ChatGPT Referrals Grew 5×: Was It AEO?

Direct answer: a fivefold increase in ChatGPT referrals does not mean an AEO change caused a fivefold effect. A 2026 single-domain natural experiment found that total ChatGPT referrals rose 5.7× while untreated pages on the same domain rose 3.5×. The treated/control analysis estimated a smaller 1.82× intervention-aligned level increase, and a conservative placebo test kept the result suggestive rather than conclusive.

The useful result is the method: subtract the platform tailwind with a contemporaneous control before calling raw channel growth an optimization win.

Read the multiples in order

The headline and controlled views answer different questions
Measure	Reported result	Interpretation
All ChatGPT referrals	5.7×	Intervention plus platform growth and other shared changes
Untreated pages	3.5×	On-domain estimate of the channel tailwind
Treated/control ratio	1.82×; 95% CI 1.31–2.54	Intervention-aligned level change in the interrupted time-series model
Placebo-in-time test	p=0.16	Not strong enough for a conclusive causal claim under the conservative check

The study used first-party analytics and server logs for one high-traffic domain, glasp.co. Its AEO intervention affected a defined subset of hundreds of thousands of YouTube question-and-answer pages in January 2026; the untreated remainder of the same site served as the control.

Why an on-domain control helps

A platform can grow because more people use it, its web-search behavior changes, links become more prominent, or analytics classification improves. Those changes can lift treated and untreated pages together. An on-domain control shares much of the brand, analytics, infrastructure and platform environment, so it can absorb more of that common movement than a historical before/after comparison.

It is still not a perfect counterfactual. Treated and untreated templates may differ, the pre-period may be short, and site-wide releases can affect the groups differently. That is why the authors describe the effect as suggestive. The study is a preprint and a single-domain field study, not a universal AEO benchmark.

Build your own referral control

Define treated URLs before publishing the change.
Select untreated URLs with a comparable reader job, template, age and baseline.
Preserve server requests, referral sessions, engaged visits and conversions separately.
Record the intervention date and every site-wide or platform-relevant change.
Chart treated and control series, then chart their ratio.
Run placebo dates and sensitivity checks; report the result that survives them.

Do not remove zero days or reclassify the control after seeing the graph. Keep Google organic traffic as a guardrail when the intervention changes important page content. The study reported preserved indexation and no treated-page Google click loss beyond the surrounding site trend, but that result belongs to its setting.

Use the AI referral workflow for GA4 alongside server evidence, and the AI visibility crosswalk to avoid collapsing mentions, citations and visits into one metric.

Calculate the naive and controlled answers

Dividing the 5.7× total increase by the 3.5× untreated increase gives a simple 1.63× ratio. The study’s interrupted time-series model estimates a 1.82× level change instead. The two numbers should not match exactly: the first uses two headline multiples, while the model uses the weekly series, intervention timing and autocorrelation-robust inference.

Reported results and transparent arithmetic
Statistic	Value	How to read it
Naive tailwind-adjusted ratio	5.7 ÷ 3.5 = 1.63×	Calculated descriptive check, not the paper’s causal estimate.
Model level change	1.82×	Reported intervention-aligned estimate.
95% confidence interval	1.31×–2.54×	The estimated range is wide enough that planning should not use 1.82× as a guaranteed return.
Engagement-filtered traffic	2.27×	Reported result after applying the study’s engagement filter.
Placebo-in-time test	p=0.16	The conservative falsification check did not clear a conventional 0.05 threshold.

This combination is more informative than “traffic grew 5×.” It shows a large raw channel movement, substantial common tailwind, a positive controlled estimate, and remaining causal uncertainty—all at once.

Primary documentation

Watanabe and Nakayashiki: Disentangling AEO from Platform Growth — preprint submitted June 3, 2026.

Your ChatGPT Referrals Grew 5×. How Much Came From AEO?