When the answer engine told people to eat rocks
The structure and markup of this piece were refreshed for current answer engines; the original analysis is preserved as written.
In May 2024, days after Google put AI Overviews in front of US searchers, it was telling them to add glue to pizza sauce and eat a small rock a day. The internet laughed — but the moment matters: it is when the public saw, plainly, that an AI can be confidently, fluently wrong.
the short answer
In May 2024, Google’s AI Overviews advised glue on pizza and eating rocks — pulling a sarcastic Reddit joke and a satirical article in as fact. The lesson is that hallucination is the nature of systems that generate plausible text, not a bug to be patched, and a confident tone is not correctness. You cannot control what AI says about your category, but the quality of the sources on your topic shifts whether it gets it right.
key takeaways
- In May 2024, Google’s AI Overviews told people to put non-toxic glue on pizza and eat a small rock a day — the moment the public saw AI invent with total confidence.
- The errors came from the AI treating sarcastic and satirical sources as fact: the glue advice traced to an 11-year-old Reddit joke, the rock advice to a piece from The Onion.
- Hallucination is not a rare glitch but the nature of a system that generates plausible text rather than retrieving verified truth — a confident tone does not mean a correct answer.
- Google added safeguards and the worst cases grew rarer, but reliability is a standing risk to manage, not a problem closed on your behalf.
- You cannot control what AI says about your category, but the clarity and quality of the sources on your topic shifts whether it gets things right or fills the gap with nonsense.
what the source environment produces
It is the same model in both rows. What differs is what it had to work with. Given thin or sarcastic material, it fills the gap and invents with confidence; given clear, credible sources, it has something solid to ground the answer on. You do not edit the engine — you change which row your category sits in.
Why this changes how you weigh AI visibility
The glue-pizza episode is easy to file under comedy, but for anyone thinking about AI and their brand it carries a sharper message. The same system that can be the place buyers discover you is a system that can state falsehoods about your category with a straight face. Both things are true at once, and a serious approach holds them together rather than picking the convenient one. It means not overpromising — no honest practitioner can guarantee what an AI will say about you — and it means not retreating either, because absence does not make the risk go away; it just removes your accurate account from the pool the model draws on. The mature posture treats AI visibility as worthwhile and imperfect: worth pursuing because the model will answer about you regardless, imperfect because the answer can be wrong no matter how well you prepare.
It also reframes what "good content" is buying you. The value is not a guarantee of accuracy, which no one can offer; it is a shift in the odds. A category where the clearest, most credible sources state the facts plainly is a category where the model has the least room to improvise something wrong. That is a probabilistic hedge, and naming it as such is part of being trustworthy about the work. The vendors who promise to "control" what AI says, or to guarantee accuracy, are selling a certainty the technology does not support. The honest offer is narrower and more durable: help make the truth about your field the most available, best-structured thing an AI can find, lower the chance of confident error, and treat reliability as a risk you manage continuously rather than a box you tick once.
The lesson, in three parts
Why confident invention is the nature of the system, why thin sources make it worse, and what a brand can actually do about it. Open each layer for the part that changes how you think about AI as a source.
01 Why hallucination is the default, not the glitch
The instinct, watching an AI suggest glue on pizza, is to call it a bug — a defect in a machine that otherwise tells the truth. That gets it backwards. A generative model does not store facts and look them up; it produces text that is statistically likely given its inputs. Most of the time, plausible and true line up, and the output is right. But the model has no internal sense of which is which: a confident falsehood is generated by exactly the same process as a confident truth, with the same fluent, certain tone. That is why the wrong answers were not hesitant or hedged — they were delivered with the same authority as everything else. The lesson the public absorbed in May 2024 is one worth keeping: the confidence of an AI answer carries no information about its correctness. Hallucination is not the system failing to work; it is the system working as designed, on an input where plausible and true happened to come apart.
