Navboost is Google's click-based re-ranking system: it moves a result up or down for a query according to how searchers behaved after clicking it. Google confirmed the system under oath in 2023. The mechanism is described in patent US 8,661,029 B1, "Modifying search result ranking based on implicit user feedback."
What is Navboost?
Navboost is a ranking signal built from aggregated click data for each query. Pandu Nayak, Vice President of Search at Google, described it in October 2023 at the United States v. Google antitrust trial as "one of the important signals" Google has. He testified that Navboost memorizes clicks over a rolling 13-month window, down from 18 months before 2017.
The patent never uses the word "Navboost." The connection comes from its content: the same inputs (clicks, time on page, query, language, country) and the same output (a new order of results for future searches on the same query).
What does patent US 8,661,029 describe?
Patent US 8,661,029 describes a relevance score computed from the ratio of long views to all views of a result for a given query. Google filed it on November 2, 2006 (application 11/556,143), and the patent was granted on February 25, 2014, with 30 claims and 5 drawing sheets. Its 4 inventors are Hyung-Jin Kim, Simon Tong, Noam Shazeer and Michelangelo Diligenti. Five continuation patents extend the same family, the latest being US 11,816,114 B1.
Claim 1 protects the principle in broad terms: a measure of relevance based on "a subset of a count of views of the document following selections of the search result" in relation to the full count of those views, sent to a ranking engine.
For every click on a result, the patent lists the information the system can record:
- Query: the search the user typed.
- Document: the result the user clicked.
- Time on document: the time between the click on the result and the moment the user comes back to the results page and clicks another result.
- Language: the interface language of the user, which the user can set.
- Country: the country where the user is likely located, identified for example by the Google domain used (google.co.uk for the United Kingdom).
- Other aspects of the user and the session.
The patent cites further data that can be logged: results that were shown but not clicked, click positions, IR scores of the results, the titles and snippets shown before the click, the user's cookie and its age, the IP address and the browser user agent. Clicks are tracked by JavaScript code in the results page, which also detects the return to that page. A toolbar installed on the user's computer can also track post-click browsing, provided the user opts in to share it.
How does Google measure a good click?
Google measures a good click by its duration. The core metric of the patent is the LC|C click fraction: long clicks divided by all clicks on a query-document pair. In the patent's words, it is "the number of Long Clicks (which may be weighted clicks) divided by the number of Clicks overall," and "a normalized measure of how long people stayed on a page given that they clicked through to that page."
The base formula is LCC_BASE = #WC(Q,D) / (#C(Q,D) + S0): the sum of weighted clicks for the query-URL pair, divided by the total number of clicks plus a smoothing factor S0. When a query has few clicks, the fraction tends toward zero. When clicks far outnumber S0, the smoothing no longer matters. A result needs enough data before its clicks count.
The patent gives an example of weights by viewing time category.
| Click type | Weight | Meaning in the patent |
|---|---|---|
| Short click | −0.1 | Indicative of a poor page |
| Medium click | 0.5 | Indicative of a potentially good page |
| Long click | 1.0 | Indicative of a good page |
| Last click (the user does not return to the results) | 0.9 | Likely indicative of a good page |
| Last click preceded by another click | 0.3 | Less indicative of a good page |
These values are an example, not constants. The patent says the time frames and weights can be tuned by comparing click logs with human generated ratings of result quality, and that the weighting can also be a continuous function of viewing time instead of fixed categories.
In this example, the short click carries a negative weight. A quick return does not simply fail to count: it lowers the sum of weighted clicks of the page.
Why does ranking position not bias the score?
The score compares a result with itself, not with the results around it. The patent notes that users tend to click results with good snippets, or that rank higher, "regardless of the real relevance of the document." It calls this presentation bias: an attractive title or snippet, and the position in the ranking. The LC|C fraction measures how long users stay once they have clicked, which makes it, in the patent's words, "relatively immune to presentation bias." The fraction "can be high for a given URL even if it gets far fewer clicks than a comparable result in a higher position." A page in position 8 that keeps its visitors can outscore a page in position 1 that users leave in seconds.
