• AI visibility

Guide · AI search

Our ChatGPT referral traffic is growing; can we credit our GEO work?

Not from the growth figure alone, because ChatGPT’s own growth lifts traffic to sites that did nothing. In the best-documented case, total ChatGPT referrals grew 5.7 times, but pages that received no optimization grew 3.5 times over the same months. To credit your work, compare the pages you changed with similar pages you left alone, and expect a far smaller, less certain effect than the headline.

By Underneath Team Guide Updated

The short version

  1. On one large website, ChatGPT referrals grew 5.7 times in five months, while pages that received no optimization grew 3.5 times (Watanabe and Nakayashiki, who work for the site’s owner).
  2. After netting out that platform growth, the optimization’s effect was about 1.8 to 2.3 times, and a stricter check could not rule out chance.
  3. A reanalysis of the same data by other researchers found the estimate ranged from 1.183 to 2.404 depending only on which week was treated as the start (Kato and colleagues).
  4. Google’s own design choices move clicks too: in a test with 1,100 US users, Google’s AI Mode cut click-through to websites by 18.8 percentage points (Wang and colleagues).
  5. A review of 45 studies concluded that claims about GEO return on investment outstrip the academic evidence (Martinez).

Does rising ChatGPT referral traffic prove your GEO work is paying off?

No: ChatGPT’s growth lifts referral traffic to almost every site, whether or not it was optimized. Watanabe and Nakayashiki (opens in a new tab) studied glasp.co, a website with hundreds of thousands of question-and-answer pages. In January 2026 the company optimized one section of the site for AI answers and left the rest alone.

On monthly totals, ChatGPT referrals to the whole site grew 5.7 times by May 2026. But the untouched pages grew 3.5 times over the same window, with no optimization at all. The authors call this the platform tailwind, and it was most of the growth.

The authors work at the company that owns the site, and this is one domain studied through ChatGPT alone. Still, it is the clearest public evidence on the question, and its lesson is general: a before-and-after figure on its own cannot separate your work from the market. The same caution applies to visibility scores; see whether GEO work raised your AI visibility.

How much of the growth was the optimization in the best-documented case?

About 1.8 to 2.3 times, against headline figures of 5.7 times or more. Read from the lowest day to the single best day, the daily series rose 38.2 times. That is the kind of number that ends up in case studies, and the authors call it a fragile basis for any estimate.

The optimized pages grew 6.1 times while the untouched pages grew 3.5 times. Measured week by week against the untouched pages, the optimized section showed a jump of 1.82 times when the work began. The authors put the defensible effect at about 1.8 to 2.3 times.

Even that is not settled. The optimized pages were already gaining share before the work started, by about 11% a month. When the researchers tested fake start dates in the months before, jumps nearly as large appeared by chance; the stricter check gave 0.16, short of the usual bar for proof. The authors suspect “many published AEO multiples are similarly inflated by conflating treatment with tailwind.”

What else can move AI referral numbers besides your work?

Analytics changes, timing choices and platform design changes can all shift the figures. In the glasp.co data, the share of ChatGPT visits counted as engaged rose from 0.486 in January to 0.896 in May. Most of that jump came from a mid-March change in how bot traffic was filtered, not from visitors behaving differently.

Kato and colleagues (opens in a new tab) reanalyzed the published data. Moving the assumed start of the work from 16 December 2025 to 6 January 2026 changed the estimated effect from 1.183 to 2.404 times. Allowing for the mid-March measurement change in their fullest models gave 1.337 to 1.455 times.

The platforms also change underneath you. In a randomized test with 1,100 US participants, Wang and colleagues (opens in a new tab) found that hiding Google’s AI Overviews raised click-through to websites by 8.8 percentage points. Switching people to AI Mode cut it by 18.8 points. We cover what that could mean for your traffic in our article on AI Mode as Google’s default.

What does a fair test of GEO look like?

Compare pages you changed with similar pages you did not, over the same weeks. Untouched pages on the same site share the same platform growth, brand and analytics settings. The difference between the two groups is what your work can claim.

Researchers use the same logic for platform effects. Khosravi and Yoganarasimhan (opens in a new tab) compared English Wikipedia articles, exposed to AI Overviews by default, with the same articles in German and French. They estimate AI Overviews cut English Wikipedia’s search traffic by 5.45% against the German comparison.

The strongest design is random assignment. Watanabe and Nakayashiki say holding back a random subset of eligible pages would have turned their estimate into a clean experiment; they did not run one. Their comparison did catch one useful result: organic Google clicks to the optimized pages fell about 25%, close to the 20% fall across the whole site, so the work did not visibly harm search traffic. Whether the reverse holds, with ChatGPT citations lifting Google, is covered in whether ChatGPT citations boost Google rankings.

Should referral traffic be your main measure of GEO?

No: people can read an AI answer without clicking, so referrals capture only part of the effect. Kato and colleagues point out that referral sessions measure a different event from being seen. They measured mentions directly instead, asking two ChatGPT models 56 questions; Glasp appeared in 33.8% of one model’s answers.

Referrals also reflect only the engines your analytics can see. On glasp.co, AI engines other than ChatGPT sent less than 1% of AI-referred visits in the study window. A brand can gain visibility in answers that send few or no visitors. Server logs add another view, showing which pages AI bots request.

The wider evidence on traffic is thin. Martinez (opens in a new tab) reviewed 45 studies and found no technique with a stable, long-term effect on discovery or on what users do. One industry study reported a 20% traffic lift against a control group, but without enough detail to judge it. The survey’s conclusion: “claims about GEO return on investment clearly outstrip the academic evidence.”

What should you do about it?

Before crediting GEO for referral growth, set up a comparison that can tell your work apart from ChatGPT’s growth.

  1. Keep a control group. Leave a similar set of pages untouched, ideally chosen at random, and compare growth between the two.
  2. Report the gap between groups, not the raw multiple. If both grew, only the difference belongs to your work.
  3. Fix the start date in advance and note any analytics or bot-filter changes, since both can move the estimate.
  4. Track AI answers directly as well as clicks: whether you are mentioned, how often, and on which engines. What predicts those mentions is examined in our study of Wikipedia, schema and recommendations.
  5. Watch Google’s organic traffic at the same time, to make sure AI-focused changes do not cost you search visits.
  6. Treat early results as suggestive until they hold for several months.

If you want help designing a measurement plan like this, see our generative engine optimization service.

What does the research not tell us yet?

There is still no randomized, multi-site study showing how much GEO work raises AI referral traffic.

  • The key evidence is one website, one engine and five months, published by the site’s own staff.
  • Its optimized and untouched pages differed in content and purpose, so they were not a perfect match.
  • The work was a bundle of changes, so no single tactic can be credited.
  • Nobody has yet linked AI mentions or referrals to sales in a published controlled study.

Frequently asked questions

How do I measure the ROI of GEO?

Compare the pages you optimized with similar untouched pages over the same period. On one site, untouched pages grew 3.5 times on their own, so only growth beyond that belongs to the work.

Why is our ChatGPT traffic growing when we have not changed anything?

ChatGPT itself is growing. In one study, pages that received no optimization still saw ChatGPT referrals grow 3.5 times in about five months.

Are published GEO case studies with big traffic multiples reliable?

Treat them with caution. In the one carefully analyzed case, a 38.2-times peak and a 5.7-times monthly rise shrank to an effect of about 1.8 to 2.3 times once platform growth was removed.

Can a GEO change hurt our Google rankings?

In the one case studied, no measurable harm was found. Organic clicks to the optimized pages fell about 25%, in line with a 20% fall across the whole site.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work