02 Why thin sources make it worse
The specific failures were revealing because of where they came from. The glue-on-pizza advice traced back to a sarcastic Reddit comment from more than a decade earlier; the eat-a-rock suggestion came from satire published by The Onion. The model was not malfunctioning at random — it was doing its best to answer questions where the available material was a joke, and it could not tell the joke from sincere advice. That points to a real pattern: a generative engine is most dangerous where the genuine sources on a topic are sparse, low-quality, or hard to distinguish from sarcasm, because it will still produce a confident answer, assembled from whatever is there. Google itself, explaining the episode, pointed to forum content interpreted without its context. The implication for any topic is direct: where good, clear, unambiguous sources are scarce, the model has nothing solid to stand on, and the odds of a confident invention rise. The quality of the source environment is not a side issue; it is the main variable.
03 What a brand can actually do
The honest answer is narrow but real: you cannot control the model, but you can improve what it has to work with. For your own category, that means being a clear, correct, well-structured source — stating the facts about your field plainly, so that when the model assembles an answer, the easiest and best-supported thing to say is the accurate one. It also means resisting the temptation to treat this as solved when a platform announces safeguards; those reduce the rate of obvious failures without changing the underlying behaviour, so the work of being a reliable source is ongoing, not one-time. None of this guarantees a correct answer on any given query — generation is probabilistic, and a hedge is not a promise. But moving a topic from "the only available sources are thin and sarcastic" to "the clearest, most credible voice says this plainly" measurably lowers the chance the model improvises something wrong. That is the realistic goal: not control, but making the truth the most available thing to find — which is the discipline the AC Group has practised for 27 years.
Reading AI answers about your category, skeptically
One practical habit follows directly from all this: check what the models currently say about your field, and read it the way you would read an confident stranger rather than an oracle. Ask the same questions a prospect would — what you do, how you compare, whether a claim about your category is true — and look not only at whether the answer is flattering but at whether it is right. Where it is wrong, ask why: is it pulling from a sarcastic thread, an outdated page, a competitor’s framing, or simply a gap where no clear source exists? The errors are diagnostic. A confident wrong answer about your category is usually pointing at a place where the credible, unambiguous source the model needed was simply not there for it to find, and that absence is itself the finding.
That turns the risk into a to-do list. Where the model invents because the sources are thin, the response is to publish the clear, correct version it was missing. Where it leans on something sarcastic or out of date, the response is to make the accurate account more prominent and current than the bad one. Where it simply gets your category subtly wrong, the response is to state the right version plainly enough to become the easiest thing to assemble. The pattern is the same each time: an error is a signpost to a missing or weak source, and the remedy is to supply the clear, credible version that should have been there, then make sure it is the most prominent and current account on the topic, not one buried beneath noise. None of this is a one-time fix, because the model and its sources keep moving, so the sensible cadence is to re-check periodically rather than assume a single pass settled it, the same way you would revisit any measurement of a moving system. Treat the AI’s answer as a live readout of how well the truth about you and your category is represented out there, in the sources it can actually reach — and treat every confident error as a specific, fixable gap rather than a verdict you are stuck with — a task on a list, not a sentence to accept, and one whose fix tends to help every model, not just the one you tested.
One question that sorts the honest from the overselling
If you take one practical thing from the glue-pizza episode into how you choose who to work with, make it this question: can you guarantee what the AI will say about us? Anyone who answers yes is either misunderstanding the technology or hoping you do. No setting, markup, or platform relationship lets a provider promise a particular output from a system that generates probabilistically and can be confidently wrong even with good inputs. The honest answer is the less comfortable one — no, we cannot guarantee it, but we can measurably improve the odds by making the clearest, most credible information about you the easiest thing for the model to find and assemble.
That distinction is not pedantry; it is the whole ethic of doing this responsibly. A provider who promises control will, at best, take credit for answers that were fine anyway and go quiet the moment the model says something odd. A provider who is clear that the work is influence over inputs and management of a standing risk is one whose claims you can actually trust, because they match how the systems behave. The episode that taught the public AI can be confidently wrong should also teach buyers to be skeptical of anyone claiming to have tamed that — and to value the partner who treats reliability as something to keep working at rather than something already solved.
