Everyone in blogging has asked the same question this year: do Google’s AI answers actually send clicks, or do they just eat them? On October 28, 2026, SEO consultant Cyrus Shepard published the first real experiment that answers it — and the answer is weirder than either side expected. His site’s AI impressions earned an estimated click-through rate of about 0.9%. That’s tiny. But when he switched his site out of AI Overviews and AI Mode, his regular organic clicks fell too.
Here’s the short answer up front: AI answers are a poor traffic source on their own, but being visible in them seems to help the rest of your search presence. Shepard calls it the “Search SERP Halo Effect” — the more places a searcher sees your brand on one results page, the more likely they are to click at least one of them. Opting out to protest AI answers may cost you the clicks you were trying to protect. This article walks through exactly what the experiment found, what it did not prove, the wider traffic data around it, and how to run the same test on your own site before you touch any setting.
What Shepard actually did
Shepard used Google’s Search generative AI control — the site-wide switch in Search Console under Settings that pulls a site out of AI Overviews, AI Mode, and generative features in Discover while leaving it in regular Search. He turned it on for one site, waited, compared the numbers before and after, then turned it back on. He described the whole thing himself as an “admittedly small case study,” and that modesty matters. This is one site, over a short period, with no disclosed traffic volume, test dates, or size of the organic decline. Treat everything below as directional evidence, not a law of nature.
Three findings came out of it, reported by Search Engine Roundtable on October 30 and Search Engine Land:
1. AI impressions went to zero immediately. The moment the site was excluded, its Generative AI impressions flatlined. No lag, no taper. The switch works exactly as documented, and fast. Before the test, AI impressions made up about 8% of the site’s total impressions — a meaningful slice of visibility, even if the clicks didn’t follow.
2. AI impressions earned an estimated 0.9% click-through rate. Shepard worked this out backward from overall click data before and after, so it is an estimate, not a number Google reported. His own verdict: “not too shabby.” Context helps here. A 0.9% CTR sounds miserable next to a classic top ranking, but AI impressions are a different animal — the answer itself satisfies many queries, so the people who do click are a self-selected, motivated minority. Later in this article I’ll show why that minority may be worth more than the percentage suggests.
3. Regular organic clicks dropped too. This is the finding that made the rounds. With the site out of AI features, its ordinary blue-link clicks fell as well. Shepard’s explanation is the halo effect: a searcher who sees your brand named in the AI answer and again in the classic results trusts you a little more and clicks a little more readily. As he put it, “earning an AI citation link/mention may influence higher clicks anywhere on the SERP, even regular organic results further down the page.” Greg Sterling of Near Media added a supporting note in the discussion: his firm’s research has seen AI overviews create brand lift that turns into later searches and clicks, “not consistently but many times.”
One more wrinkle from the data keeps the story honest. Sitewide, the average position actually rose during the test — but Shepard attributes that to a couple of new posts plus brand search, not to the opt-out. On the individual URLs with the most AI visibility, the pattern was cleaner: position dropped while AI visibility was off, then rose again when it came back. That’s the closest thing in the study to a controlled comparison, and it points the same way as the click data.
What the experiment did NOT prove
Before you change strategy on the basis of one test, know its limits — because the people who reported it were careful about them, and you should be too.
First, the AI Overviews themselves never left the results page. During the test, searchers still saw AI answers for those queries; only Shepard’s site was absent from them. So this is not a test of “a world without AI Overviews.” It is a test of “your site in the answer versus your site out of it.” The distinction matters because it rules out the simplest reading — that AI answers hurt everyone equally — and replaces it with a subtler one: within a page that already has an AI answer, being named in it beats not being named.
Second, the 0.9% figure is derived, not measured. Google’s Generative AI report shows impressions, not clicks, so Shepard inferred the CTR from overall before-and-after click changes. If any other factor moved during the test window — a new post, a brand mention, a seasonal shift — the estimate moves with it. Search Engine Land flagged exactly this: without the site’s volume, dates, or the size of the organic decline, firm conclusions are out of reach.
Third, one site is one site. A site with 8% of impressions from AI features is fairly AI-exposed; a site with 1% would likely see a smaller halo, and a site with 30% might see a bigger one. Your mileage will vary with your niche, your query mix, and how often your pages get cited rather than merely summarized.
None of this makes the experiment useless. It makes it a first data point — and the first data point we have that measures the opt-out decision directly instead of arguing about it in theory. If you read our earlier guide to the opt-out control itself, you know the switch is reversible and takes effect within a day or two. That reversibility is what makes the test repeatable: you can run it yourself, which is exactly what the last section of this article shows you how to do.
