- Two German courts have now treated an AI generated search answer as the search company's own statement rather than a summary of other people's pages.
- The Munich injunction of 28 May 2026 carries a penalty of up to 250,000 euros per further violation, and Google was ordered to bear 80 percent of the costs.
- Neither ruling is final. The Munich case settled before appeal, so no higher court has tested the reasoning.
- A US court reached the opposite result in 2025, and one reason was that the person affected had never asked for a correction first.
- Whatever the jurisdiction, the same two things decide your position: a dated record of what was said, and a written request to fix it.
In late August 2026 a German doctor was told by a colleague that Google's AI search was linking him to a sexual assault allegation and to court proceedings that did not exist. The allegations belonged to a different person. He emailed Google. In mid September Google replied that it would take no action, and ignored a formal legal demand sent the same day. By the time a court looked at it, the text was still being produced.
On 24 September the regional court in Aachen granted an injunction without even hearing Google first. Most small businesses will never face anything that serious. Plenty will face the smaller version: an answer that says you closed down, that you charge a fee you do not charge, or that you are the subject of a complaint that belongs to a company with a similar name. The question that used to have no good answer is what you can actually demand, and from whom.
Can you hold the company responsible for what its AI says?
In Germany, on current rulings, yes. Two regional courts have now decided that an AI generated answer is the operator's own speech, which removes the defence that has protected search results for twenty years.
The reasoning matters more than the outcome, because it is the part that might travel. The Munich court set out four grounds in case 26 O 869/26, decided on 28 May 2026. The feature presents a structured thematic answer rather than a list of links. The operator built it and alone controls the algorithms behind it. The claims were original, in the sense that a plain search for the company name produced nothing of the kind and the cited sources did not support them. And the court held that traditional search is essential infrastructure for finding things online, while an AI summary is supplementary, so it does not inherit the same protection.
The plaintiffs were two Munich publishers whose names, searched alongside the German word for fraud scheme, returned summaries describing subscription traps and invoices for services never delivered. Some of the links the summary drew between those companies and others appeared in none of the pages it cited.
Why did the AI label not help?
Because the court decided a label does not move the duty to check onto the reader. It leaned on a German idea of the reader who absorbs only the headline level impression, and concluded that someone who sees a confident paragraph has no particular reason to click through and audit it.
Google's other defences went the same way. Being a hosting provider under the EU's Digital Services Act did not help, because that framework expressly preserves a court's power to order an injunction, and because passive hosting is a poor description of content your own system wrote. The argument that linking to sources was itself sufficient notice failed for the same reason the label failed: the answer presents itself as complete.
What exactly was ordered, and is it settled law?
An injunction against repeating specific statements, backed by a fine of up to 250,000 euros for each further violation, with 80 percent of the costs falling on Google. The Aachen court later barred three disputed variants of the text about the doctor, with the same ceiling.
It is not settled. Both were preliminary injunctions. Google announced an appeal in the Munich case and then settled instead, which means no higher court has examined the four grounds above. Neither side has said what the settlement contained. Anyone telling you the law is now decided is reading more into two first instance orders than they hold.
What they do establish is a direction, and a cost. A company that can be enjoined and billed has an incentive to answer a complaint it would otherwise file away, and that incentive is the practical thing a small business can use.
Does any of this help outside Germany?
Less than you would hope, and a 2025 American case shows why. A Georgia court granted summary judgment to OpenAI over a ChatGPT output that falsely accused a radio host of embezzlement.
The court gave three reasons, and the first is the one that clashes directly with Munich. Under Georgia law a statement must be reasonably readable as describing actual facts, and the court held a reader in that position could not have taken the output as factual, citing the product's own disclaimers and the fact that the journalist who saw it checked and knew within about ninety minutes that it was false. Where the German court said a disclaimer does not shift the burden, the American court treated the disclaimers as part of why no reasonable reader would be misled.
The second reason was that no standard of care was established, and the court declined to treat shipping a model with a known error rate as negligence, calling that a strict liability theory the state does not recognise. The third was damages: the plaintiff admitted he had suffered none and was seeking none.
Buried in that third point is the most portable lesson in either case. Punitive damages were unavailable partly because the plaintiff never asked OpenAI to correct or retract the output, which Georgia law requires first. The cheapest legal step available to you is also the one that preserves your options, and it is a written request for a correction.
What should you capture before you complain?
Everything that proves what was said, when, and to whom it was shown. An AI answer is not a page: it is generated afresh, it varies by wording and region, and it can disappear while you are writing your complaint, which is why the record has to be made first.
Two of those deserve a note. Recording whether a plain search for your business name produces the same claim matters because originality was one of the four Munich grounds: the court cared that an ordinary search did not return it. And checking a second device or account matters because an answer shown only to you is a much weaker complaint than one a stranger can reproduce.
Keep the cited pages too. If the claim came from a real page that is genuinely wrong about you, your fastest fix is that page rather than the summary, and the routine for that sits in correcting what assistants say about your business.
Which route actually gets something removed?
