Because they measured different populations over different periods using different methods of identifying what counts as an AI referral, and none of those choices is wrong.
In October 2026 two benchmarks landed within days of each other. Both were careful. Both published their methods. They are roughly three and a half times apart.
The two figures
Marie Haynes analysed twelve months of GA4 data across 100 sites, more than 145 million sessions. ChatGPT came to 0.62 percent of Google organic volume. All assistants together came to 0.73 percent.
SEO Works analysed 95 GA4 properties from September 2025 to August 2026. AI referral sessions were 72,095 against 2,825,594 organic, which is 2.6 percent of organic volume.
Both describe a small channel. One describes it as roughly a fortieth of organic and the other as roughly a hundred and fortieth.
Three reasons the numbers separate
Eligibility thresholds change the population. SEO Works required at least 100 AI referral sessions and 1,000 organic sessions over the period for a site to qualify. Of 95 properties, 66 met the volume thresholds and 63 had recorded organic conversions. A threshold like that removes sites with almost no AI traffic, which raises the average of what remains. Haynes’s 100 sites were described as randomly sampled.
The period is not identical. SEO Works covers September 2025 to August 2026. Haynes covers twelve months without the window being specified in the published piece. Haynes also reports that ChatGPT referral growth plateaued around August 2025, so where a twelve month window starts can matter a great deal.
Identification is a judgement call. Google’s own AI Assistant channel grouping only exists from June 2026. Any analysis covering the period before that had to identify AI referrals another way. SEO Works matched referring domains with a regex and said so. Which domains go in that pattern, and whether in-app browser referrers are caught, changes the count.
Aggregation method changes the result. SEO Works calculated percentages from summed sessions and conversions per group rather than averaging per-site rates. Those two approaches can produce materially different figures from identical data, because averaging per-site rates gives a tiny site the same weight as a large one.
Conversion figures should not be stacked at all
This is where comparison goes wrong most often.
Haynes reports ChatGPT referrals converting at 13.8 percent against Google organic at 3.02 percent, roughly four and a half times. SEO Works reports AI referrals at 1.18 times the pooled organic rate, 7.18 percent against 6.08 percent, with a median of 1.21 times.
Those are not two estimates of one quantity. Haynes compares ChatGPT specifically against Google organic. SEO Works compares a combined AI-referral category against organic. Different numerators, different baselines. Reading them as a disagreement about one thing misreads both.
There is a further limit underneath both. Haynes’s conversion comparison rests on the 55 of her 100 sites that had tracked conversions in GA4. The other 45 could not contribute conversion evidence through the tracking she examined, though they may measure outcomes in a CRM or offline.
What the researchers say their instruments cannot see
The most useful caution in this material is not in anyone’s rebuttal. It is in the research.
SEO Works states that their measured AI contribution is a floor rather than a ceiling. Clicks from Google’s AI Overviews arrive as ordinary organic sessions and are excluded. Referrals from in-app browsers and stripped referrers go untracked. So does the reader who asks ChatGPT or Perplexity about a category, then searches the brand by name and arrives as organic or direct.
They also note that GA4 counts key events rather than people, so conversion figures are upper bounds as an estimate of converting people, while adding that the AI-to-organic ratio is unaffected because both channels use the same definition.
Note what makes their floor claim legitimate. It is not a general rule about referral data. It follows from the specific exclusions they listed, each of which removes AI sessions and none of which adds any. A limitation only has a direction when something concrete establishes the direction, and most limitations in this category do not have one. Missing observations can move a percentage either way depending on what went missing.
How to use conflicting benchmarks
Three practical rules.
Read the thresholds before the headline. A benchmark that filtered for minimum volume is describing a different population than one that did not.
Never compare two conversion multiples without checking their numerators and baselines. Most published comparisons in this category fail that test.
Treat the spread as the finding. Two careful studies several times apart tell you that your own distribution is the only one that describes you, and that a plan which only works at one end of the range is not a plan.
What your own numbers would have to show
The useful version of this takes about twenty minutes and does not require either study.
Open your analytics for the last twelve months. Pull sessions from AI assistant referrers and sessions from organic search, and divide one by the other. That single figure tells you where in the published range your business actually sits, which is the only version of the number that can inform a decision about your business.
Then note the same limitations the researchers did. Your figure excludes AI Overview clicks, which are sitting inside your organic number, and it excludes anyone who read an answer from ChatGPT, Claude, Gemini or Perplexity and then searched your brand by name, because they are counted somewhere else entirely. Both of those omissions remove AI sessions without adding any, so your figure understates on those two counts specifically. That is as far as the reasoning goes, and it is not a claim that every limitation in your analytics points the same way.
If your share comes back well below the published range, the acquisition question is settled for you. If it comes back above it, you have something the aggregates cannot explain, which is worth understanding before anyone budgets against it.
Frequently asked questions
Which AI traffic study is correct?
Both are internally consistent and neither has been independently reproduced for this review. They describe different populations, so there is no single correct figure to pick between them.
Is 2.6 percent or 0.73 percent the right number for my site?
Neither, necessarily. Both are aggregates across many industries, and category concentration varies enough that an individual distribution can sit well outside the published range.
Why does Google’s AI Assistant channel grouping matter?
It only exists from June 2026, so any study covering earlier months had to identify AI referrals by matching referring domains, which is a judgement call that affects the count.
Does a higher conversion rate from AI referrals mean I should invest?
Not on its own. The reported multiples use different numerators and baselines, and the denominators behind them are small, so they cannot support a spend decision by themselves.
When can you say a figure understates or overstates something?
Only when something concrete establishes the direction. SEO Works can call their AI contribution a floor because the exclusions they listed all remove AI sessions, but most limitations have no knowable direction and should be reported as unknown rather than guessed.
Can ChatGPT, Claude, Gemini and Perplexity be compared to each other directly?
Only on the same question under the same conditions, recorded with the date. Different engines execute different retrieval, so a difference between them is not automatically a difference about your brand.
ABOUT AXIS SUITE
Axis Suite by TrendAxis is the independent intelligence layer that explains what AI believes about your brand, why it believes it, and what decision that belief ultimately drives. See the recorded observations and the method behind them at axissuite.ai and in the Proof Center.