The wider traffic picture is worse — and more complicated
Shepard’s test landed in a week of grim traffic news, and the combination is what makes the halo finding so uncomfortable. If AI answers were merely a small new traffic source, you could ignore them. The problem is that they are simultaneously shrinking the old traffic source.
Reddit told shareholders in late October that its search referrals fell by more than a million visits a week during the month, with CEO Steve Huffman naming AI Overviews and AI Mode as the main suspect and warning that daily users could drop by a low-double-digit percentage this quarter if the trend holds. Ahrefs compared 300,000 keywords and found that by December 2025, the top result on a page with an AI Overview got 58% fewer clicks than it would have without one — up from a 34.5% gap a year earlier. Google, for its part, says AI Overviews now reach more than 2.5 billion people a month and AI Mode more than a billion. The audience for AI answers is enormous; the click rate out of them is small. Both things are true at once.
But there is a counterweight, and it showed up the same week. The agency First Page Sage studied 218 of its client sites and found that visitors who clicked through from a page with an AI Overview converted to leads at 2.6%, against 1.7% without one. That fits the halo story from a different angle: the clicks that survive the AI answer are pre-qualified. Someone who reads the summary, still wants more, and clicks your link arrives warmer than someone who clicked a bare blue link cold. A 0.9% CTR looks very different if those clicks convert half again as well.
Put the three data points together and a coherent picture emerges. AI answers shrink total clicks from a results page. Being cited in the answer recovers a small, high-intent slice of them. And being cited may also lift your regular results through familiarity and trust. None of that makes AI search a good traffic source in the old sense. It makes visibility in AI answers a defensive asset: not the thing that grows your traffic, but one of the things that stops it shrinking faster.
The citation-card wrinkle you should watch
One more October development belongs in this story because it could change the math again. On October 29, SEO analyst Brodie Clark reported that the citation cards linking to sources in AI Overviews were appearing far less often. Search Engine Roundtable could not replicate the drop, and Google had not said whether it was a bug or a test. As of this writing, nobody knows.
Why does it matter? Because the halo effect depends on searchers seeing your name. If citation cards shrink — fewer visible source links, less prominent branding — the mechanism behind the halo weakens even if your pages are still technically “in” the answer. This is the kind of change that won’t show up as an announcement. It shows up as a slow drift in your Generative AI impressions, which is one more reason to check that report regularly rather than assuming today’s visibility is tomorrow’s.
The practical takeaway is simple: don’t set your AI strategy once and forget it. The surfaces, the link displays, and the citation formats are all still moving. A decision that was right in October may need revisiting in January, and the only way to know is to watch your own numbers.
Should you opt out? A decision worksheet
Given everything above, here is how I’d work through the decision for a typical small blog. This updates the framework from our opt-out guide with the new evidence; the control itself hasn’t changed, but the cost-benefit around it has.
Stay in if: most of your traffic comes from Google, your content is the citable kind (specific products, real prices, dated rules, step-by-step processes, original data), and your income depends on full-page reads, affiliate clicks, or email signups. The experiment suggests leaving costs you twice — the AI clicks and some regular clicks — and hands the answer slot to a competitor. For most blogs, this is still the default, and now it has experimental support rather than just intuition.
Consider opting out if: your content is the kind an AI answer can fully replace in two sentences (a definition, a single fact, a simple conversion), your Google dependence is low because newsletters, social, or direct visits carry you, or you have a principled objection to your work powering answers for free. Go in with eyes open: the experiment says this likely won’t restore lost clicks, and it may cost you regular ones. For some publishers that trade is still worth making on principle. Just don’t make it expecting a traffic rebound the data doesn’t promise.
Go page-level if: only some of your content is at risk. The site-wide switch is all-or-nothing, but snippet controls work per page: max-snippet, data-nosnippet, and nosnippet directives can keep a specific page’s text out of AI answers while the rest of your site stays in. This is the scalpel for the common case — a handful of pages that AI answers fully replace, on a site that otherwise benefits from citation. Our earlier guide to getting cited in AI search covers the flip side: what makes a page the cited source instead of the summarized one.
Whatever you choose, measure it. That brings us to the most useful part of this article.
Run the same test on your own site
You don’t need Shepard’s traffic to replicate his method. The Search generative AI control is reversible, takes effect within a day or two, and the Generative AI performance report gives you the before-and-after numbers. Here’s a protocol that avoids the traps in the original test:
1. Pick a baseline window first. Record four weeks of data before you change anything: total clicks, total impressions, average position, and Generative AI impressions, all from Search Console. Export the numbers — don’t trust memory, and don’t rely on screenshots. Four weeks smooths out weekday-weekend swings; two weeks is the minimum if you’re impatient, and anything shorter is noise.