There are three, they are not interchangeable, and most people only ever use the weakest one. Here they are side by side, with what each is actually for.
| Route | What it is | What it achieves | When to use it |
|---|---|---|---|
| In product feedback | The thumbs down and Report a problem control on the answer itself | Trains the feature. No case number, no reply, no record you can rely on | Always, as a first click, and never as your only step |
| Legal removal request | Google's Report Content for Legal Reasons form, filed per product | Produces a confirmation email with a reference number and a human review | When the claim is defamatory, false or breaches your rights |
| Formal written demand | A letter or email stating the claim, the evidence and a deadline | Creates the notice that later matters if a court looks at whether the operator had the chance to act | When the first two produce nothing, or the harm is active |
| Court application | An injunction against repeating specific statements | A binding order with a per violation penalty attached | Rarely, and with advice, when the harm justifies the cost |
The ordering is not bureaucratic politeness. In Munich, the plaintiffs had written on 2 February and used the reporting form, and a search on 10 February still returned substantially similar content. That gap between notice and inaction is what let the court find the operator had its chance. In Georgia, the absence of any correction request narrowed what the plaintiff could recover.
What does the in product route actually do?
It sends feedback, and Google's own help page describes it as a thumbs down followed by Report a problem and a category, with your most recent query attached. The same page states plainly that AI responses may include mistakes and that the overviews can and will make mistakes. Useful, and not a remedy.
And the legal form?
It is a different pipeline with a different output. The legal reporting page asks you to pick the product and the reason, give the exact URL rather than a homepage, and explain specifically what is unlawful and why. It promises a confirmation email with a reference number, warns that you may be asked follow up questions, and tells you to file separately for each product the content appears in. That last instruction is easy to miss and matters here, because a claim can surface in more than one surface of the same search product.
What if the claim is partly true, or came from your own old page?
Then you are not in defamation territory and the legal route will not help you. A summary that reports an accurate but outdated fact, a closed location or a discontinued product, is a data problem with a data fix.
This is the commonest version by a distance, and it usually traces back to something you control: a stale listing, an old price on a page nobody deleted, an archived announcement that still ranks. A removal request aimed at the summary will be refused, correctly, because the summary is doing its job on the material available. The work is to change the material. Our own policy on how we use and disclose AI exists for the same reason: what a machine can find about you is mostly what you published.
The grey case is a claim that is technically defensible but misleading in context, which is exactly the kind of thing the Munich court objected to when it noted that connections drawn between companies appeared in none of the cited pages. If the individual facts are right but the conclusion drawn from them is not, say precisely that in your complaint, because a reviewer skim reading a vague objection will close it.
What should the written demand actually say?
Four things, in plain language, and no more. A long letter gets skimmed; a short one with the right elements gets actioned or creates a usable record when it is not.
State the exact words complained of, reproduced verbatim rather than paraphrased, with the query that produced them and the date. Say why each statement is false, in one sentence each, and attach whatever proves it. Say what you want, which is normally that the specific statements stop being produced, not that the feature stops mentioning you. And give a date by which you expect a reply.
Keep the tone flat. The German rulings did not turn on how strongly anybody objected; they turned on whether the operator had been told and had the opportunity to act. A letter that is calm and specific does that job better than one that is angry and broad, and it reads better later if a third party ever has to assess whether your complaint was reasonable.
One thing to avoid: do not threaten anything you have no intention of doing. That is a bad habit generally, and in the review context it is also a regulated one in several places, since groundless legal threats used to suppress criticism are themselves a prohibited practice in both the UK and the US regimes on consumer reviews.
How often is this actually the operator's fault?
Less often than the aggrieved version of the story suggests, which is why the triage matters before the letter does. Most wrong answers about a small business trace to a wrong input rather than an invented claim.
The Munich case is useful precisely because it is the uncommon kind. The court made a point of recording that a plain search for the company name did not return the claims, and that some of the connections drawn between companies appeared in none of the cited pages. That is the signature of something generated rather than repeated, and it is the test to run first on your own case.
The timeline is the other detail worth copying. Searches on 20 January, 26 January and 10 February 2026 all produced substantially similar content, with a formal letter sent on 2 February in between. Three dated observations spanning three weeks, with a written notice in the middle, is a far stronger record than one screenshot, and it is the difference between a complaint that can be dismissed as a one off glitch and one that shows persistence after notice.
If your own checks show the claim does appear in a source page, your effort is better spent on that page. If the source is a third party you cannot edit, a directory entry or an old news item, then you are in the slower work of getting that corrected, and the summary will usually follow once the underlying material does.
Does this cut the other way, for your own bot?
It does, and the symmetry is worth sitting with before you demand much from anybody else. The argument that an AI generated paragraph is the operator's own statement is not reserved for large companies.
If you run an assistant on your site that tells a customer something false about a product, a price or a competitor, the same reasoning points back at you, and in at least one tribunal it already has. We went through that in what happens when your chatbot makes a promise, and the practical conclusion is the same in both directions: keep the transcript, answer the complaint in writing, and fix the source rather than the sentence.
What none of this gives you
It does not give you speed. The German doctor was alerted in late August and had an order on 24 September, with a refusal in between, and that was a serious case pursued with lawyers. A small business complaint about a wrong opening time will move at the pace of a support queue.
It does not give you a guaranteed removal either. Both orders bar specific statements, not categories, which is why the Aachen court listed three variants of the text rather than one. A system that generates language afresh each time can produce a near miss that nobody has enjoined.
And it does not give you certainty about the law. Two preliminary injunctions in one country, one of them settled before appeal, sitting against an American summary judgment that reasoned in the opposite direction, is not a settled position. What it is, is enough to make the written complaint worth sending, and enough to make the dated screenshot worth taking before you send it.