Does Google use clicks to rank?

Whether Google uses clicks to rank is one of the longest-running arguments in SEO, and most of the heat comes from the two sides talking past each other. Google’s public denials are about raw click-through rate: it does not take a page’s CTR, compare it to its neighbours, and promote the one that gets clicked more. That is true. It is also not the whole story. Clicks are collected, aggregated, classified by whether they stuck, and fed into both the relevance calculation and a re-ranking step. The distinction between “raw CTR as a direct factor” and “click data in the system” is where the confusion lives.

This is a Google-specific question. Other engines have their own retrieval systems, and the evidence below is about Google.

What does Google actually say?

Google’s representatives have consistently denied that click-through rate is a direct ranking factor. The careful phrasing matters: they deny that raw CTR is used the way practitioners often imagine, as a dial that rewards results for being clicked more than expected. Taken literally, this is not a lie, and it is not an admission either. The public debate treats “do clicks matter?” as a yes/no question, while Google is answering a narrower one about a specific mechanism.

The antitrust disclosures support that narrower denial rather than overturning it. The September 2025 remedies opinion lists clicks alongside page content, human rater scores, and query terms as a raw signal, the lowest-level kind of data point, which is then aggregated and processed into the higher-level signals that ranking systems actually consume.¹ A raw click is not a ranking factor. What it becomes several steps later is.

So the useful reframe is not “is Google lying?” but “if not raw CTR, then what?”

What the evidence actually shows

Three bodies of evidence converge, and they are not equally strong. Sworn testimony outranks a document leak, which outranks a conference disclosure. Take them in that order.

The US Department of Justice antitrust case. Interviews with Google engineers Pandu Nayak and HJ Kim, released as trial exhibits in the 2025 remedies hearing, describe the core relevance calculation directly. Topicality, Google’s judgement of how well a document matches a query, is built from what they called ABC signals: Anchors (links pointing at the page), Body (the terms in the document itself), and Clicks, defined as how long a user stayed on the clicked page before returning to the results.² These combine into a topicality score that sits alongside a separate quality score.

The significance of that is easy to miss. Clicks are not a bolt-on adjustment applied to a finished ranking; they are one of the three components of the relevance calculation itself.

Two details from the same exhibits cut against the tidier version of this story that circulates in SEO. Ex-Googler Eric Lehman, asked whether NavBoost trains on 13 months of user data, testified that “trains” is misleading: “NavBoost is not a machine learning system. It’s just a big table. It says for this search query, this document got two clicks. For this query, this document got three clicks.”² And the exhibits record that Google “avoids simply predicting clicks”, because predicted clicks are easily manipulated and do not reliably measure whether a user was served well.² Whatever is happening, it is closer to counting and classifying observed clicks than to modelling a click rate.

The 2024 Content Warehouse leak exposed thousands of internal ranking-related attributes, including a module of click and impression signals.³ The named fields are the useful part: good clicks, bad clicks, last longest clicks, and unsquashed variants of each, alongside a squashing function whose documented purpose is to stop any one large signal dominating the others.⁴ That is a system classifying clicks by quality and normalising them against manipulation, not one counting them.

A late-2024 endpoint exploit disclosed by Mark Williams-Cook of Candour surfaced roughly 2TB of data covering more than 90 million queries and over 2,000 internal properties, reported to Google’s Vulnerability Reward Programme for a $13,337 payout.⁵ His reading of the click question was that Google factors in how likely it thinks someone is to click a result, an inference he drew partly by analogy with the estimated click-through rate Google Ads Keyword Planner reports.⁶ Treat that as one practitioner’s interpretation of the data rather than a documented mechanism, particularly given the later exhibits saying Google avoids predicting clicks.

The through-line is consistent. There is no “more clicks equals higher rank” dial. There is a large aggregation of observed click behaviour, sorted by whether the click satisfied the searcher, feeding both the relevance score and a re-ranking step. NavBoost is described in the disclosures as the re-ranking system built on that data, drawing on roughly the last 13 months of clicks, segmented by location and device.²

What did the antitrust remedies confirm?

The clearest public confirmation that this data matters is not a leak at all: it is what the court ordered Google to hand over.

