Someone turned AI Overviews off. The ordinary clicks moved too.
A late-October opt-out test put AI impressions at zero and watched organic clicks fall. It is one site, the overview stayed on the page, and it is not a reason to flip your whole domain off.

On 28 October 2026 Cyrus Shepard posted a small case study from Zyppy. He used the Search generative AI control in Search Console and took the site out of AI Overviews and AI Mode. Three things happened. AI impressions went to zero, immediately. He estimated those impressions had been worth about a 0.9% click-through rate, worked out from the gap in overall clicks once they disappeared. And ordinary organic clicks dropped as well. He called the third one a halo: showing up in more than one place on the page makes a click somewhere more likely. He also wrote the sentence the screenshots tend to lose. Results are not guaranteed. Test your own data.
Two days later Barry Schwartz wrote it up at Search Engine Roundtable, and Shepard filled in the limits. AI Overviews were on the results page before the test, during it, and after it. The site was cited before and after, and missing during. This experiment does not measure a world without AI Overviews. It measures one site losing its citation while the summary stayed up for everyone else.
0
AI impressions after the opt-out
They dropped immediately. The control does what the settings screen says.
0.9%
Estimated click rate of those AI impressions
Inferred from overall clicks, not a click column Google reports.
8%
Share of impressions that were the AI ones
What Shepard told Search Engine Roundtable. One site.
What 0.9% actually is
Search Console still does not split AI clicks out of the rest. The Generative AI performance report, the one that went worldwide on 31 August, shows impressions from AI Overviews and AI Mode. It does not show clicks, queries, or a position of its own. Shepard's 0.9% is an estimate: overall clicks and impressions, with the AI-impression period subtracted, on a site where those impressions were about 8% of the total. That is a clever use of a blunt report. It is also one before-and-after, on one site, with no second site sitting still beside it. Schwartz called the sample small. Believe the direction. Do not tattoo 0.9% onto a quarterly plan.
He told Schwartz something else that keeps the story honest. On the URLs with the most AI visibility, average position worsened while the site was opted out and improved when the citation came back. Across the whole site, average position actually got better, which he put down to new posts and brand search, not to the opt-out. Hold both. The clean chart and the messy one arrived in the same conversation.
- AI impressions8%
- Other impressions92%

Two pitches this test cuts down
The first pitch says opt out. The summaries steal the click and give nothing back, so leave the feature. On this site, removing the citation lined up with fewer organic clicks, not more. If the halo is real, the summary and the blue link are one appearance. We said on 5 October to leave the control on if you want the company name answered properly. This is the first public measurement of turning it off, and it does not argue for a company-wide opt-out. Google has also said the control is not a ranking signal for ordinary results. Shepard's position movement on the high-visibility URLs is still worth watching, because "not a ranking signal" and "the page behaves differently when you vanish from the summary" can both be true.
The second pitch says AI Overviews are now a fifth of your traffic, so buy the package. That percentage is not in this test. Eight percent of impressions, clicked at a higher rate than a sleepy site usually manages, can be a real slice of attention. It is one SEO publication that has been working to be cited. Your site is not that site. A local-intent query and a how-to query do not even earn overviews at the same rate, which is the September finding this test does not replace.
How to get your own number without blinding the domain
The control applies to the Search Console property you set it on. Opting the whole company property out, to see what happens, is how you spend a month invisible in the one place buyers have started reading your name. Through September, branded queries were the place overviews spread fastest. That is the page you least want to take out of the experiment by accident.
If you want a measurement of your own:
- Put one section on its own URL-prefix property in Search Console.
- Turn the generative AI control off for that property. Leave the rest of the site on, over the same dates.
- Compare clicks on the opted-out URLs with clicks on the URLs you left in. A vibe from one chart is how the 0.9% escapes its footnote.
- Turn it back on when you have the gap. On Shepard's site the high-visibility URLs recovered position when the citation returned. You want the same week available to you.
- Keep a separate log for ChatGPT and Perplexity. This test is about Google's summary. Rank still feeds that summary. It does not feed ChatGPT the same way, which is the split in the March playbook we corrected in October.
Before
the site is cited inside the overview
the overview is on the page
|
v
During
the site is missing from the overview
the overview is still on the page
AI impressions go to zero
ordinary clicks fall with them
|
v
After
the citation comes back
the URLs that had the most AI visibility recover position
What to keep doing
Nothing in the late-October test retires the work from September. A page that ranks, and that says something specific in the first sentences, is still what Google's shortlist is built from. Turning that page off to prove a point is a research design. It is a poor way to run the site. The opt-out is there for a section you have a reason to exclude, or for a prefix property you are willing to treat as a test. It is not a strategy for the domain.
The off switch works. That is a different claim from the off switch being wise.
The branded check from October still comes first. Search the company name logged out, read what the overview claims, and fix the sentence on the page it should have cited. Then, if you want a number for the halo on your own URLs, run it on a prefix and put the control back. The page that ranks and can be quoted is the asset. The control is a light switch for that asset, not a new channel.


