Every year, someone publishes a new ranking factor study. It pulls millions of search results, measures something like backlinks or page speed, and reports a correlation with rank. Then the headline turns that number into a rule: do this, rank higher. Here is the honest verdict on that kind of study: mixed. Most Google ranking factor studies can be a reasonably accurate account of what was associated with rank inside their own sample. They are weak evidence that the measured thing causes rankings, and they should never be read as a checklist. Keep those two statements separate and you can actually use these studies instead of either trusting them blindly or throwing them out.
That distinction matters because of how often it gets collapsed. Moz's Rand Fishkin laid out the correlation vs causation SEO problem directly in a January 2018 explainer: correlation is a useful hint for what to investigate next, not proof of what causes what. Google's own ranking systems guide, last updated December 2025, describes many systems and signals working together, not a simple formula that a correlation study can reverse-engineer from the outside. The studies are not worthless. They are also not the thing most people treat them as.
What a ranking factor correlation study actually measures
A typical study works like this. Researchers pull a large set of Google results, measure some characteristic of each result's page or domain, then check whether that measurement moves together with rank position. The output is an average relationship inside the sample they collected. It is not a window into Google's private scoring mechanism.
Moz's 2015 methodology is a clear example of what this actually looks like in practice. It started with 17,600 suggested queries pulled from 22 Google AdWords categories, trimmed duplicates down to 16,521, and pulled the top 50 English-language U.S. results for each one in early May 2015. The team calculated a Spearman correlation for each query and then averaged those correlations across the whole set. That is a specific, bounded exercise: one country, one language, one week in 2015, a defined list of queries. It tells you what moved together with rank inside that frame. It does not tell you what Google's algorithm actually weighs, and it was never built to.
This is the gap that gets lost in every recap post. A correlation describes an association. A ranking factor, in the sense most readers mean it, is a causal claim: changing this thing will change your rank. Those are two different statements, and a study built to answer the first one cannot answer the second just because the sample size is large. That question is also distinct from asking which confirmed Google ranking factors are proven versus speculation, a separate evidence question that starts from primary sources rather than correlation alone.
Why a correlation can mean several different things
A reported association between a page characteristic and rank is consistent with more than one explanation, and a correlation study usually cannot tell you which one is true.
The measured feature could genuinely affect rank. But rank could also affect the measured feature instead: a page that already ranks well gets more visits, more links, and more engagement simply because more people see it, not because those things caused the ranking in the first place. Fishkin's 2018 piece uses exactly this example with search visits, and it is the same logic that applies to backlinks, social shares, and dwell time. A third variable, like the quality of the underlying content or the authority of the site publishing it, could be driving both the measured feature and the rank at once. Or the pattern could simply be incidental to that particular sample and not show up again in a different one.
None of these explanations can be ruled out by a correlation coefficient alone. Ruling them out takes a different kind of study design, one built to test cause and not just measure association, and most published Google ranking factor studies were never built that way. That is the real shape of the correlation vs causation SEO debate: not a fight over whether the data is real, but over what kind of question the data can answer. It is not a knock on the researchers. It is a description of what the method can and cannot do.
Where the methodology actually breaks down
A few specific weaknesses in SEO studies methodology show up again and again in this genre, and knowing them turns you from someone who reads the headline into someone who can evaluate the study underneath it.
Many studies sample only pages that already rank, often just the top ten or twenty results. Roger Montti's April 2020 critique in Search Engine Journal makes this point directly: conclusions drawn only from top-ranked pages can fail to describe how ranking actually works across the much larger set of pages that exist, including the ones that never made the sample at all. This is a version of the same problem survivorship bias describes elsewhere: when you only look at the winners, you cannot see what separated them from everyone who did not win, because the people who did not win were never in your dataset to begin with.
Search intent muddies the comparison too. Montti uses the word "jaguar" as his example: it could point to a car, an animal, a sports team, or breaking news, and each of those intents pulls a different kind of result into the top of the page. A study comparing rank positions across a large query set is often comparing pages that are answering completely different questions, not pages competing on the same characteristics for the same goal.
Then there is the proxy problem. A study can only measure what it has access to, and what it has access to is not always the same as what the headline claims to measure. The Backlinko study of 11.8 million results, run with Ahrefs as the main data partner, used Alexa data from roughly 10,000 domains to estimate page speed, not a direct measurement of each individual page's load time. This is a different kind of evidence than Google's patents revealing possible mechanisms, which describe what a company could build rather than what it actually weighs in production. Its time-on-site analysis worked the same way, pulling a domain-level number to stand in for something that varies page by page. That is a reasonable thing to do when nothing better is available, but a domain-level proxy is not the same thing as the page-level signal the article's headline implies it measured.
