📄Optimizing Search Engines using Clickthrough Data
Joachims, 2002. ACM SIGKDD 2002.
Brass-SEO cites this as the academic foundation for treating clicks as a ranking signal. Joachims showed that clickthrough data — the query log paired with which results users clicked — can train a retrieval function, and that a meta-search engine trained this way outperformed Google for a specific group of users 'after only a couple of hundred training examples.' The core idea is that a click on a lower result when a higher one was skipped is implicit evidence that the lower result was more relevant. This is the intellectual root of every later click-based ranking system: user behavior is relevance feedback that is 'virtually free and substantially more timely' than human judgments.
- Examines:
- The foundational method for using clickthrough logs as implicit relevance feedback to train search ranking functions, with an SVM-based approach evaluated against Google.
- Brass-SEO draws on:
- The clicks-as-implicit-relevance-feedback principle — the basis for Brass-SEO's explanation that interaction data has informed ranking for over two decades.
📄A Dynamic Bayesian Network Click Model for Web Search Ranking
Chapelle & Zhang, 2009. WWW 2009. DOI 10.1145/1526709.1526711.
Brass-SEO references this for the central caveat about click data: it is biased by position. Chapelle and Zhang's click model addresses the well-documented problem that results in lower positions are less likely to be clicked even when they are equally relevant, so a raw click count overstates the appeal of whatever already ranks first. Their model separates the probability that a result was examined from the probability that it was genuinely relevant. Brass-SEO uses this to temper click-based reasoning: a page that ranks first will collect more clicks regardless of quality, so chasing clicks without improving the underlying result is circular.
- Examines:
- A probabilistic click model that corrects for position bias when inferring relevance from click logs, a core problem in using interaction data for ranking.
- Brass-SEO draws on:
- The position-bias correction — cited when Brass-SEO explains why raw click counts are an unreliable proxy and ranking position itself inflates clicks.
📊U.S. v. Google — Testimony on Navboost (Pandu Nayak)
U.S. District Court, D.D.C., No. 1:20-cv-3010. Trial testimony, October 18, 2023.
Brass-SEO cites this as the strongest direct evidence that Google uses click data in ranking. Under oath in the federal antitrust trial, Google's VP of Search Pandu Nayak confirmed that Navboost is one of Google's 'core' ranking systems, that it 'memorizes past clicks that have been issued for past queries,' and that it 'is trained on user data.' He testified that Navboost retains this click information for queries received over a rolling window — 13 months at the time of testimony, previously 18. Nayak characterized it as 'certainly a factor' in ranking, 'but by no means the only factor.' This moves the click question from inference to sworn record.
- Examines:
- Sworn trial testimony from Google's VP of Search confirming Navboost, a core ranking system that memorizes and uses past click data for queries.
- Brass-SEO draws on:
- Nayak's confirmation that Navboost memorizes clicks and is 'certainly a factor' in ranking — cited as the primary evidence that click signals influence Google results.
🏢How Search Works — Rigorous Testing
Google. How Search Works.
Brass-SEO uses this for Google's own, careful description of how it watches clicks. Google states that it runs 'live traffic experiments' on a small fraction of users and then looks at 'a very long list of metrics, such as what people click on, how many queries were done, whether queries were abandoned, or how long it took for people to click on a result.' Google frames this as evaluation — measuring whether a change improves results before launch — rather than a per-page ranking dial. Brass-SEO presents this alongside the Navboost testimony to give users the accurate, bounded picture: click data measurably shapes Google's systems, but Google publicly describes it as aggregate evaluation, not a direct per-URL score.
- Examines:
- Google's official description of using click and engagement metrics in live-traffic experiments to evaluate search changes before launch.
- Brass-SEO draws on:
- Google's framing of click metrics as aggregate evaluation — cited when Brass-SEO explains the difference between clicks informing systems and clicks being a direct ranking lever.
🏢United States v. Google — 2024 Liability Ruling
U.S. District Court for D.C., Judge Amit Mehta. August 2024.
Brass-SEO monitors this for adversarial, court-tested confirmation that user-interaction data shapes search quality. On August 5, 2024, Judge Amit Mehta ruled Google an illegal monopolist in general search and text advertising. The court found that Google's roughly $10 billion a year in default-placement payments locked in a scale of user data rivals could not match, and the 2025 remedy ordered Google to share search-index and user-interaction data with qualified competitors. That is a federal court, weighing sworn evidence, treating click and interaction data as a real driver of result quality. Brass-SEO presents this next to the 2023 Navboost testimony so users see the click-signal case is corroborated by evidence, not vendor spin.
- Examines:
- The federal antitrust ruling finding Google a search monopolist, whose findings of fact and 2025 remedy treat user-interaction (click) data as central to search quality and a barrier to entry.
- Brass-SEO draws on:
- The court's treatment of user-interaction data as material to search quality — cited alongside the Navboost testimony when Brass-SEO explains that click behavior genuinely informs ranking systems.