2. Freeze everything else during the test. Don’t publish new posts, don’t redesign pages, don’t launch a newsletter push. Shepard’s own test was muddied by new posts and brand search lifting sitewide position. Your test is only interpretable if the opt-out is the only thing that changed. Tell anyone else who touches the site: nothing goes live until the test ends.
3. Flip the switch and wait at least two weeks. Go to Search Console, Settings, Search generative AI, and choose Exclude. Expect AI impressions to hit zero within a day or two — if they don’t, check whether a parent property is overriding your setting through inheritance. Let the test run a full two weeks minimum, four if your traffic is small. Watch total clicks and regular organic clicks separately; the halo hypothesis predicts both will dip, not just AI clicks.
4. Flip it back and compare the mirror image. Restore Include and watch for the rebound. The strongest version of the halo finding is a symmetric pattern: clicks down while out, clicks back up when restored, on the same URLs. If clicks stay down after restoration, something else changed during the test window, and your result is inconclusive — say so, don’t force a conclusion.
5. Decide on your data, then revisit quarterly. If your regular clicks fell while out and recovered when back, staying in is your answer, and you can stop agonizing about it. If nothing moved, your site may be less AI-exposed than Shepard’s, and the decision becomes genuinely free — make it on principle, not traffic fear. Either way, put a quarterly reminder on it: citation displays are shifting (see the citation-card wrinkle above), and today’s result may not hold in six months.
One honest note before you start: this test costs you something. If the halo effect is real, two to four weeks out of AI features means two to four weeks of slightly lower clicks. That’s the price of knowing instead of guessing. For a site where Google is a large share of traffic, run the test during a quiet period, not a launch week. And if the thought of deliberately cutting your own traffic for a month makes you queasy, that’s useful information too — it tells you how much you actually depend on the traffic you’re debating.
The bottom line
The first opt-out experiment gave bloggers two numbers that pull in opposite directions: a 0.9% click-through rate from AI impressions, and a drop in regular clicks when those impressions disappeared. The honest reading is that AI search is a weak traffic source and a surprisingly useful trust signal at the same time. For most small blogs, that means staying visible in AI answers, fighting to be the cited source rather than the summarized one, and checking your own Generative AI report instead of taking anyone’s study — including this one — as gospel. The switch is reversible, the test is repeatable, and your own data beats every pundit, including me.
For a video walkthrough of Shepard’s experiment, including the impressions chart and the caveats, see this breakdown:
Frequently Asked Questions
What is the AI Overviews click-through rate?
In Cyrus Shepard’s small October 2026 case study, AI impressions earned an estimated 0.9% click-through rate, worked out from overall before-and-after click data rather than measured directly by Google. It is one site’s estimate, not an industry benchmark, and Google does not publish click data for AI features.
What is the Search SERP Halo Effect?
Shepard’s term for his finding that appearing in both AI answers and classic results on the same page seems to lift clicks to the classic results. The idea is that seeing a brand twice builds familiarity and trust, making a searcher more likely to click at least one of its listings. It is a hypothesis supported by one small test, not a proven ranking mechanism.
Does opting out of AI Overviews hurt regular Google rankings?
Opting out via the Search generative AI control is not a ranking signal, according to Google — your classic results should not drop in position because of it. What Shepard observed was a drop in regular organic clicks, not necessarily rankings, which he attributes to lost visibility and trust signals on the page rather than a demotion.
How do I check my site’s AI impressions?
In Google Search Console, open the Generative AI performance report. It shows how often your pages appeared in AI Overviews, AI Mode, and generative features in Discover, broken down by page, country, device, and date. It shows impressions only — no clicks, average position, or queries — so pair it with the standard Performance report for click data.
Should small blogs opt out of Google AI search?
For most small blogs, no — the first experiment suggests opting out costs both AI clicks and some regular clicks. Consider it only if your content is fully replaceable by a two-sentence answer, your Google dependence is low, or you object on principle and accept the likely traffic cost. Whatever you choose, test it on your own data using the protocol above.
Is the 0.9% figure reliable?
It is directional, not definitive. It comes from one site over a short period, derived from before-and-after totals rather than direct measurement, with undisclosed traffic volume and test dates. Treat it as the first real data point on AI click rates, and weigh it alongside the Ahrefs, Reddit, and First Page Sage data — not as a number to quote as gospel.