Exclusions change what a correlation is actually describing, and they are easy to miss in a skim. That same Backlinko study found that about 95% of the pages in its dataset had zero backlinks, and the authors excluded those zero-backlink pages from the backlink analysis because they said the zeroes were skewing the result. That might be a reasonable call for that specific comparison. It also means the backlink finding describes a smaller, different population than the headline sample size suggests, and that detail rarely survives into the recap post that cites the number.
Finally, every one of these studies is a snapshot, pinned to a specific country, device type, and collection date. The Moz 2021 study, from Tom Capper, sampled the first 20 organic results for 10,000 keywords from a single suburban U.S. location. The Backlinko data was pulled in the first week of January 2020. Google's systems have changed since both of those dates, and a pattern observed in one snapshot is not automatically still true, let alone true in a different market or on a different device. This is also why a traffic drop after an update needs its own confirmation step instead of blaming whichever factor an old correlation study pointed at.
What the primary sources actually say
It helps to separate two kinds of source here, because they are answering different questions. Google's own ranking systems guide describes the existence of many systems and signals operating together, including ones built for specific contexts like freshness or local results, and it says the company regularly tests and adjusts those systems. That is Google telling you the shape of the thing. It is not a public formula, and it was never meant to be one. Google's own public statements have shifted over time too, and Google's top 3 ranking factors claim is a good example of how an offhand remark hardens into industry gospel long after the context it was said in has changed.
The correlation studies are a separate kind of source: third-party researchers measuring what they can observe from the outside and reporting what moved together. Moz's own explainer is notably careful about this boundary, treating correlation as a lead worth investigating rather than a conclusion. The Backlinko methodology document says plainly that correlation does not always indicate causation and that the study cannot assign causation to any single factor, a disclosure that is easy to find in the methods section and easy to lose by the time the number reaches a listicle. Montti's Search Engine Journal piece takes the strongest stance of the three, arguing the industry should largely set this kind of study aside. That is his stated opinion, not a universal verdict the other sources share, and it is worth reading as one voice in a real disagreement rather than the final word.
How to read one of these studies without getting fooled
Reading SEO studies methodology closely, rather than skimming the headline number, is most of the work here, and a short checklist gets you most of the way there. Ask what the study actually claims: does it say "correlates with," or did a headline somewhere upstream quietly turn that into "causes" or "is a ranking factor"? Check who ran it and who supplied the data, since a commercial publisher can still produce a useful study, but its incentives should be visible rather than hidden. Look at what was sampled: which countries, how many queries, what rank range, and whether non-ranking pages were included at all or left out of the frame entirely. Separate direct measurements from proxies, and watch for subsets, since a secondary analysis inside a big study often runs on a much smaller slice of the data than the headline number implies. Note the collection date and ask whether the system being described has changed since then. And always ask what alternative explanation the study ruled out, because a correlation that has not ruled out reverse causation or a shared third cause has not actually ruled out much.
A single sentence captures what any of these studies can honestly support: this study found an association in its sampled results, and it does not by itself show that the measured feature caused rank or that changing it will move your own rankings.
Specific named factors deserve the same scrutiny individually. Our piece on whether backlinks still matter for rankings applies this checklist to the backlink correlation literature specifically. The same goes for click-through rate as a ranking signal, where leaked documents add a different kind of evidence than any correlation study can. Dwell time and pogo-sticking evidence gets the same treatment elsewhere on the site. So does evidence on social signals as a ranking factor, which is one of the clearest examples of a correlation that faded as the web changed.
The verdict, and what would change it
The honest grade stays mixed for Google ranking factor studies generally, and it should stay mixed for any individual study that has not addressed the gaps above. A correlation study can describe a real pattern in a real dataset, and that pattern can be a legitimate starting point for a hypothesis worth testing further. What would move a claim from mixed to confirmed is a design built to answer the causal question directly: a controlled or quasi-experimental setup, a longitudinal design with a real comparison group, or convergent results from independently replicated studies that are not all leaning on the same data provider or the same sampling frame. Short of that, treat every correlation study, including the ones cited in this piece, as a lead rather than a verdict.
If you are the one producing content that cites studies like these, the same discipline applies in reverse: a claim is only as strong as the source behind it, and a reader who checks your citation and finds it does not hold up stops trusting the rest of the page too. DeepSmith's writing pipeline builds citation checks and sourcing into article production itself, so claims trace back to a named, dated source instead of a recycled stat nobody re-verified.
