AI Overviews Earned One Site a 0.9% CTR, and Leaving Still Cost It Clicks

AI Overviews Earned One Site a 0.9% CTR, and Leaving Still Cost It Clicks
Search Console's opt-out removes a site from AI Overviews, but the block stays on the page without you.

SEO consultant Cyrus Shepard used Google's Search Console opt-out to pull one site from AI Overviews and AI Mode for about 25 days in September. AI impressions fell to zero, and he estimated they had earned only a 0.9% CTR, yet regular organic clicks fell too. Before anyone copies him, run a 28-day test with a control group, and not during BFCM.

The 0.9% is going to get screenshotted a lot this week, because it seems to confirm what publishers already feel: AI answers take the attention and hand back almost nothing. If that's the number you read, the opt-out looks like a free decision. You'd give up a sliver of low-quality exposure and lose nothing that matters.

The more useful half of Barry Schwartz's write-up is the part that cuts against that. According to Shepard, "regular organic clicks ALSO DROPPED when AI features were disabled." He calls the idea a "Search SERP Halo Effect," meaning that showing up in more places on a results page makes a click more likely anywhere on it.

I think he might be right. I also think this test can't prove it, and he said as much himself, calling it a small case study and telling people to test their own data. So let's look at what it can actually carry.

Six clicks a day is carrying a lot of weight

The chart in the Roundtable piece shows AI impressions running somewhere around 400 to 1,100 a day, then flatlining from roughly September 3 to September 28, then coming back. The report totals 42.6K AI impressions across the window. Spread that over the 65 or so days the site was actually opted in and you get about 650 a day.

Apply the 0.9% and that's about six clicks a day. (My arithmetic from the chart, so treat it as a ballpark, because Shepard hasn't given exact daily figures in anything I've read.)

Six clicks a day should not be the reason anyone changes a sitewide setting.

Those AI impressions were about 8% of the site's total, which also explains part of why overall impressions dropped during the opt-out. Some of that fall is just the removed AI impressions leaving the total. Google's own Search Console help page says standard impression rules apply to AI Overviews, and that clicking an external link in one counts as a click.

That last detail matters more than it sounds. The main Performance report blends AI and regular clicks, and as of the June launch of the generative AI reports, those reports showed impressions with no click data. Shepard's 0.9% was estimated by subtracting known AI impressions from the site's overall CTR and impressions during the test. So "AI clicks" and "regular clicks" are being carved out of one pool of numbers, and any error in the first estimate lands in the second one. On a site this size, it doesn't take much of an error.

The position drop may be mostly a reporting quirk

One finding looks like a smoking gun: the URLs with the most AI visibility fell in position during the opt-out, then rose when visibility came back. Read quickly, that suggests Google punished the site for leaving, despite saying the control "won't be used as a ranking signal."

There's a much duller explanation. Search Console's position is the topmost spot occupied by a link to your site, and Google says an AI Overview occupies a single position in the results, with every link inside it sharing that position. John Mueller has conceded that "position for these is hard to do in a way that makes it useful."

So if a URL ranks fifth organically but also gets cited in an AI Overview sitting above that, it reports the block's position. Pull it from the block and it reports fifth. Nothing about its ranking changed, but the graph would look exactly like a penalty and a recovery. I can't say that explains all of Shepard's movement, and to be fair he reports his sitewide average position went up over the same stretch, which he credits to new posts and brand search. I'd just want to rule this out before telling anyone that opting out costs rankings.

New posts and brand search are a confound on their own, too. They were changing at the same time as the thing being tested. That's normal for a real site. It's also why one before-and-after line can't settle much.

The halo idea has decent company, though not proof

Here's where I end up leaning Shepard's way anyway. Note the setup: AI Overviews stayed on the results pages throughout. His site just wasn't in them. So the test compared being cited to being absent from a block that still pushes everything else down, which is closer to the choice publishers actually face than the usual with-and-without comparison.

And Seer Interactive's 2026 update, drawn from 53 accounts and about 5.47 million queries, points the same direction. Across informational queries in 2025, organic CTR averaged 2.07% when a brand was cited in the AI Overview and 0.94% when it wasn't. Seer is careful here, saying it "cannot claim causation" because stronger brands are more likely to get cited in the first place, and one account makes up 47% of the not-cited impressions. Without that account the figure is closer to 1.61%, which shrinks the gap a good deal.

Greg Sterling of Near Media offered another mechanism in the Roundtable piece: AI Overviews "create a kind of brand lift or awareness that translates into subsequent searches and clicks." Honestly, that one feels most plausible to me. It's a bit like taking your store off the mall directory because only one in a hundred people who read it walks straight to you. The other 99 still saw your name, and some of them turn up later through a different door, and you never get to attribute them.

Which is also what makes this so annoying to measure.

If you want the answer for your own site, test it after Cyber Monday

The control has been spreading since its UK debut in June. Herdl reported in July that owners in countries including the US were seeing it, and that changes typically take a few days to apply. Check your own property before planning around it. It's also domain-level for now. PPC Land reports that page-level grounding controls under the UK rules don't come into force until March 2027, so you can't simply split your pages into test and control through the toggle.

You can still get something better than one line on a chart. Personally, I wouldn't touch the toggle today. Here's what I'd do instead, and the first part takes about ten minutes:

  • Size the stake. Open the generative AI report and export AI impressions for the last 28 days. Multiply by 0.9% as a placeholder, then divide by your total clicks for the same period. Under 1% of clicks means you're not deciding about traffic at all, so any opt-out case has to rest on licensing or content-use grounds.
  • Build two page groups before you change anything. Group A is pages with the most AI impressions. Group B is pages with similar click volume and close to zero AI impressions. Group B is your control, since the toggle won't give you one.
  • Split branded queries out. A regex filter on your brand name will do. If the brand-lift idea is real, it should show up there first, and it's the part a sitewide total hides.
  • Run three phases of 28 days each: in, out, back in. Shepard's window was around 25 days out, so this isn't far off. Skip the holiday stretch, because BFCM demand would swamp everything. A January start finishes around the end of March, right as page-level controls should be arriving for UK sites.
  • Set your threshold before you look. My rule: Group A has to lose at least 10% more clicks than Group B during the opt-out, and recover most of that afterward. One without the other is noise. I'd also want at least 1,000 clicks a week in each group. Below that, you're probably back to reading six-clicks-a-day tea leaves.

The recovery phase is the part people will skip, and it's the part that makes Shepard's test worth reading at all. His AI impressions went away and came back on command, and a lot of before-and-after SEO tests never get that far. Opting out also carries its own cost, because you're giving up a month of citations to learn something, so if your site earns most of its leads from informational content I'd think hard about whether you need the answer this year.

If you're trying to work out where you stand first, our breakdowns of how far AI Overviews have spread into review searches and why daily citation snapshots mislead cover the two numbers people most often misread in this report.

My bet on the next few of these tests

Schwartz notes nobody else has published similar data, and that won't last, since the toggle is sitting right there and SEOs love this sort of thing. I'd guess we see at least five public opt-out tests by the end of March 2027, and I'd put money on most of them landing within 10% either way on regular clicks once someone controls for seasonality. That would make the opt-out less of a traffic decision than either camp wants it to be.

And that might be the actual finding here, even if it's a less exciting one to post. Shepard may well have found something real, or he may have found six clicks a day and a quirk in how Search Console labels position. I don't know yet, and I'm not sure he does either, which is more or less what he said in the first place.