Google Click Signals and SEO: The Facts

Clicks have been a long-running topic in SEO discussions, often treated as either an important ranking factor or dismissed entirely depending on who you ask. Over time, the general understanding within the industry has become clearer: clicks are not a direct ranking factor in Google Search. Instead, they should be seen as raw behavioural data that feeds into larger systems, rather than something that directly determines where a page appears in results.

This distinction is important because it changes how clicks should be interpreted. Rather than acting as a simple “vote” for rankings, clicks are part of a much broader data ecosystem that helps search systems understand user behaviour at scale. On their own, they do not directly move a page up or down the rankings.

Recent references in the September 2025 US Department of Justice antitrust documentation describe clicks as a type of “raw signal” used by Google. This category also includes other basic inputs such as search queries and webpage content. A raw signal is essentially the most basic level of data collected by a system before it is processed, interpreted, or used in machine learning models.

These raw signals are important because they represent direct user interactions. However, they are not immediately meaningful in a ranking sense. They first need to be aggregated, cleaned, and processed before they can contribute to anything more complex, such as training models or generating higher-level ranking signals.

Within the same documentation, expert testimony explains that signals exist on a spectrum. At the simplest level are raw inputs like click counts or page content. These can be collected through straightforward methods, such as tracking how often a result is clicked for a particular query. At the more advanced end are deep learning systems that analyse large volumes of data to identify patterns, relationships, and relevance signals across millions of searches.

This distinction matters because it highlights how far removed raw clicks are from actual ranking decisions. While they are part of the data pipeline, they are not themselves ranking instructions. Instead, they contribute to systems that help build those instructions.

The documentation also references systems such as Navboost, which is associated with measuring popularity and user feedback. In this context, it is described more as an aggregated measure of user behaviour and intent rather than a mechanism that directly applies ranking changes based on individual clicks. It helps identify broader trends in how users interact with search results rather than acting on single interactions.

In addition to this, the same legal materials explain how Google uses large datasets to train machine learning systems that support both traditional search and AI-driven features. These datasets include user-side information used to build models such as RankEmbed, as well as data used to improve generative AI systems. Importantly, this data is used for training purposes rather than being applied directly as live ranking signals.

This is where clicks sit in the overall structure. They are part of the training material used to help systems learn patterns between queries and webpages. Similar to human rater scores, they provide input that helps models improve over time, rather than directly influencing search results in real time.

A commonly referenced example in this discussion is the use of “70 days of search logs”. While this figure is often quoted on its own, it is only a small part of a much larger dataset used for training ranking systems. These logs are combined with other inputs, including human quality evaluations, to help improve models like RankEmbed and RankEmbedBERT.

However, it is important to understand that this does not mean click data is directly applied to live rankings. Instead, it is processed in aggregate and used to train systems that later generate ranking signals indirectly. The clicks themselves are not acting as ranking factors, but rather as input data that helps improve model accuracy.

RankEmbed is a particularly relevant example of how this works in practice. It is a deep learning system designed to improve how search understands language and relevance. It helps match queries with documents even when exact keyword matches are not present, relying instead on patterns learned from large datasets.

The training process for RankEmbed includes a combination of query data, click behaviour, and human rater assessments. This allows the model to learn which types of results tend to satisfy user intent, based on aggregated historical behaviour rather than individual actions.

In practical terms, this means clicks contribute to shaping how the system learns, but not how it reacts in real time. The model does not adjust rankings based on a single click or even a small group of clicks. Instead, it identifies patterns across large datasets and uses those patterns to improve relevance over time.

This approach is also reflected in earlier Google patent filings related to click data. One example describes how clicks are aggregated into what is known as a “click fraction”, which is then used as part of a relevance measure. This involves combining multiple user interactions for a specific query and webpage pair, applying weighting, and then normalising the results.

The key idea in this system is aggregation. Individual clicks are not meaningful on their own. Instead, they are grouped together across many users to form a statistical representation of behaviour. This aggregated measure is then passed through additional systems that may influence ranking indirectly.

The process also includes safeguards such as smoothing factors, which help prevent skewed results caused by unusual or low-volume interactions. For example, a single click on a rare query would not disproportionately influence results. Instead, the system looks for consistent patterns across many interactions.

Even in these earlier models, the focus is not on individual clicks but on long-term behavioural trends. These trends are used to create signals that can be fed into ranking systems, but only after significant processing and aggregation.

By the time any of this data reaches a ranking system, it has already been transformed. Raw clicks have been converted into structured signals, statistical measures, or training inputs for machine learning models. At no point is a single click acting as a direct ranking instruction.

This leads to a clearer understanding of how clicks function within search systems. They are not ranking signals in the traditional sense, but rather foundational data points that contribute to broader systems responsible for generating those signals.

They are similar in structure to human rater data in that both are used to help train and refine systems, rather than directly influencing live rankings. Human raters assess quality, while click data reflects user behaviour. Both are valuable, but neither operates as a direct ranking switch.

The key takeaway is that clicks exist at the bottom of the data hierarchy. They are collected as raw signals, aggregated into meaningful patterns, and then used to train systems that ultimately help determine rankings. The process is layered, and each stage adds interpretation and structure.

In simple terms, clicks do not directly control rankings. Instead, they contribute to the systems that help improve ranking accuracy over time. They are part of the learning process, not the final decision-making process.

Understanding this distinction is important for SEO because it shifts the focus away from chasing individual behavioural signals and towards improving overall content quality, relevance, and user satisfaction. Clicks matter, but not in the simplistic way they are often described.

Ultimately, they are one part of a much larger ecosystem of signals, models, and systems that work together to deliver search results.