Opting Out of AI Answers Cost One Site Its Regular Clicks, Too

The standard fear about AI Overviews goes like this: Google answers the question at the top of the page, nobody clicks, and the sites that supplied the answer starve. The defensive move that follows from that fear is to opt out — block your site from the AI answers and protect the organic clicks underneath.

Cyrus Shepard actually ran that move. Using Google’s own generative-AI opt-out controls, he blocked a site from appearing in AI Overviews and AI Mode, and compared what happened before, during, and after. AI Overviews still appeared on those results pages throughout — his site simply wasn’t in them while the opt-out was on. What the data showed is the opposite of protection.

What the test found

The AI slot itself was worth less than the panic assumes — and more than zero. AI impressions were about 8 percent of the site’s total impressions, and Shepard estimates the click-through rate on those impressions at roughly 0.9 percent. That number is derived, not dialed in: he calculated it from overall before-and-after click and impression data, subtracting the known AI impressions from the opt-out period. Call it an informed estimate, not a Search Console readout.

The surprise was underneath. When the opt-out switched on, AI impressions went to zero immediately — expected. But regular organic clicks dropped too. And on the individual URLs with the most AI visibility, rankings and clicks sagged during the opt-out and recovered when AI visibility was switched back on.

One confound, disclosed by Shepard himself: sitewide average position actually rose during the test, thanks to a couple of new posts and brand search. The halo signal lives at the URL level — the pages that had been most visible in AI answers are the ones that lost ground when they vanished from them.

The halo effect

Shepard names it the Search SERP Halo Effect: the more places your brand appears on a results page, the more likely a searcher clicks at least one of them. An AI citation, on this reading, isn’t just a link inside the answer — it’s an introduction. The searcher meets the source in the AI answer, then recognizes it in the organic results below and clicks there.

Greg Sterling of Near Media, commenting on the study, described the same pattern from their research: AI Overviews appear to create a brand lift that translates into subsequent searches and clicks — “not consistently but many times.”

The honest limits

This is one site, one test, and a small sample — Shepard calls it a “small case study” himself, and the 0.9 percent figure is subtraction math, not a measurement Google hands you. The comparison is also narrower than the headline version: it tests a site being in or out of AI answers that still appear either way, not a results page with AI answers versus one without. Nobody should rewrite a strategy on this alone.

But weigh the other side of the ledger: almost nobody else has produced data like this at all, because running the experiment means volunteering to disappear from the fastest-growing surface in search. The evidence base for “AI answers kill clicks” is mostly vibes and correlation. This is, so far, the closest thing to a controlled test — and it points the other way.

Why this matters to how we run Tygart Media

This site’s whole strategy is citation authority: structure pages so AI systems cite them, and measure the citations. Our most-cited page has been referenced hundreds of thousands of times in AI answers this year. The standing question from skeptics is what those citations are worth if the click happens — or doesn’t — somewhere else.

The halo test offers the first real answer: a citation is not the consolation prize for losing the click. It appears to be one of the things that earns the click, including clicks on the plain organic listing further down the page that everyone assumed the AI answer was cannibalizing. Being the named source in the answer and being a familiar name in the results are the same asset compounding.

And the flip side is a warning for anyone tempted by the opt-out: in the only public test we have, leaving the AI answers didn’t defend the organic traffic. It taxed both channels at once.

One small study doesn’t settle a question this big. But it changes which side has to produce evidence. The claim “opt out and you’ll be fine” now has a controlled test standing against it — and the halo has a name, a mechanism, and numbers, however small. We’ll take that trade.

Sources: Cyrus Shepard’s opt-out case study, as reported by Bitcomme (Nov 1, 2026) and discussed on LinkedIn; Near Media comment via Greg Sterling. Study limits are stated in the piece: single site, small sample, CTR estimated by before/after subtraction.