The September 2025 remedies ruling requires Google to share portions of its search index and its user interaction data, though not its ads data, with “qualified competitors” at marginal cost.⁷ The opinion names the datasets specifically, including the user-side data used to build and operate the Glue statistical models, the user-side data used to train the RankEmbed models, and the user-side data used to train the generative models behind AI features in Search.¹ The same document refers to NavBoost and Glue as measures of “popularity as measured by user intent and feedback systems”.

A court identifying user-interaction data as the asset rivals need in order to compete settles the direction of the argument, even though it settles nothing about weighting. Whatever clicks are doing inside Google’s systems, they are doing enough of it to be treated as a competitive moat.

Why is there no universal CTR benchmark?

If Google compared your click-through rate against a fixed target for your position, you could publish that target. Nothing in the disclosures describes such a number, and the structure of the data explains why.

The click data is aggregated per query and document, and segmented by location and device.² The comparison that exists is between the results competing for one query in one segment, not between your CTR and a cross-site average. Add the eight query semantic classes the endpoint exploit surfaced, Short Fact, Boolean, Definition, Instruction, Reason, Comparison, Consequence, and Other, and what a satisfying result looks like varies query by query.⁶

So the same CTR can be good or bad depending on what it is sitting next to:

• On a fast-moving news query, where users scan and click readily, a 5% click-through rate may put you behind every other result on the page.
• On a slow evergreen query, the same 5% might be the strongest performance in the set.

This is why comparing your click-through rate to an industry-average table tends to mislead. The nearest usable proxy is your own Search Console data over time for a given query, not a benchmark chart.

Does clickbait work, then?

Not durably, and the leaked field names are the reason. A system that records good clicks, bad clicks, and which result held the longest click in a session is measuring whether the click resolved the query, not whether it happened.⁴ A headline engineered to win the click without delivering on it can lift interaction briefly and lose ground as the return-to-results behaviour accumulates.

Dwell time is often dismissed as folklore, but the testimony defines the click component of topicality as how long a user stayed on the page before returning to the results.² The long click is the substance of the click signal, not a separate theory alongside it.

What is not supported is the scaffolding built around it: a dwell threshold to hit, a number of seconds to target, or the idea that Google temporarily promotes a result to “test” it. The mechanism is evidenced; the tactics devised to game it are not. The squashing function exists precisely so that a burst of manufactured clicks cannot run away with the signal.⁴

What the exploit revealed beyond clicks

The same endpoint exploit surfaced related findings that bear on other common assumptions. All of these are strong evidence rather than confirmed live weightings, and the field names describe internal signals, not published ranking rules.

• A site quality score. A per-subdomain quality value on a 0 to 1 scale. Sites below roughly 0.4 appeared ineligible for features like featured snippets and People Also Ask.⁵ This undercuts the “high third-party Domain Authority is enough” assumption: the relevant quality signal is Google’s own, built from inputs like brand-search demand and how often a site is selected when it is not the top result. The antitrust exhibits describe an equivalent document quality score, Q*, that Google engineers called “incredibly important”.²
• Consensus scoring. Google appears to count how many passages in a piece of content agree with, contradict, or stay neutral to a general consensus, with a particular role for debunking queries.⁶

The honest summary

Does Google use clicks to rank? The most accurate answer is: not as raw click-through rate, but yes as aggregated click data, classified by whether the click satisfied the searcher, feeding both the relevance score and a re-ranking step. Google’s denials and the practitioner belief have been arguing about different things.

Two cautions keep this honest. Leaked and exploited field names confirm that signals exist; they do not confirm exactly how, or how heavily, each is weighted in live ranking. And this is Google’s system specifically. The takeaway for a working SEO is not to game a CTR number, which no fixed benchmark supports, but to earn the click with an accurate, compelling result and then deliver on it, so the follow-through runs in your favour rather than against you.

Footnotes

1. The facts about Google click signals, rankings, and SEO — Search Engine Journal ↩ ↩²

2. The ABCs of Google ranking signals: what top search engineers revealed — Search Engine Land ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. An anonymous source shared thousands of leaked Google Search API documents with me — SparkToro ↩

4. Secrets from the algorithm: Google Search’s internal engineering documentation has leaked — iPullRank ↩ ↩² ↩³

5. Google’s site quality scoring system revealed through endpoint discovery — PPC Land ↩ ↩²

6. Exploit reveals how and why Google ranks content — Search Engine Land ↩ ↩² ↩³

7. Court issues remedies ruling in United States v. Google search case — Hughes Hubbard & Reed ↩