Clicks as a ranking-related sign have been a topic of debate for over twenty years, though these days most SEOs perceive that clicks usually are not a direct rating issue. The straightforward reality about clicks is that they’re uncooked knowledge and, surprisingly, processed with some similarity to human rater scores.
Clicks Are A Uncooked Sign
The DOJ Antitrust memorandum opinion from September 2025 mentions clicks as a “uncooked sign” that Google makes use of. It additionally categorizes content material and search queries as uncooked alerts. That is essential as a result of a uncooked sign is the lowest-level knowledge level which is processed into greater degree rating alerts or used for coaching a mannequin like RankEmbed and its successor, RankEmbedBERT.
These are thought of uncooked alerts as a result of they’re:
- Straight noticed
- However not but interpreted or used for coaching knowledge
The DOJ doc quotes professor James Allan, who gave skilled testimony on behalf of Google:
“Indicators vary in complexity. There are “uncooked” alerts, just like the variety of clicks, the content material of an online web page, and the phrases inside a question.
…These alerts will be created with easy strategies, reminiscent of counting occurrences (e.g., what number of occasions an online web page was clicked in response to a specific question). Id.
at 2859:3–2860:21 (Allan) (discussing Navboost sign) “
He then contrasts the uncooked alerts with how they’re processed:
“On the different finish of the spectrum are modern deep-learning fashions, that are machine-learning fashions that discern complicated patterns in massive datasets.
Deep fashions discover and exploit patterns in huge knowledge units. They add distinctive capabilities at excessive value.”
Professor Allan explains that “top-level alerts” are used to supply the “ultimate” scores for an online web page, together with recognition and high quality.
Uncooked Indicators Are Information To Be Additional Processed
Navboost is talked about a number of occasions within the September 2025 antitrust doc as recognition knowledge. It’s not talked about within the context of clicks having a rating impact on individal websites.
It’s known as a technique to measure recognition and intent:
“…recognition as measured by person intent and suggestions methods together with Navboost/Glue…”
And elsewhere, within the context of explaining why among the Navboost knowledge is privileged:
“They’re ‘recognition as measured by person intent and suggestions methods together with Navboost/Glue’…”
Within the context of explaining why among the Navboost knowledge is privileged:
“Underneath the proposed treatment, Google should make out there to Certified Rivals …the next datasets:
1. Person-side Information used to construct, create, or function the GLUE statistical mannequin(s);
2. Person-side Information used to coach, construct, or function the RankEmbed mannequin(s); and
3. The Person-side Information used as coaching knowledge for GenAI Fashions utilized in Search or any GenAI Product that can be utilized to entry Search.
Google makes use of the primary two datasets to construct search alerts and the third to coach and refine the fashions underlying AI Overviews and (arguably) the Gemini app.”
Clicks, like human rater scores, are only a uncooked sign that’s used additional up the algorithm chain to coach AI fashions to higher in a position match net pages to queries or to generate a high quality or relevance sign that’s then added to the remainder of the rating alerts by a rating engine or a rank modifier engine.
70 Days Of Search Logs
The DOJ doc makes reference to utilizing 70 days of search logs. However that’s simply eleven phrases in a bigger context.
Right here is the half that’s steadily quoted:
“70 days of search logs plus scores generated by human raters”
I get it, it’s easy and direct. However there may be extra context to it:
“RankEmbed and its later iteration RankEmbedBERT are rating fashions that depend on two predominant sources of knowledge: [Redacted]% of 70 days of search logs plus scores generated by human raters and utilized by Google to measure the standard of natural search outcomes.”
The 70 days of search logs usually are not click on knowledge used for rating functions in Google, AI Mode, or Gemini. It’s knowledge in mixture that’s additional processed with a purpose to practice specialised AI fashions like RankEmbedBERT that in flip rank net pages primarily based on pure language evaluation.
That a part of the DOJ doc doesn’t declare that Google is straight utilizing click on knowledge for rating search outcomes. It’s knowledge, just like the human rater knowledge, that’s utilized by different methods for coaching knowledge or to be additional processed.
What Is Google’s RankEmbed?
RankEmbed is a pure language method to figuring out related paperwork and rating them.
The identical DOJ doc explains:
“The RankEmbed mannequin itself is an AI-based, deep-learning system that has sturdy natural-language understanding. This enables the mannequin to extra effectively determine one of the best paperwork to retrieve, even when a question lacks sure phrases.”
