A Munich court held that AI Overviews are Google’s own words, and that Google is liable for the false ones.

This is the emergency page, not the program

Our AI reputation management service is an ongoing measurement and correction program. This is narrower. Something specific and false is being said about you by an assistant, somebody senior has seen it, and there is a meeting about it tomorrow.

What follows is what we would do in the first week, and what we would tell you honestly about your options.

Capture the evidence before it changes

This is the step people skip and the one that cannot be recovered later. AI answers are generated fresh each time. Ask again tomorrow and the wording moves. Nothing is stored on the engine side, so there is no history to subpoena and no log to request.

Before anything else: the exact prompt, the engine, whether search was on, the full answer text, and the date. Screenshots help for a meeting, and they are weak evidence because they are trivially edited. We store the raw text with metadata in your client portal, which is the record you will need if this ever becomes a regulatory complaint or a legal matter.

If you do only one thing after reading this page, do that.

Where the law actually stands

Two data points, pointing in different directions.

Walters v. OpenAI was the first US defamation claim over AI output to reach judgment. In May 2025 the Georgia court granted OpenAI summary judgment, holding that a reasonable reader in that context would not have taken the output as a statement of fact, and that the plaintiff had not met the required fault standard. Litigation is not currently a reliable route to a corrected answer.

The Munich ruling went the other way on a related question, treating AI Overview output as Google’s own statement rather than a neutral summary of somebody else’s, and holding Google responsible for false content in it. Different jurisdiction, different mechanism, and it is the strongest signal yet that the liability position is not settled.

If you are an individual rather than a company, there is a third avenue worth asking a lawyer about: jurisdictions with a right to rectification of inaccurate personal data. How that applies to a generated answer rather than a stored record is an open question and a live one.

None of this is legal advice. What we can tell you is that the supply-side work is available now, regardless of how the law settles, and that the two are not alternatives.

The four claims and what each needs

A stale fact. True once, not true now. A closed matter reading as open, a former director reading as current, a discontinued product reading as available. The most common, the most fixable, and the one where a clear dated correction on a page an engine can reach does most of the work.

A merged identity. Correct facts about somebody else who shares your name. Not a content problem at all. The fix is entity work: structured data, authoritative records, and a single canonical identity across every listing.

An invented specific. A detail with no source anywhere. A case that never happened, a certification never held, a figure nobody published. There is no wrong page to correct, so the fix is to make the correct fact easier to find than the gap that invited the invention.

A resurfaced old story. An accusation or dispute that was resolved, reappearing without the resolution, because the accusation was covered far more widely than the outcome was. The resolution needs to be as findable as the beginning, stated factually and dated. A rebuttal that repeats the accusation in its heading hands the engine a fresh source containing the claim.

What we will tell you

If the claim is retrieval-driven, we can usually show movement within weeks and Perplexity typically first. If it sits in training memory, nothing you publish removes it today and it waits for a model release on a date nobody outside the lab knows. We will say which one you have rather than let you assume the faster answer.

And if the damaging thing is accurate, we will tell you that in the first week. No amount of source work outruns a real operational problem, and you should hear it before the invoices start rather than after six months of them.