Every deck about AI traffic engagement this summer has carried the same slide: visitors from ChatGPT and Perplexity stay 67.7% longer than visitors from Google. It is a good stat. It is also an average of a wildly skewed distribution, and the study that produced it quietly publishes the corrective one paragraph later.
SE Ranking ran the analysis across 101,574 websites in 250 countries, using aggregated Google Analytics data over sixteen months from January 2025 through April 2026. It is one of the largest public looks at where AI referral traffic goes and what it does when it lands. The engagement finding was the part that traveled: mean session duration of 9 minutes 19 seconds for AI referrals against 5 minutes 33 seconds for organic search. That is the 67.7%.
Then the same study gives the medians. AI: 2 minutes 24 seconds. Organic: 1 minute 53 seconds. The advantage is still there, but it shrinks to about 27%, and the absolute numbers collapse by roughly three quarters. A visitor who stays two and a half minutes is a normal engaged reader. A visitor who stays nine minutes is a research session, a tab left open, or a bot that GA4 did not filter.
The number everyone quoted
The 67.7% figure spread because it is directionally useful and it flatters a story marketers want to tell. AI traffic volume is still small. In the same dataset, AI platforms accounted for 0.32% of total website traffic in 2026, up from 0.24% in 2025 and 0.02% in 2024. That is a sixteenfold rise in two years and still a rounding error against organic search. When the volume is that thin, quality is the only argument left, so a big engagement multiple does a lot of work in a budget conversation.
| METRIC | AI REFERRALS | ORGANIC SEARCH | GAP |
|---|---|---|---|
| Mean session duration | 9m 19s | 5m 33s | +67.7% |
| Median session duration | 2m 24s | 1m 53s | +27% |
| Mean vs median ratio | 3.9x | 2.9x | AI skews harder |
| Share of total site traffic (2026) | 0.32% | Majority | Volume is not the case |
Look at the third row, because it is the one that matters. In organic search the mean runs about 2.9 times the median. In AI referrals it runs about 3.9 times. AI traffic is not just longer on average, it is more unequal. A small number of very long sessions is doing more of the lifting than in organic, which means the average is a worse summary of the typical visit than it already was.
What the median says about AI traffic engagement
A 27% median lift is not a disappointing result. It is a very good one, and it lines up with what the rest of the evidence base has been saying. AI referrals arrive later in the buying process, land on decision-stage pages, and convert at multiples of organic. We have argued before that the 23x AI conversion figure is one company's n=1 data and that the honest range across seven published studies is closer to 1.4x through 23x depending on how you define a conversion. The engagement data follows the same pattern: real advantage, wildly overstated headline.
It also fits the page-type evidence. When more than 80% of AI referral traffic lands on homepages, product pages, and comparison content rather than blog posts, you would expect somewhat longer sessions, because those are pages people evaluate rather than skim. Our teardown of where AI referral traffic actually lands found exactly that concentration. A 27% lift is what page-type mix plus late-funnel intent should produce. A 68% lift is what page-type mix plus late-funnel intent plus a long tail of outliers produces.
SE Ranking, June 2026: share of AI referral traffic by platform, 101,574 sites.
Platform mix matters here too, because the engines do not send the same visitor. ChatGPT held 74.78% of AI referral traffic in the study, growing 27% year over year. Gemini took 11.56% on 231% growth. Claude sat at 2.62% but grew 320%, and jumped 159% in March 2026 alone. Perplexity slipped, falling from 11.42% of US AI traffic in 2025 to 6.85% in 2026. When one platform supplies three quarters of the sample, the blended engagement number is mostly a ChatGPT number wearing a category label.
Why the mean runs so hot
There are three plausible causes for a mean that sits nearly four times its median, and they call for different responses. Two of them are measurement artifacts you should strip out before reporting. The third is the real thing you want more of, and it is the smallest of the three.
You cannot separate these three from a public benchmark. You can separate them in your own property, and that is the point. Segment AI referrals in GA4, then compare mean and median session duration for that segment against organic. If your ratio looks like the study's 3.9x, you have a long tail worth investigating before you build a narrative on it. Teams running a proper reporting and analytics program should already have the segment; if the medians are missing from the dashboard, that is a one-afternoon fix.
“A mean tells you what the outliers did. A median tells you what your buyer did. Pick the one you are willing to defend under questioning.”
The dataset swap almost nobody noticed
Here is the detail that should change how you cite this study. The headline engagement comparison does not come from the same sixteen-month window as the traffic-share numbers. SE Ranking notes that the engagement section draws on a separate dataset covering January through April 2025. The traffic-share figures run January 2025 through April 2026.
So the widely quoted claim is that AI visitors were 67.7% more engaged during the first four months of 2025, presented alongside platform-share data that is a year newer. In a category where Claude's traffic moved 159% in a single month and Perplexity's US share nearly halved in a year, a fifteen-month lag is not a rounding issue. Nobody is being dishonest. The study says so plainly. It is just that the caveat did not survive the trip into everyone else's slide deck.
This is the same discipline we applied when four independent AI visibility datasets disagreed about basic questions of measurement. The datasets were not wrong. They measured different things over different windows and got reported as if they were interchangeable. AI traffic engagement is now in that phase: plenty of numbers, not enough attention to what each one actually counts.
How to report AI traffic engagement honestly
The fix is not to stop quoting the research. It is to report AI traffic engagement as a pair of numbers with a stated definition, the same way you would report any distribution that skews. Three lines on a dashboard, and the conversation with finance gets easier rather than harder.
Pair that with conversion and pipeline rather than time on site wherever you can. Session duration is a proxy, and a weak one. It is only load-bearing because AI referral volume is too small to produce statistically comfortable conversion counts for most sites, which is precisely the situation where a proxy metric gets over-trusted. If you have the volume, skip engagement and go straight to tying organic and AI-cited traffic to pipeline.
What to do Monday
Open GA4, build the AI referral segment, and pull mean and median session duration for the last 90 days against organic search. Write both numbers down with the date range attached. If your mean-to-median ratio clears 3x, spend an hour on the longest sessions and find out whether they are humans, tabs, or machines. Then rewrite whichever slide currently says 67.7% so it says what your own property says.
After that, stop optimizing for the engagement number and start optimizing for the thing underneath it. AI referrals arrive on comparison, product, and pricing pages because that is what engines cite for buying questions, which is the same reason comparison content earns the largest share of AI citations. Earn more of those citations and both your volume and your engagement improve for the same reason. That is a better use of a quarter than defending an average.
Source for all figures cited: SE Ranking's AI traffic research study, published June 18, 2026, covering 101,574 websites (study).
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.