Getting cited is not the goal if the citation is false
Every GEO conversation is about getting mentioned. Almost none is about what the mention says. But an AI that confidently tells a buyer something false about your product has not helped you — it has handed you a reputation problem you cannot see, in a channel you do not control.
the short answer
Being cited is worthless if the citation is wrong. AI hallucinations can attribute false claims to your brand, invent quotes from your executives, or misstate your product — in the same confident tone as the truth. They are not rare (2024 studies found 58-88% on legal questions), and some newer models hallucinate more, not less. The defence is editorial: monitor what AI says about you, and publish clear, sourced canonical content that gives models true facts to ground on instead of inventing.
key takeaways
- The goal is not just to be cited but to be cited accurately: hallucinations can attribute false claims to your brand, invent executive quotes, or misstate your product.
- Hallucinations are neither rare nor trivial: 2024 studies found 58-88% on legal questions and GPT-4 inventing 28.6% of citations in medical references.
- Counterintuitively, some of the newest, most capable models hallucinate more, not less — up to 48% on one 2025 benchmark; fluency improves faster than accuracy.
- The harm is real: defamation (ChatGPT falsely accused a mayor of bribery), bad decisions, eroded trust, and PR crises.
- The defence is editorial, not technical: monitor what AI says about you, and publish clear, well-sourced canonical content that reduces the ambiguity hallucinations feed on.
hallucination is not rare — measured rates by context
In words, so the bars do not carry it alone: measured hallucination rates make clear this is not an edge case. A 2024 Stanford study found general-purpose chatbots hallucinated on 58 to 88 percent of legal questions. A 2024 medical study found GPT-4 invented 28.6 percent of the citations it produced for systematic-review references. And a 2025 benchmark found newer reasoning models hallucinating up to 48 percent of answers on a fact-retrieval task. Different domains, different methods, one conclusion: a meaningful share of confident AI answers are simply false, and your brand is not exempt from being the subject of one.
Why this is a brand problem, not just a tech one
It is easy to file hallucination under "AI is still immature" and move on, but that framing misses where the cost lands. When a model invents a claim about your brand, the damage is yours, not the model maker’s. A buyer who asks an engine about your product and is told something false — that it lacks a feature it has, that it failed a test it passed, that your founder said something they never said — makes a decision on that falsehood, and does so with no idea they were misinformed. You may never learn it happened. There is no server log on your side, no referral telling you a sale was lost to a sentence a machine made up. The harm is real and the visibility is near zero, which is the worst combination a reputation risk can have.
The cost is not hypothetical. A model once fabricated a bribery accusation against an Australian mayor who was in fact the whistleblower in the case, bringing him to the edge of a defamation suit against the model’s maker. Courts have sanctioned professionals who relied on invented citations, and a database of such cases has grown steadily as more surface. For most brands the stakes are quieter than a lawsuit but no less corrosive: a slow drip of subtly wrong claims, each one shaping a buyer’s impression before you ever enter the conversation. Treating that as someone else’s engineering problem is how it goes unmanaged.
The accuracy problem, in three parts
How AI gets your brand wrong, why newer models do not automatically fix it, and what you can actually control. Open each layer for the part that changes how you manage the risk.
01 How AI gets your brand wrong
A hallucination about your brand is not one failure but several. It can fabricate — invent a product feature, a partnership, or a statistic that never existed. It can misattribute — put words in your executive’s mouth, assign you a stance you never took, or credit you with a competitor’s misstep. And it can misstate — get a real fact subtly wrong in a way that sounds plausible, which is the most dangerous kind because no one thinks to check it. The unifying thread is confidence: a model presents the false claim in exactly the same fluent, authoritative tone as a true one, so a reader has no signal that anything is off. The error does not look like an error. That is what makes it a brand problem and not just a technical curiosity.
