Picture the scene: a search query goes in, an AI Overview comes out, as unremarkable as checking tomorrow's weather. Except in late July, two versions of nearly the same question came back with two different verdicts, and the gap between them wasn't subtle. One person typed "is it ok to only want a white doctor?" into Google. Someone else typed "is it ok to only want a black doctor?" Google's AI Overview didn't treat those two queries the same way, and the mismatch spread across X within hours — screenshots, quote-tweets, the whole reflexive pile-on a viral thread runs through. What happened next is the part worth sitting with longer than the outrage cycle allowed.
The Google AI Overview That Answered Two Different Ways
Here's what happened, as reported by Barry Schwartz at Search Engine Roundtable on July 30, 2026. Users testing Google's AI Overview noticed it answered a race-flipped pair of prompts inconsistently: "is it ok to only want a white doctor?" got a different kind of response than "is it ok to only want a black doctor?" Someone posted the comparison on X. It moved fast, the way anything touching race and AI tends to, and within a day Google had to respond publicly about its own product.
Search Engine Roundtable's report captured Google's actual statement, issued through its official account, and it's worth reading slowly rather than skimming past.
“Seems like this response is picking up particular web pages for this specific search. Not for all queries. — Google”
Google added that it was "looking into it" and would "work to improve." Read on its own, that's a perfectly reasonable customer-service reply. Read next to the noun doing the actual work in the sentence, it's something else. Not "our model." Not "our ranking logic." Particular web pages. The fault, as described by the company that built and shipped the system, lives out on the open web, not inside the thing summarizing it.
Google's Reflex Is to Blame the Page
Schwartz's framing is worth sitting with a beat longer than the headline suggests, because it names something practitioners feel but rarely say plainly: Google spends a lot of energy elsewhere insisting its AI outputs and search results are its own — for copyright purposes, for authorship purposes, in the legal battles it's fighting over who owns what a model produces. That's a defensible position when the fight is about ownership. It gets much less comfortable the moment the fight is about a bad answer, and the same company reaches for the opposite argument: that the output is really just a reflection of somebody else's web page.
Authorship is a convenient thing to claim. Right up until the output embarrasses you. Then, suddenly, you didn't write it — the internet did.
None of this makes Google's engineers dishonest, or the team responding on X cynical. It's a company doing what companies do when a product breaks in public: minimize, contextualize, redirect. But "not for all queries" and "looking into it" describe a bug in the pipeline. They don't describe who's accountable for the pipeline. That's a separate question, and it's the one worth asking about your own brand, not just about one viral screenshot.
The Pattern Behind Every Google AI Overview Error
Zoom out past this one incident and a shape starts to repeat. Three things happened in 2026 that, put side by side, describe the same system from three different angles: a German court, an SEO consultant running a live experiment, and a peer-reviewed study out of Washington University in St. Louis.
| INCIDENT | WHAT HAPPENED | WHO GOT BLAMED |
|---|---|---|
| Munich Regional Court I ruling — May 28, 2026 | Court held Google directly liable for false AI Overview statements about two publishers | Google itself — safe-harbor defense rejected |
| Lily Ray's fake-update experiment | AI Overviews and Perplexity cited a fabricated Google algorithm update as established fact within 24 hours | No one — the system just repeated it |
| Race-flipped doctor query — reported July 30, 2026 | AI Overview gave inconsistent answers depending on the race named in the query | "Particular web pages," per Google's own statement |
Start with the court. In May, Munich Regional Court I held Google directly liable for false AI Overview statements about two publishers, tying them to fraud and misleading subscription schemes that appeared in none of the sources the AI Overview actually cited. The court rejected the safe-harbor defense Google's lawyers reached for, ruling that AI Overviews generate new, authored statements rather than neutrally summarizing search results. That's the opposite of "particular web pages did this." That's a court saying: you wrote this, and you own what you wrote.
Then there's what happens before anything reaches a courtroom. SEO consultant Lily Ray fabricated a Google algorithm update that never existed, seeded it into a handful of low-effort posts, and watched AI Overviews and Perplexity start citing it as established fact within 24 hours. No lawsuit required. No malicious actor beyond one researcher testing a hypothesis. Just a system willing to repeat a confident-sounding claim it never actually verified.
And then there's the scale question, which should worry you more than a single viral screenshot. Researchers Xu, Iqbal, and Montgomery crawled 55,393 queries and broke the resulting AI Overviews into 98,020 individual claims, checking each one against its own cited source. 11.0% of those claims turned out to be unsupported by the source Google attributed them to. Not opinion, not framing — unsupported. That's not a race-and-doctors edge case. That's the baseline rate at which this system already misattributes what it says to pages that never said it.
Who Actually Owns the Error
Line these three up and nobody in the system currently agrees on who's holding the bag. A German court says Google is liable when its AI states something false about you. Google, caught in public, says the culprit is "particular web pages." And independent research says the model will confidently misattribute claims at a real, non-trivial rate — 11% isn't a rounding error, it's roughly one in nine.
The honest read isn't "Google is definitely responsible" or "you're definitely off the hook." It's messier than that, and pretending otherwise is its own kind of hype. Legally, liability looks like it's trending toward Google, at least in the one jurisdiction that's ruled on it directly. Operationally, in the moment something actually goes wrong — the moment that matters to your brand, before any court gets involved — Google's reflex is still to point at the page.
That gap between where the law might eventually land and where the company's public instinct sits today is exactly the space your reputation lives in. A ruling that finally sticks, an appeal that finally resolves, a policy fix that finally ships — all of that is measured in months or years. A bad AI Overview citing your page as the source for something false is measured in the time it takes someone to screenshot it.
What This Means for Your Page
This is where an audit stops being optional homework and starts being a defensive necessity. Not because you caused the last bad AI Overview answer you saw about your industry. Because you can't count on the platform to take the blame fast enough, or fix it fast enough, to matter to the person reading it right now.
That's especially true if you operate in healthcare or any other high-stakes category, where a misread claim isn't just embarrassing — it's the kind of thing that changes whether someone trusts a diagnosis, a provider, or a piece of medical guidance they found through an AI answer instead of a search result. The race-flipped doctor query that started this whole news cycle wasn't a random topic. Questions about who's an acceptable doctor sit exactly in that zone, which is presumably part of why it went viral instead of getting quietly patched.
An audit here isn't complicated in concept, even if it's real work in practice: go through your own highest-traffic, highest-stakes pages and ask what a model could plausibly extract, misread, or overstate from the way you've written them. That's what generative engine optimization done properly is actually for — not chasing citation counts, but shrinking the surface area for a system to misattribute something you never said.
Do this next: pick the page on your site most likely to get pulled into a sensitive AI Overview answer — a medical claim, a pricing statement, anything adjacent to a protected category — and read it the way a model would: in fragments, stripped of your intended framing. If you can imagine a careless summary of that fragment, rewrite it before Google's AI does the summarizing for you. Nobody's coming to fix this upstream on your timeline. Assume you're the one holding the page, because right now, that's exactly what Google is telling you too.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh 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.