What is correlation vs causation in SEO?
Correlation vs causation in SEO is the difference between a feature that shows up on pages that rank well and a feature that makes pages rank well. A correlation says two things travel together. Causation says changing one changes the other.
A real example: pages with the current year in the title are about 1.9 times more likely to be quoted by Google's AI Overviews than pages on the same search without one [2]. That is a correlation. Compare only pages of the same age and the advantage falls to 1.05 times, with a range that includes no difference at all [2]. The year was traveling with something else, and that something else was doing the work.
Most published "ranking factor" findings are correlations. They are useful, but they answer "what do ranking pages look like?" and not "what should I change?"
Why do most SEO ranking studies only show correlation?
Most SEO ranking studies only show correlation because they observe pages that already rank instead of changing a page and watching what happens. Collecting the top 20 results for a thousand keywords is cheap. Running a controlled change across hundreds of pages, and waiting for Google to recrawl them, is slow and expensive.
Observation has a built-in weakness: the pages in each group were never chosen at random. Pages carrying a feature differ from pages without it in many ways at once. Google says its ranking systems "look at many factors and signals" [4], so any single feature a study measures sits alongside dozens it does not.
Search Insider labels its own studies plainly. The year-in-title ranking study is described as observational, because no title was changed to see what happened [1].
What is a confounder in SEO research?
A confounder in SEO research is a third factor that drives both the feature being studied and the ranking or citation outcome, creating a link between them that is not cause and effect. Page age is the most common one.
In Search Insider's year-in-title study, pages with the current year in the title had a median age of 93 days, while pages with no year had a median age of 449 days [1]. Newer pages carry the current year, and Google runs freshness systems that favor newer content for queries where it is expected [4]. So a raw comparison of year-titled and undated pages partly measures freshness and labels it "the year." Within the results Google showed a date for, holding page age constant removed roughly a third of the measured ranking advantage [1].
Page type is a second confounder. In the AI Overview follow-up, year-titled pages were 94.6% editorial articles, against 73.1% of pages without a year, and the no-year group held far more videos and shop listings, which were quoted less often whatever their titles said [2].
Other common confounders in SEO studies:
- Ranking position. AI Overview citations track what already ranks [3], so any feature common among top results looks like a citation factor.
- Site strength. Large, well-linked sites tend to adopt every best practice at once, so each practice inherits their rankings.
- Query mix. Comparing pages from easy searches with pages from hard ones measures the keyword, not the page. Comparing pages within the same results page avoids this [1].
- Industry habits. A year in the title was 3.2 times more common in marketing and SEO than in beauty [1]. A feature that maps onto publishing culture is a convention first.
What is reverse causation in SEO?
Reverse causation in SEO is when the ranking produces the feature, rather than the feature producing the ranking. Backlinks are the usual case. Pages that rank first are seen by more writers, who then link to what they found. A study showing that top pages have more links cannot tell, by itself, how much of that is links lifting rankings and how much is rankings attracting links.
Traffic-based metrics have the same problem. Engagement, branded searches and social shares all grow once a page ranks, so a page's rank can explain them as easily as they explain the rank.
What is the difference between correlation and causation in SEO?
Correlation in SEO tells you what ranking pages have in common, while causation tells you what will happen if you change your own page. The table sets out how each is established and what it is good for.
| Correlation | Causation | |
|---|---|---|
| Question answered | What do ranking pages look like? | What happens if I change this? |
| How it is measured | Observing pages that already rank | Changing some pages and not others, then comparing |
| Typical evidence | Ranking factor studies, SERP surveys | Holdout tests, SEO split tests, Google's live experiments |
| Main risk | Confounders and reverse causation | Too few pages, too short a test, noise |
| Good for | Spotting norms, ruling ideas out, choosing what to test | Deciding whether a change is worth rolling out |
Google itself relies on the causal approach. Its live traffic experiments switch a change on for a small share of people, usually starting at 0.1%, and compare that group with a control group that did not get the change [5]. In 2023 Google reported 16,871 live traffic experiments and 4,781 launches [5].
How do you test whether an SEO factor causes better rankings?
To test whether an SEO factor causes better rankings, change the factor on a randomly chosen set of pages, leave a matched set unchanged, and compare how both groups move over the same period. Random assignment is what breaks the link with confounders, because age, site strength and query difficulty end up spread evenly across both groups. Randomized experiments of this kind are the basis of A/B testing [6].
When an experiment is not possible, an observational study can still get closer to the truth by matching. The AI Overview study shows the method, adding one control at a time [2]:
| Comparing year-titled pages with undated pages that share | Times more likely to be quoted | Range |
|---|---|---|
| The same search | 1.91x | 1.65 to 2.22 |
| and a similar ranking position | 1.74x | 1.47 to 2.06 |
| and similar freshness | 1.46x | 1.22 to 1.75 |
| and the same page age | 1.05x | 0.81 to 1.36 |
Each row removes a difference between the groups that has nothing to do with the title. By the last row, the advantage is gone. Matching only removes the confounders a study can measure, so a matched result is stronger evidence than a raw one, but still weaker than a controlled test.
How has SEO research dealt with correlation vs causation?
SEO research has moved slowly from raw correlations toward matched comparisons and controlled tests, while Google has used controlled experiments on its own results for over two decades. The dates below mark that shift.
2000: Google runs its first A/B test
Google engineers ran their first A/B test in 2000, to find the best number of results to show on a page [6]. Controlled experiments, not correlations, became the way Google decides what ships.
2011: experiments at scale
Google ran more than 7,000 different A/B tests in 2011 [6].
2023: tens of thousands of live experiments
Google reported 16,871 live traffic experiments, 124,942 side-by-side experiments with raters and 4,781 launches in 2023 [5]. Each live experiment compares a test group against a control group [5].
September 2026: one feature, two answers
Search Insider's year-in-title study found the current year worth about 0.62 positions within the same results page, with roughly a third of that explained by page age in dated results [1]. A week later, the AI Overview follow-up found a raw citation advantage that disappeared once page age was matched [2]. The same feature on the same pages was a modest ranking correlate and no measurable citation cause.
Are correlation studies useless for SEO?
No. Correlation studies are useful for SEO when they are read as descriptions rather than instructions. They show what is normal in a results page, they can rule tactics out (if a feature is rare among ranking pages, it is not required), and they point to what is worth testing properly.
A correlation study is also strongest when it builds in its own baseline. Search Insider's stop word study searched 80 keywords twice each, seconds apart, to measure how much a results page changes on its own (about 1 in 10 websites), and read every deletion and swap against that figure [3]. A difference smaller than the background churn is not evidence of anything.
What a correlation study cannot do is tell you that adding a feature to your page will move it. That gap is where most ranking factor advice goes wrong.
How should you read an SEO study's claims about causation?
Read an SEO study's causation claims by checking how it compared pages, not by how large its dataset is. A large observational dataset with no controls repeats the confounders at scale. Before acting on a finding, check:
- Was anything changed? If no page was edited, the result is a correlation, whatever the headline says.
- Were pages compared within the same results page? Comparing across different keywords mixes in query difficulty.
- Were the obvious confounders held constant? Page age, ranking position, page type and site strength are the usual suspects.
- Is there a baseline? A study should show how much results vary with nothing changed. See statistical significance in SERP testing.
- Is the effect big enough to matter? Half a ranking position can be real and still not worth a sitewide change [1].
- Does the study say what it cannot show? A credible study lists its limits, including that it is observational.
When a finding passes those checks and still matters to your site, the next step is not to roll it out everywhere. It is to run a holdout test on your own pages and measure the incremental difference.