02 Why newer models can be worse
The intuitive assumption is that hallucination is a teething problem that each model generation cures, and the data unsettles it. Some of the newest, most capable reasoning models have posted higher hallucination rates than their predecessors, not lower — one 2025 benchmark put a newer model near half its answers wrong on a fact-retrieval task. The reason is that fluency and accuracy are different axes, and they improve at different speeds. A model can get better at sounding right faster than it gets better at being right, which makes its mistakes more convincing, not less. For a brand, the implication is uncomfortable: you cannot assume the problem is shrinking on its own as the engines upgrade. The confident wrong answer may get more confident before it gets more correct.
03 What you can actually control
You cannot control whether a model hallucinates, but you can shrink the space it improvises in and shorten the time a false claim survives. The space shrinks when the clear, authoritative facts about you are easy to find, consistent across sources, and stated plainly — a model grounding on solid facts invents less than one filling a vacuum. The time shortens when you are actually watching: monitoring what the engines say about you, so a damaging claim is caught while it is one answer rather than after it has spread. Neither is glamorous, and neither is a guarantee. But between them they convert hallucination from a thing that happens to you in the dark into a risk you can see and manage, which is the difference between a brand that gets blindsided and one that does not.
Measure accuracy, not just presence
Most AI-visibility work counts mentions: how often you appear, in which engines, against which competitors. That is necessary and incomplete, because a mention can be wrong, and a wrong mention is worse than no mention at all. The missing measurement is accuracy — not just whether the engines name you, but whether what they say is true. It is a different, harder metric: you cannot fully automate it, because judging whether a claim about your product is correct still needs someone who knows the product. But it is the metric that maps to real risk, and the brands that track only presence are watching the wrong number while the dangerous one goes unmeasured.
In practice this means reading the actual answers, not just the dashboards. Pose the questions your buyers ask — about your product, your pricing, your people, your comparisons — to each major engine, and check the claims against the truth, engine by engine, because they diverge and drift. The point is to find the false claim while it is still a single answer, before it propagates into the sources other models train on and hardens into a consensus error that is far harder to undo. Accuracy monitoring is unglamorous and partly manual, which is exactly why most brands skip it — and why the ones that do it catch problems the others never see.
The defence is clarity, not control
You cannot edit a model’s output, but you can change what it grounds on, and that is where the real defence lives. Hallucinations feed on ambiguity: when the authoritative facts about your brand are scattered, inconsistent or simply absent, a model fills the vacuum with a confident guess. Remove the vacuum and you remove much of the temptation. That means canonical pages that state your key facts plainly and consistently, question-and-answer blocks that directly address the questions where errors are most damaging — what your product does and does not do, who your people are, how you compare — and the same facts repeated, unchanged, across the sources a model is likely to read. Give the engine something true and easy to lift, and it reaches for invention less often.
When a damaging error does appear despite all that, speed is the whole game. A false claim caught early, with a clean canonical source ready to point to, can be corrected before it spreads; the same claim found late, after it has seeded a dozen downstream repetitions, becomes a slow and expensive cleanup. The brands that handle this well are not the ones that argue hardest about whose fault the hallucination is — they are the ones that decided the risk was theirs to manage, built the clear sources and the monitoring in advance, and treated a false AI claim the way they would treat any other reputational incident: as something to detect fast and correct faster.
When it has already happened: a correction playbook
Detection is the first move, and it has to be specific: confirm the exact false claim, which engine produced it, and for which prompts it appears, because a fix you cannot reproduce is a fix you cannot verify. Then point the truth somewhere durable — a canonical page on your own site that states the correct fact plainly, with a date and a source — so that every later step has something authoritative to reference. Where a platform offers a feedback or correction channel, use it, citing that canonical page; where it does not, the canonical source still matters, because it is what the model and its downstream copies can eventually re-ground on. Speed compounds here: the earlier you act, the fewer places the false claim has been repeated into.
