You finally got the AI Overview citation. The dashboard lit up, someone screenshotted it for Slack, and for about four minutes the whole team felt like the GEO work had paid off. Then someone asked the pointed question nobody wants asked in that meeting: cited for what, exactly? A new study of AI search citations answers that in a way that should make every team pump the brakes before reporting citation count as a win on its own.
We Couldn't Reach Reddit, So We Went Looking for the Receipts
Normally this is where we'd drop into a Reddit thread and show you a marketer venting about a citation that turned out to be worthless. We couldn't reach Reddit again this run — down for the fourth day straight on our end — so instead we pulled a peer-reviewed study that documents the exact same anxiety practitioners have been posting about for months. Same pain point, better sourcing. Researchers at Washington University in St. Louis spent 40 days, from mid-March through late April 2026, crawling 55,393 trending queries and capturing 7,583 Google AI Overviews for the paper, arXiv:2605.14021. Then they did something almost nobody in our industry has the patience to do: they checked every individual claim against its cited source.
They broke those Overviews into 98,020 atomic claims — one fact, one attribution, checked line by line. That's the kind of grinding, unglamorous audit that produces numbers you can actually build a reporting framework on, instead of another vibes-based take on whether AI Overviews cite you or recommend a competitor in the same breath. If you've felt like the ground keeps moving under GEO reporting, this is the paper that explains why: citation and accuracy turned out to be two different questions with two different answers.
The 11% Problem: What AI Search Citations Actually Support
Here's the number that should reframe every GEO report on your desk: 11.0% of the claims inside Google AI Overviews were not supported by the source Google cited for them. Not "loosely supported." Not "supported with caveats." Unsupported. Split that open and it gets more specific — 7.0% of claims were simply absent from the source material entirely, and 4.1% were directly contradicted by it. Read that twice. Roughly one in every ten claims Google attributes to a domain either says nothing of the sort or says something different. If your page is the source on one of those, you didn't get misquoted by a person. You got misquoted by a system, at scale, automatically, with your brand name attached to the mistake.
| CLAIM OUTCOME | SHARE OF ATOMIC CLAIMS | WHAT IT MEANS |
|---|---|---|
| Supported by source | 89.0% | Cited page actually says what the Overview claims |
| Omitted from source | 7.0% | Claim never appears anywhere in the cited page |
| Contradicted by source | 4.1% | Cited page says something different or opposite |
None of this means the system is malicious. It means claim generation and source attribution are running on separate tracks that don't always check each other's work before publishing. For the practitioner, the practical fallout is the same either way: a citation next to your domain is not a fact-check, it's a pointer, and pointers can be wrong. Track this alongside how AI citation and search visibility are already diverging in your own reporting.
The 1% Click: Why AI Search Citations Rarely Send You Anyone
Say your citation clears the accuracy bar — the claim is real, the attribution is fair, nothing's contradicted. You still have a second problem, and it's the one that hits the P&L harder. Pew Research measured how often users actually click through to a source cited inside an AI answer, as reported by Tech Times. The number: 1%. Ninety-nine times out of a hundred, the person reading the AI Overview absorbs your claim and never opens your site. The citation did its job for the user. It did almost nothing for you.
That 1% is worth sitting with, because it's the number most citation-count dashboards quietly skip. A spike in "AI Overview mentions" looks like a win the same way a spike in impressions used to look like a win before anyone checked click-through rate. If this feels like the not-provided moment all over again, that's because it is — reach without attribution data. If your reporting stack still treats a citation as equivalent to a visit, you're one CFO question away from an uncomfortable meeting. This is exactly the kind of gap our reporting and analytics work is built to close before that meeting happens.
Source credibility: AI-Overview-cited domains vs. co-displayed organic results (PC1 credibility scale)
Here's the part that should complicate the panic a little, in the reassuring direction: the domains Google actually cites inside AI Overviews score higher on credibility metrics than the organic results sitting right next to them — 0.732 versus 0.645 on the study's composite scale. Google isn't citing garbage. It's citing better sources than the blue links it's burying below the fold. The catch: about 29.8% of those cited domains never show up in the regular first-page organic results at all, which means the AI Overview is running its own separate source-selection logic, not just promoting whatever already ranks.
The Ad Load Sting: Who Pays While Google Keeps Its Ads
Now the part that turns this from an accuracy story into an economics story. Over 50.6% of the pages AI Overviews cite carry their own display advertising — meaning more than half the publishers being used as Google's raw material are counting on click-through traffic to fund the content Google is summarizing. And that traffic isn't coming, remember: 1% click-through. Meanwhile Google's own ad inventory keeps running on those same results pages, undisturbed. The researchers put it plainly: AI Overviews suppress the organic clicks that drive publisher ad revenue while leaving Google's own ad inventory intact. One side of that trade loses reach it used to monetize. The other side keeps monetizing the page either way.
If you're the one arguing for content and GEO budget internally, this is the sentence to bring into the room: being the correct, credible, frequently-cited source is now decoupled from getting paid for it in traffic. That's not a reason to stop investing in AI/ML-adjacent visibility work or to pull back on structured, citable content. It's a reason to stop measuring the investment with a metric — raw citation count — that was never built to capture whether the citation converted into anything.
What to Report Instead of Citation Count
None of this is a case for giving up on GEO. It's a case for reporting it honestly. Citation count alone answers "did the system notice us." It doesn't answer "did the system represent us correctly" or "did anyone act on it." Those are two different questions, and the study above just handed you the framework to ask both. Pair every citation-volume number with a claim-fidelity spot-check and a referral-traffic reality check, the same way you'd pair organic and AI pipeline attribution instead of reporting either channel in isolation.
Getting cited was never the finish line — it was Google telling you your page cleared a bar worth citing. What happens after that is still on you: whether the claim next to your name is accurate, whether anyone reads past it, whether the traffic shows up. Keep chasing citations. Just stop reporting them like a page view. The teams that win the next stretch of GEO won't be the ones with the highest citation count — they'll be the ones who can prove what that count actually bought them.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.