
  • The candidates are narrowed to tens of thousands and then cut down to a few hundred (around 200–300), but Nayak cautioned that this narrowing isn’t Navboost’s work alone. Trial transcript
  • Click data is stored separately by country and state, and by mobile and desktop. SparkToro

Later, in evidence from the remedies hearing (the phase where the winning side argues over what corrective measures to obtain), clicks were described as one of three pillars, along with link anchors (A) and body text (B), that make up the basic relevance score. Search Engine Land

Glue and Instant Glue

  • Nayak defined Glue as another name for Navboost that handles every element on the SERP other than the web document results. Trial transcript
  • Instant Glue is a fast version that looks only at the last 24 hours of logs and updates about every 10 minutes, as the plaintiffs’ expert, Professor Douglas Oard, explained in the liability trial. DOJ record

How did it become public?

The name Navboost came out through two events. They’re often cited together, but they’re separate events and should be read separately.

Testimony at the 2023 antitrust trial

  • The US Department of Justice and 11 states sued Google in October 2020, and the trial ran for 9 weeks starting in September 2023. Court opinion
  • On October 18, 2023, VP Nayak acknowledged on the witness stand that Navboost memorizes past clicks and is trained on user data. Trial transcript
  • On August 5, 2024, Judge Amit Mehta ruled that Google is a monopolist and violated Section 2 of the Sherman Act. Court opinion
  • The opinion cited an internal document saying that even after the arrival of signals based on large language models, Navboost has historically remained one of the most powerful ranking components. Court opinion

The May 2024 API documentation leak

Separately from the trial, the leak began on March 13, 2024, when an automated bot mistakenly posted Google’s internal API documentation to a public GitHub repository. The documents ran to more than 2,500 pages with about 14,000 attributes. An industry figure who found them passed them to Rand Fishkin, who shared them with Mike King, and they were analyzed publicly in late May 2024.

  • The click signal module in the 2024 leaked documents does contain fields such as goodClicks, badClicks and lastLongestClicks. Leaked docs mirror
  • However, the original documents don’t define these fields, so interpreting them as satisfied and unsatisfied clicks is the work of analysts such as Fishkin and King. Leaked docs mirror
  • Mike King counted 84 mentions of Navboost across the 2024 leaked documents. iPullRank
  • Fishkin judged that the documents directly contradict Google’s repeated denials that it uses click-centric user signals. SparkToro

A Google spokesperson acknowledged the documents were authentic, but cautioned against making inaccurate assumptions about Search based on out-of-context, outdated or incomplete information.

Google’s official position vs. what came to light

Google’s official positionGoogle spokespeople have long denied click theories. At a 2016 conference, Gary Illyes said clicks are generally incredibly noisy. Search Engine Journal In a 2019 public Q&A, he dismissed dwell time and click-through rate theories as generally made up. Impression Digital (AMA summary) In 2021, John Mueller also answered that if click-through rate moved rankings, results would be full of clickbait titles. John Mueller’s statement
What the trial and the leak revealedAt the 2023 trial, VP Nayak acknowledged that Navboost is an important signal built by memorizing click data, and the court’s opinion established it as fact. The 2024 leaked documents even revealed fields that store clicks in several categories. Court opinion

Strictly speaking, Google’s official documentation also says it uses aggregated and anonymized interaction data to assess the relevance of results. Google (official)

So the focus of the denials wasn’t clicks themselves, but the claim that individual click-through rates aren’t fed directly into rankings.

Why click manipulation doesn’t work

Once it’s confirmed that clicks are used in rankings, it’s tempting to think of click-exchange schemes or manipulation programs. But the testimony and leaked documents actually show that there are layers of defenses.

  • Google uses cookie history, signed-in Chrome data and pattern detection as defenses against manual and automated click spam. SparkToro
  • Because click data is split by country, state and device (geo-fenced storage), the effect of remote click farms is diluted. SparkToro
  • Squashing (normalization that keeps one big signal from overwhelming the rest) is also applied. iPullRank
  • Google deliberately avoids simple click prediction, because it’s easy to manipulate and doesn’t reliably measure user experience. Search Engine Land

In fact, when Fishkin ran a public experiment in 2014 asking his followers to search for a specific query and click, the result rose from #7 to #1 in about 3 hours, but he himself cautioned that the effect could be temporary and was not conclusive proof. SparkToro

What it means for SEO practice

1. Satisfaction after the click is the real signal

In the remedies hearing evidence, the click signal was defined as the time a user stays on a document before returning to the search results page. DOJ record

If you push up your CTR with clickbait titles alone and visitors bounce right back (pogo-sticking, bouncing back to the results page), you only pile up bad records. First make sure the body delivers what the title promised, in line with search intent.

2. Invest in site-level quality

A Google engineer said that the page quality score is largely static regardless of the query, tied to the site level, and still very important today. DOJ record

In the long game, what matters more than a few clicks on individual posts is building up the trust represented by E-E-A-T , in other words, assets in the nature of domain authority .

3. Click manipulation services are a waste of money

Click manipulation programs and click-exchange services that promise higher rankings get filtered out by the defenses above, and their effect is scattered by storage that’s split by region and device. Even if rankings move briefly, they don’t hold.

4. Use real data to find mismatches

You can see which queries your site appears and gets clicked for in Google Search Console. It doesn’t show behavior after the click, but it’s a starting point for finding pages with many impressions but few clicks, or suspected bounces after the click, and fixing mismatches between titles and body content.

References

Aug.2026By Studio JT

More from Basics

Like this term?
Give it a like!

(167)