Then document and watch. Record what was claimed, where, and when, both as evidence if the matter ever turns legal and as a baseline to measure whether your correction took hold. Re-run the prompts that surfaced the error on a schedule, because a claim that disappears from one engine can persist in another or resurface after a model update. The goal is not a single heroic fix but a standing process: a known canonical source, a habit of checking, and a short path from "we found a false claim" to "we have corrected the record." Brands that treat AI misinformation as an ongoing operational risk, rather than a one-off surprise, are the ones that keep it small.
When AI gets it wrong: quick answers
Can I sue if AI defames my brand?
It is a live and unsettled area, not a clear yes. There have been real cases — an Australian mayor threatened a defamation suit after ChatGPT falsely described him as guilty of bribery when he was in fact the whistleblower — and regulators are paying attention to consumer-harm angles. But the law is still working out who is liable when a model invents a damaging claim: the AI company, the user who published it, or no one. Litigation is slow, expensive, and uncertain, and a lawsuit does nothing about the next answer the model generates tomorrow. For most brands the practical response is not legal first but operational: detect the false claim quickly, correct the source material that feeds it, and document everything. Treat the legal route as a backstop for the most damaging cases, not as the primary remedy, because by the time a case resolves the misinformation has already done its work.
How do I find out what AI says about me?
You ask it, systematically, the way your customers would. Most brands have never run the obvious test: pose the questions a buyer would ask about your category and your company to each major engine, and read what comes back — not just whether you are mentioned, but whether what is said is true. Check the claims about your products, the quotes attributed to your people, the comparisons with competitors, the basic facts about what you do. Do it across engines, because they diverge, and repeat it, because answers drift as models and sources change. This is monitoring, not a one-time audit: the goal is to catch a damaging false claim while it is one answer rather than after it has propagated into a hundred. Visibility tools that track mentions are a start, but the part that matters most — accuracy — usually still needs a human reading the actual answers.
Does good content reduce hallucinations about my brand?
It helps, though it cannot eliminate the risk, and the mechanism is worth understanding. Hallucinations thrive on ambiguity: when the clear, authoritative facts about your brand are scattered, inconsistent, or missing, a model fills the gap by guessing, and a confident guess reads exactly like a fact. Publishing canonical pages that state your key facts plainly, consistently and with sources gives the model something true to ground on instead of inventing. Question-and-answer blocks that directly address the high-risk questions — what your product does, what it does not do, who your executives are — reduce the space where a model improvises. None of this guarantees accuracy, because the model can still err, but it lowers the odds and gives you a clean source to point to when you need a correction. Clarity is not a cure; it is the best available prophylactic.
Is this the AI company’s problem or mine?
Morally, theirs; practically, yours — and confusing the two is how brands get hurt. It is entirely fair to say the model should not invent claims about you, and the companies building these systems carry real responsibility for reducing hallucination. But fairness does not protect your reputation in the meantime. The false claim reaches your customers whether or not it is the AI maker’s fault, and waiting for them to fix the underlying model is waiting on a timeline you do not control. The mature stance is to hold both truths at once: push for better systems and accountability at the industry level, and own the operational work of monitoring and correction at the brand level. The brands that do well here are the ones that stop arguing about whose fault it is and start managing the risk as theirs to manage.
A note on sources and certainty
The figures here — 58 to 88 percent on legal questions, 28.6 percent of invented medical citations, up to 48 percent on a newer-model benchmark — come from 2024 and 2025 studies of hallucination across legal, medical and general domains. We present them as evidence that the problem is real and non-trivial, not as a single rate that applies to your brand: actual risk varies enormously by query, domain and model, and these numbers are illustrative of scale rather than predictive of your exposure. The rates will shift as models change, and not always downward. What is durable is the underlying point: confident AI answers are sometimes false, the harm of a false one about your brand lands on you, and the only parts you control are clarity and vigilance. The AC Group has spent 27 years arguing that reputation is built on what others can verify about you; this is that argument meeting a channel that will confidently say things about you that are not true.