It’s educated on much less knowledge than earlier fashions. The information partially consists of question phrases and net web page pairs:
“…RankEmbed is educated on 1/one centesimal of the information used to coach earlier rating fashions but gives greater high quality search outcomes.
…Among the many underlying coaching knowledge is details about the question, together with the salient phrases that Google has derived from the question, and the resultant net pages.”
That’s coaching knowledge for coaching a mannequin to acknowledge how question phrases are related to net pages.
The identical doc explains:
“The information underlying RankEmbed fashions is a mix of click-and-query knowledge and scoring of net pages by human raters.”
It’s crystal clear that within the context of this particular passage, it’s describing the usage of click on knowledge (and human rater knowledge) to coach AI fashions, to not straight affect rankings.
What About Google’s Click on Rating Patent?
Manner again in 2006 Google filed a patent associated to clicks known as, Modifying search end result rating primarily based on implicit person suggestions. The invention is concerning the mathematical method for making a “measure of relevance” out of the aggregated uncooked knowledge of clicks (plural).
The patent distinguishes between the creation of the sign and the act of rating itself. The “measure of relevance” is output to a rating engine, which then can add it to present rating scores to rank search outcomes for brand spanking new searches.
Right here’s what the patent describes:
“A rating Sub-system can embrace a rank modifier engine that makes use of implicit person suggestions to trigger re-ranking of search outcomes with a purpose to enhance the ultimate rating
introduced to a person of an data retrieval system.Person alternatives of search outcomes (click on knowledge) will be tracked and reworked right into a click on fraction that can be utilized to re-rank future search outcomes.”
That “click on fraction” is a measure of relevance. The invention described within the patent isn’t about monitoring the press; it’s concerning the mathematical measure (the press fraction) that outcomes from combining all these particular person clicks collectively. That features the Brief Click on, Medium Click on, Lengthy Click on, and the Final Click on.
Technically, it’s known as the LCIC (Lengthy Click on divided by Clicks) Fraction. It’s “clicks” plural as a result of it’s making choices primarily based on the sums of many clicks (mixture), not the person click on.
That click on fraction is an mixture as a result of:
- Summation:
The “first quantity” used for rating is the sum of all these particular person weighted clicks for a particular query-document pair. - Normalization:
It takes that sum and divides it by the whole rely of all clicks (the “second quantity”). - Statistical Smoothing:
The system applies “smoothing components” to this mixture quantity to make sure that a single click on on a “uncommon” question doesn’t unfairly skew the outcomes, particularly for spammers.
That 2006 patent describes it’s weighting method like this:
“A base LCC click on fraction will be outlined as:
LCC_BASE=[#WC(Q,D)]/[#C(Q,D)+S0)
the place iWC(Q.D) is the sum of weighted clicks for a question URL…pair, iC(Q.D) is the whole variety of clicks (ordinal rely, not weighted) for the query-URL pair, and S0 is a smoothing issue.”
That method describes summing and dividing the information from many customers to create a single rating for a doc. The “query-URL” pair is a “bucket” of knowledge that shops the press habits of each person who ever typed that particular question and clicked that particular search end result. The smoothing issue is the anti-spam half that features not counting single clicks on uncommon search queries.
Even approach again in 2006, clicks is simply uncooked knowledge that’s reworked additional up the chain throughout a number of levels of aggregation, right into a statistical measure of relevance earlier than it ever reaches the rating stage. On this patent, the clicks themselves usually are not rating components that straight affect whether or not a website is ranked or not. They had been utilized in mixture as a measure of relevance, which in flip was fed into one other engine for rating.
By the point the knowledge reaches the rating engine, the uncooked knowledge has been reworked from particular person person actions into an mixture measure of relevance.
- Enthusiastic about clicks in relation to rating isn’t so simple as clicks drive search rankings.
- Clicks are simply uncooked knowledge.
- Clicks are used to coach AI methods like RankEmbedBert.
- Clicks usually are not straight influencing search outcomes. They’ve all the time been uncooked knowledge, the start line for methods that use the information in mixture to create a sign that’s then blended into rating resolution making methods at Google.
- So sure, like human rater knowledge, uncooked knowledge is processed to create a sign or to coach AI methods.
Learn the DOJ memorandum in PDF type right here.
Examine 4 analysis papers about CTR.
Learn the 2006 Google patent, Modifying search end result rating primarily based on implicit person suggestions.
Featured Picture by Shutterstock/Carkhe
