NavBoost: what the DOJ trial and the API leak actually confirmed

Two disclosures in 2024 landed on the same subject from opposite directions. In November 2023, Google VP of Search Pandu Nayak took the stand in United States v. Google and described a system called NavBoost. Six months later, an internal Google document reached SparkToro’s Rand Fishkin from an anonymous source and named click categories that line up with what Nayak described under oath. Neither event was staged to corroborate the other. Read together, they are the closest the public has to independent confirmation of how click data moves through Google’s ranking pipeline.

What Nayak testified to#

Google’s VP of Search told the court that NavBoost is “one of the important signals” Google uses, not the retrieval system itself. Search Engine Land’s trial coverage quotes him describing it as a memorization system that works from clicks on a query going back roughly 13 months, a window that was 18 months before a change in 2017.

Its job in the pipeline is culling, not fine-tuning. Nayak’s account, as reported, places NavBoost early: an initial retrieval stage returns a large candidate set of documents, and NavBoost uses historical click behavior to cut that set down before later ranking stages run. That is a different claim than “clicks nudge the final order,” and it is worth keeping the two apart when the testimony gets summarized.

It cannot help pages with no click history. By Nayak’s own description, a new or rarely-clicked page gets nothing from NavBoost, because the system has no aggregated behavior to memorize for it. That limitation was part of the testimony, not an inference added afterward.

What the leaked document named#

The Content Warehouse API leak is a separate chain of custody entirely. An anonymous source sent internal documentation to Erfan Azimi, who passed it to Rand Fishkin at SparkToro, who published his findings in May 2024 alongside analysis from iPullRank’s Mike King. Search Engine Land’s coverage of the same leak confirms the same field names independently: badClicks, goodClicks, lastLongestClicks and unsquashedClicks appear in the documentation as distinct click-related attributes.

The document names features; it does not state their weight. Both outlets are explicit on this point. Search Engine Land reports that the leak “did not specify how any of the ranking features are weighted, just that they exist.” Fishkin went further in his own post, warning readers directly not to treat a named API field as proof that Google currently uses it the way the name implies - a field can be retired, reserved for testing, or never shipped to production.

Treat the exact terms as leaked shorthand, not a public spec. goodClicks, badClicks and lastLongestClicks read as plainly as the terms in this piece describe them, but they come from internal engineering documentation that Google has never authenticated in public. That caution applies to the names themselves, not just to their weighting.

The tension with Google’s own public position#

For roughly two decades, Google’s public statements treated a direct ranking role for click data as something between unconfirmed and denied. A 2019 Google response reported by Search Engine Land said interactions are used “for personalization, evaluation purposes and training data,” a sentence built to avoid the word ranking rather than address it directly.

This piece states the tension; it does not resolve it. Sworn testimony describing NavBoost as an important signal sits uneasily next to years of public phrasing that stopped short of confirming any such role. That is not the same as proof that every earlier public statement was false: Google’s representatives may have been describing a narrower system, working from incomplete internal knowledge, or using “ranking signal” more narrowly than NavBoost’s actual function. What can be said without overreaching is that the public framing and the sworn testimony do not match, and neither this trial nor this leak is the last word on exactly how today’s ranking pipeline weighs a click.

Why we care#

If a page has real search demand but no click history, in Nayak’s own account of NavBoost, this particular signal has nothing to offer it - which puts the weight back on other ranking inputs and on technical SEO fundamentals that do not depend on accumulated behavior data. Neither the testimony nor the leak is grounds for chasing a specific click-pattern tactic; both sources here explicitly warn against reading more precision into named fields or a single number than the disclosure supports. Read the primary coverage linked above before repeating a specific figure, and hold the 13-month window and the exact leak terminology as reported claims, not settled specification.