Navboost, manual signals, pogosticking: the hidden side of Google ranking

New documents from the US Department of Justice's trial against Google shed light on the internal workings of the search engine. Far from the company's public statements, these revelations point to widespread use of manual signals, a database called Navboost, and internal debugging interfaces that allow reading the ranking of pages signal by signal.

 This article in audio format

Navboost: not an artificial intelligence, but a giant click table

Contrary to what one might think, Navboost, one of the pillars of Google's ranking system, is not a machine learning algorithm in the strict sense. According to Dr. Eric Lehman, a former distinguished engineer at Google, " Navboost is not a machine learning system. It's just a big table. " He explains that this database records simple data, like the number of clicks a document received for a given query. In short, a huge table that maps user behaviors, query by query, page by page.

This system contradicts Google's public statements, which have often downplayed the role of clicks and behavioral signals in page ranking. Yet Eric Lehman's statements seem to prove that these metrics are indeed at the heart of the system.

Signals largely handcrafted

Another striking element in these documents: the vast majority of signals used in ranking do not come from machine learning, with the notable exception of RankBrain and DeepRank, two systems based on LLM-type models (Large Language Models). The other signals are explicitly "hand-crafted," that is, made by hand.

This means engineers analyze the data, choose a mathematical function, often a sigmoid, and manually define the trigger thresholds. This process of manual tuning allows Google to have fine-grained control over its algorithm's behavior, far from the image of a runaway autonomous AI.

Pogosticking exists (and Google knows it very well)

Another particularly sensitive point: the famous short and long clicks, which Google has always publicly rejected as direct ranking criteria. Again, the document does mention the time a user spends on a page before returning to the results page (SERP): a behavior nicknamed «pogosticking» in SEO jargon.

This type of interaction, coupled with clicks recorded in Navboost, clearly factors into the relevance evaluation and optimization systems.

A highly advanced internal debugging interface

One of the most interesting passages concerns an internal tool used by Google's engineers. An example is illustrated by the query « James Allan UMass » entered into Google, followed by the opening of a « debugger window ». This internal panel lists the top ten blue links shown, along with the scores assigned for each signal and an overall 'Final IR' (Information Retrieval) score.

This internal transparency, which is obviously not available to the public, allows engineers to understand exactly why a page is ranked at a certain position. This reinforces the idea that the search engine is carefully calibrated, query by query, sometimes even document by document.

Other intriguing elements: Q*, RankEmbed and the Twiddlers

Among the terms mentioned, but still nebulous, there is also Q*, RankEmbed and the "Twiddlers". These would be score adjusters, modules capable of dynamically modifying certain signals or applying post-processing corrections. They resemble intermediate layers whose role is to rebalance rankings after the raw signals are applied.

These modules suggest that Google's algorithm is less a closed, fully automated system than a gigantic hybrid engine, fed by data, manual rules, and fine-tuned edits.

This set of documents, uncovered by Brett Tabke (WebmasterWorld/Pubcon), offers a unique look into the guts of Google Search. They confirm what many SEO specialists have long suspected: behind the image of an all-powerful algorithm lies a far more manual, empirical, and tightly controlled system than commonly believed!

The article “Navboost, manual signals, pogosticking: the hidden side of Google ranking” was published on the site Abondance.