Why every AI Overviews study seems to say something different

Source: The Influence of Retrieval, Fan-out, and Google SERPs on ChatGPT Citations, AirOps, Branded vs unbranded AI citations, Foundation Marketing, How AI engines search and cite, Prefer, AIO Impact on Google CTR: 2026 Update, Seer Interactive

TL;DR: Our read: most AI search studies don't really disagree. They measure different engines, different questions and different things, then get squeezed into one headline. Before trusting a number, check which engine, which questions, and what exactly was counted.

One week a post says AI cites the top Google result. The next says it mostly cites small sites. A third says AI Overviews cost you clicks, and a fourth says they bring more.

They can all be true at once.

The same study, read two ways

Take AirOps' 2026 study of ChatGPT. It covered 15,000 prompts and 82,108 citations. Pages ranked #1 on Google were cited 43.2% of the time, 3.5 times the rate for pages outside the top 20.

That has been retold online as "half of ChatGPT's citations come from #1 pages." It flips the numbers around. The study counted how often #1 pages got cited, not what share of all citations they made up.

The same report also found that about 74% of citations went to sites with DA under 80, and that 85% of retrieved pages were never cited. So "rank first" and "small sites win" both come out of one dataset. Neither is the whole story.

Different questions, different answers

Foundation Marketing and AirOps studied 50 B2B brands on five AI platforms. When people asked about a brand by name, 77.6% of answers cited the brand's own site. For unbranded category questions, only 2.2% of cited links went to brand-owned sites.

Those are two different measures, a share of answers and a share of links, so dividing one by the other gives a multiplier that means nothing. But they show how much the question type changes the result. A study made of branded questions and one made of category questions will reach opposite conclusions, both correctly.

The engines don't behave alike

Prefer ran 80 questions three times on four engines. Two runs of the same question shared, on average, 37% of the sites cited across both runs on ChatGPT, and 96% on Perplexity. Those figures come from API answers, which can differ from the consumer apps. Most cited sites, 72.7%, were cited by only one engine.

So a finding about ChatGPT may not hold for Perplexity, or for Google's AI Overviews. And a ChatGPT study that runs each question once is partly measuring chance.

Clicks: losing and winning together

Seer Interactive tracked 53 brands and 5.47 million queries. On informational queries, brands cited in an AI Overview got about 120% more organic clicks per impression than brands that weren't. They also got about 38% fewer than on queries with no AI Overview at all.

"AI Overviews cut clicks" and "AI Overviews boost clicks" are both in that one report, depending on what you compare against. Seer also notes it is correlation: higher-authority brands get cited more to begin with.

What our own small run shows

We ran our tool, QueryTrace, on 233 AI answers about real businesses across five engines. The business was named in 163. Its own site was cited in 68 of those, the AI read its page without crediting it in 33 more, and 62 named it without its site. Count the last two together: 95 answers name the business without citing its site, 33 because the page was read and not credited, and 62 because the site was not involved at all.

It's a small sample and we only share totals. It sits comfortably next to AirOps: being found and being credited are separate steps, and counting one is not counting the other.

Why the headlines drift

Part of the problem is the retelling. A study's own page usually carries its method and caveats: how many prompts, which engines, which dates. A social post has room for one number. By the third retelling the number has lost its conditions, and sometimes its direction.

Timing matters too. These engines change their search and citation behavior often, and a study is a snapshot of the weeks its data was collected. Two honest studies a few months apart can disagree simply because the systems moved in between. When you see a figure, look for the collection dates before you look at the figure itself.

Some claims that don't hold up

A few popular claims don't survive a look at the sources. The "half of citations from #1" line is a misreading, as above. Claims that adding schema markup or an FAQ section reliably boosts AI citations aren't backed by the studies we checked. Treat them as guesses until someone shows the data.

How to read the next study

1. Which engine? ChatGPT, Perplexity and AI Overviews behave differently.
2. Which questions? Branded, category, or about the AI search niche itself.
3. What was counted? Share of pages, answers, links or clicks.
4. How many runs? One run per question on ChatGPT is mostly noise.
5. Who ran it? Most of these are vendors or agencies. That doesn't make them wrong, but it shapes what they measure.
6. Correlation or cause? Almost none of these are experiments.