How does the score change rankings?
The LC|C fraction can be applied directly to the result's information retrieval (IR) score, or turned into a boosting factor applied to the IR score. The example in the patent is a sigmoid: IRBoost = 1 + M / (1 + e^(X·(LC|C − 0.5))), with constants such as (M, X) = (100, −100). A low LC|C fraction produces a boost of about 1, so the ranking barely moves. Above 0.5, the boost climbs fast toward its maximum. The patent also gives linear and exponential variants, and says the transform is chosen by tuning against human relevance ratings.
The LC|C fraction can also be used on top of the traditional click fraction, which compares the clicks on a result with the weighted clicks on all results for the query.
How does the patent segment the score by language and country?
The system computes the fraction at 3 levels: base (all clicks), per language, and per country and language. Each level is smoothed toward the level above it, so "if there is less data for the more specific click fractions, the overall fraction falls back to the next higher level for which more data is available."
The smoothing factors are not uniform either. The patent raises them for high-traffic segments (it cites US-English queries) and "for query sources that have historically generated more spamming activity (e.g., queries from Russia)."
Does a good click last the same time for every query and every user?
No. The patent adjusts the thresholds, or the formula, for "what constitutes a good click" with viewing length differentiators: the category of the query and the type of user.
- Query category. A navigational query targets one page or site (the patent's example is "BMW"). An informational query has many equally useful pages ("George Washington's Birthday"). The patent splits informational queries further: a birthday takes seconds to read, while "Hilbert transform tutorial" needs a good deal more time. Categories can be detected from the skew of IR scores or click fractions (one dominant result suggests a navigational query), by regression on historical clicks, or by K-means clustering of average dwell times.
- User type. Computer savvy users often click faster than less experienced users, so users can receive different weighting functions, down to a single user. A user who almost always clicks the top result can have his good clicks weighted lower than a user who more often clicks results further down. A user who issues many queries on a topic (the patent's example is law) can be presumed to have expertise, and his clicks on that topic weighted accordingly. Weighting clicks per user brings the score close to search personalization, where Google builds a profile for each user.
Can fake clicks improve rankings?
No, the patent is designed to prevent it. It anticipates "spammers (users who generate fraudulent clicks in an attempt to boost certain search results)" and describes safeguards. The first is "a user model that describes how a user should behave over time": click data from users who do not conform to it can be disregarded. The safeguards have 2 objectives:
- Democracy in the votes, for example "one single vote per cookie and/or IP for a given query-URL pair."
- Removing unnatural traffic: cookies or IP addresses with an abnormal distribution of click positions, click durations or clicks per minute, hour or day are removed entirely.
Queries that look spammed, for example with an abnormal distribution of user agents or cookie ages, need not use click signals at all. Click bots and click farms produce exactly these traces: repeated votes from the same sources, uniform durations and unnatural click rates.
What does Navboost change for your SEO?
Navboost rewards pages that satisfy the click, not pages that only win it. The patent leads to 5 practical consequences:
- Write titles that match the page. A misleading title wins the click and loses the score, because the user returns fast and logs a short click.
- Put the answer on the first screen. A visitor who sees right away that the page answers the query has a reason to stay. Give the answer before the context.
- Aim to be the last click. The last click weighs 0.9 in the patent's example, against 0.3 when the user clicked another result first. Complete the task (the answer, the comparison, the next step) so the user has no reason to return to the results.
- Match the expected depth of the query. The threshold for a good click depends on the query category: a quick fact is satisfied in seconds, a tutorial needs a longer visit.
- Read Search Console by country. The score is segmented by language and country, so the same page can satisfy searchers in France and fail them in Belgium on the same query.
In our on-page SEO work, we compare each page's title with what its first screen delivers, since that gap is what turns a won click into a short click.
Navboost does not replace relevance. A page has to be retrieved and scored by Google's other systems, such as PageRank and the authority that flows through links, before clicks can move it. Clicks decide the final order among pages that already qualify for the query.
The patent describes what Google's system can do. The 2023 testimony confirms that Navboost exists and matters, not that it uses these exact weights.