ChatGPT traffic grew even on pages we did not optimize
Visits from ChatGPT to pages we changed grew 6.1 times. We aimed the changes at services that answer questions and link to sources. Pages we left alone grew 3.5 times too. The changes may have added a further step up, but a stricter timing test leaves that only suggestive.
In short
- Much of the growth came without changes. Visits from ChatGPT grew faster on the changed pages, but the rest of the site grew substantially too.
- The extra step up is suggestive. Allowing for the rest of the site and the earlier trend leaves a step of about 1.8 times at the rollout. Jumps almost as large appear at made-up dates before it.
- We tested several changes together on one site. We cannot say which change worked, or whether another website or service would see the same gain.
An answer engine answers questions and may link to sources, as ChatGPT does. Answer engine optimization, or AEO, means changing a website to help its pages appear in those answers. If visits from ChatGPT rise afterwards, the changes may have helped. ChatGPT may also be sending more visitors as its own use grows. A before-and-after chart mixes those possibilities together.
A large traffic increase also happened without the changes
We changed Glasp's pages that answer questions about YouTube videos, starting in January 2026, and compared them with the rest of our website. Every number here is an aggregate of site traffic; no individual visitor is described. We counted visits arriving from ChatGPT links, not mentions or citations in its answers.
Monthly visits from ChatGPT grew 5.7 times overall. Changed pages grew 6.1 times; pages we left alone grew 3.5 times. That second number is the clue: much of the growth came without our changes (Figure 1).
Figure 1
Visits from ChatGPT rose on changed and unchanged pages alike
- Pages receiving the changes
- Other pages on the same site
Show the numbers
| Page group | January index | May index | Reported growth |
|---|---|---|---|
| Changed pages | 100.0 | 609.8 | 6.1 times |
| Comparison pages | 100.0 | 348.5 | 3.5 times |
| May observation ends | May 28, 2026 | Partial month |
Comparing the lowest single day with the highest would give a 38.2-times increase. That peak did not last; the sustained late-May level was much lower. We used monthly totals to reduce the influence of that brief spike.
The test measures all the changes together
We merged duplicate web addresses for the same content under one preferred address. Requests by AI bots for pages that did not exist helped us choose new pages to create. Titles became questions, and opening summaries became short answers that made sense on their own. We kept pages already earning meaningful Google traffic out of the rewriting queue.
We made these changes together, so we cannot separate their effects. The bundle also added pages: the estimated gain combines changes to existing pages with a larger collection of pages.
We did not randomly choose which pages to change. Questions about videos may attract different visits from articles or discovery pages. If ChatGPT starts favoring a particular kind of content, those groups could grow at different rates even without our changes.
Allowing for the earlier trend leaves a possible step up at the rollout
The simple ratio of growth rates is 1.75 times. Our weekly model estimates a 1.82-times step at the rollout. The range the data allow (a 95% interval) is 1.31 to 2.54 (Figure 2). The 1.75 compares growth over the whole period. The 1.82 estimates the step at the rollout after allowing for the earlier trend.
Figure 2
The estimated step up appears in every version of the analysis
- 95% interval for the main estimate
Show the numbers
| Version of the analysis | Step-up multiplier | 95% interval |
|---|---|---|
| Main | 1.82 | 1.31 to 2.54 |
| Visits passing the engagement filter | 2.27 | Not reported in the paper's text |
| Early-May spike weeks excluded | 1.83 | Not reported in the paper's text |
| Pre-period from September 2025 | 2.07 | Not reported in the paper's text |
The changed pages were already gaining on the rest of the site before January. Continuing that trend would produce growth without any extra effect from our changes. The weekly model finds a possible step up at the rollout, but no detectable acceleration afterwards.
All versions reuse the same rollout and assumptions. We also used visits that pass the analytics system's engagement filter, a measure less affected by non-human referral traffic. That version still gives a positive step. So do removing the early-May spike weeks and measuring the earlier trend from a later date.
Jumps almost as large appear at made-up dates before the rollout
A short, noisy history can make an ordinary fluctuation look like a change caused by the rollout. We moved the supposed rollout to made-up dates before January and ran the analysis again. This checks whether January's jump is unusually large compared with jumps in the earlier traffic.
Figure 3
The made-up-dates test gives a less convincing result
Show the numbers
| Test | p-value | What the comparison asks |
|---|---|---|
| Weekly model | 0.001 | Is January's step clear under the weekly model's assumptions? |
| Made-up dates | 0.16 | Is January's jump unusual compared with jumps at earlier dates? |
The tests give different answers (Figure 3), leaving the effect suggestive, not conclusive. The weekly model allows for neighboring weeks moving together and finds a clear jump under its assumptions. The made-up-dates test finds that the earlier traffic can produce jumps nearly as large.
Google traffic fell much as it did across the rest of the site
We also checked whether rewriting hurt visits from Google search. Clicks to changed pages fell about 25%, while site-wide clicks fell about 20%, consistent with the rule protecting pages that already earned Google traffic. The pages still appeared in Google search, and the number of times they appeared stayed near its normal range.
For a stronger causal test, we propose a follow-up that randomly leaves some eligible pages unchanged. We did not run it here. We also found no sign of a cycle in which ChatGPT visits then lift Google traffic: Google traffic was flat or declining.
What this study does not show
- We did not randomly choose pages to change. The rollout date is approximate, the groups contain different kinds of content, and the earlier trend is short and noisy. The made-up-dates test leaves the effect suggestive.
- The adjustment makes an assumption. Changes at ChatGPT must lift both groups by the same percentage. If our changes also help other pages on the site, the comparison can understate the bundle's effect.
- The bundle includes more pages. Its components were not varied separately, and the estimate combines changes to page performance with expansion of the page collection.
- The result comes from one site. The mix of traffic changed during the window, and the evidence is predominantly about ChatGPT. We do not know whether other engines or websites would see the same result.
The paper
Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic
Keisuke Watanabe and Kazuki Nakayashiki · June 2026 · arXiv:2606.04362
@misc{watanabe2026aeo,
title = {Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic},
author = {Watanabe, Keisuke and Nakayashiki, Kazuki},
year = {2026},
eprint = {2606.04362},
archivePrefix = {arXiv},
primaryClass = {cs.IR},
url = {https://arxiv.org/abs/2606.04362}
}This is a plain-language summary of the paper above. More studies are on the Glasp Research page. Glasp is a social web highlighter: see how it works.