AI reliability: quick answers
Can I stop AI from saying wrong things about my category?
Not directly, and anyone promising you control is overselling — but you are not powerless either, and the distinction matters. You cannot reach into a model and edit what it says, and you cannot guarantee any given answer. What you can influence is what the model has to draw on. A generative engine builds its answer from the sources available on a topic, and when those sources are thin, contradictory, or laced with jokes and sarcasm, it fills the gap with whatever sounds plausible — which is exactly how a decade-old Reddit joke became culinary advice. So the lever is the source environment around your category: clear, correct, well-structured information published by credible sources, including you, gives the model solid ground to stand on instead of guesswork to improvise from. It is influence over inputs, not control over outputs. That is less than you might want, but it is real, and it is the only honest version of managing this risk.
Isn’t this just a temporary bug Google will fix?
Google did add safeguards after the May 2024 episode, and the most absurd results became rarer — but treating hallucination as a one-off bug misreads what it is. The strange answers were not a glitch in an otherwise truthful machine; they were the visible edge of how these systems normally work. A generative model produces text that is statistically plausible given its inputs, not text it has verified to be true, so a confident wrong answer is the same kind of output as a confident right one — the model cannot tell them apart from the inside. Patches can reduce the rate and catch the most obvious failures, and they have. But the underlying property — fluent, confident generation that can be wrong — does not get "fixed," because it is not a defect bolted onto the system; it is the system. The responsible stance is to treat reliability as a standing risk to manage, not a problem someone else has closed on your behalf.
Does good content guarantee the AI gets it right?
No, and it is important to be honest about that rather than sell certainty. Even with excellent, clear sources available, a model can still misread, misattribute, or blend them into something wrong, because generation is probabilistic and the engine is summarising, not quoting. What good content does is shift the odds: it makes the correct answer the easiest, best-supported thing for the model to assemble, and it reduces the room for the model to reach for a joke or a bad source to fill a gap. Think of it as lowering the probability and the severity of error, not eliminating it. That is genuinely worth doing — moving from "the only sources on this are thin and sarcastic" to "the clearest, most credible source says this plainly" measurably improves what AI tends to produce. But it is a hedge, not a guarantee, and the moment a vendor promises guaranteed accuracy is the moment to be skeptical.
Should I avoid AI visibility because of this risk?
That would be the wrong lesson, and an expensive one. The risk that AI can get things wrong is real, but stepping back does not remove it — the AI will still answer questions about your category whether or not you are a clear source in the mix; it will just do so with you absent, drawing on whatever else it finds. Absence does not protect you; it cedes the ground. The episodes that went viral are an argument for engagement, not retreat: the better and clearer the credible information available about you and your topic, the less room there is for the model to improvise something wrong, and the more likely your accurate account is the one it leans on. The right response to a system that can confidently invent is to make sure the truth is the most available, best-structured thing it can find — which is the same work as earning visibility, approached with clear eyes about its limits. We have helped clients do exactly that for 27 years.
A note on sources and timing
This is written in June 2024, weeks after Google’s AI Overviews launched to US searchers and immediately produced the glue-on-pizza and eat-a-rock answers that spread across social media. We have described what was documented then: the specific errors, their origin in sarcastic forum posts and satire, and Google’s response, in which Liz Reid acknowledged that some odd results were genuine and announced safeguards against treating humour and user-generated content as fact. We have deliberately not cited the later studies that put numbers on hallucination rates, or claims that newer systems hallucinate more, because those came afterward and would misdate a June 2024 view. What is solid now is the structural point the episode made plain: these systems generate confident text that can be wrong, the quality of available sources shapes the outcome, and reliability is a risk to manage rather than a solved problem — a stance the AC Group has held about every information system for 27 years.