When AI Gets Your Business Wrong: The Correction Playbook

This Is Already Costing Businesses Money

The clearest evidence is in Google's own support forums, where business owners have been describing the same problem for over a year. A pest-control company found an AI Overview serving another firm's pricing under its name. A business found its phone number attached to a former client. Another watched an AI Overview attribute a competitor's complaints to it. The threads recur every few weeks, and the responses follow a pattern: try the feedback button, claim your Business Profile, keep your details consistent, add structured data.

In one of those threads the business owner replied that he had already done all of it, that his profile had been in place for years, and that the problem had "impacted my business grossly." The thread was locked and tagged as a non-issue. In another, a Google product expert answered honestly: there is no published threshold for how many feedback submissions change anything, and repeated feedback "can help, but there is no guarantee or timeline."

The harm is not hypothetical. In the most-reported case, a Minnesota solar installer sued Google after AI Overviews stated it had been sued by the state attorney general, which had not happened. The company says a six-figure contract was cancelled as a result. That case has produced one order so far, a procedural one, sending it back to state court after Google missed a deadline. Nothing has been decided about whether Google is liable.

First: Is It Real, or Is It Variance?

Before escalating anything, establish that the claim is persistent. A single bad screenshot is close to worthless as evidence of a systematic problem, because these systems are non-deterministic by construction. The cause is not the temperature setting most people blame. It is that inference kernels are not batch-invariant, so the numerics of a forward pass shift with server load. Even at temperature zero, what you get can depend on who else happened to be querying at the same moment.

Our own measurements put numbers on how much this moves. Across repeated runs, whether a brand was mentioned at all flipped 6.8% of the time between consecutive measurements, while the sentiment framing flipped 45.5% of the time, roughly 6.7 times noisier. Around 77.5% of brand-prompt-engine cells were stable. That gives you a usable threshold: a one-off unflattering answer is probably variance, and a false factual claim that survives repeated runs is a real problem.

So measure a rate, not an anecdote. Fix a prompt set, run each prompt several times across the engines you care about, log the model version and the date, and record the share of runs in which the false claim appears. Our guide to checking what AI says about you covers the method. This matters practically as well as evidentially: without a before-and-after rate you will never know whether anything you did worked.

The Rights Gap: Why There Is No Correction Button

Here is what actually exists across the major assistants, and it is a thinner set than most advice implies.

Two details are worth drawing out. First, the OpenAI path is a personal-data right, and it filters rather than fixes: the argument made by the privacy group that first pressed this point is that the system has no means of correcting false information and can only hide it at the output stage. Second, Anthropic states plainly that Claude "can write things that might look correct but are very mistaken" and that users should not treat it as a single source of truth, which is candid, and is also not a remedy. Its formal privacy-request path covers personal data, and is disarmingly clear that models "do not store text like a database." Microsoft is equally direct that reporting a concern will not necessarily result in removal, and points you back to whoever published the page.

The exception is Google's legal removal troubleshooter, which is the only channel we found anywhere that names AI Overviews and AI Mode alongside a defamation category covering a business. It is a legal notice reviewed against local law rather than a support ticket, and filing one is not the same as removal. It is still the strongest formal lever a company has.

What the Courts Have Actually Done

Coverage in 2026 has suggested the law shifted decisively in favour of the wronged party. The underlying documents say something more measured, and the difference matters if you are deciding where to spend effort.

On 24 July 2026 a Delaware Superior Court judge denied Google's motion to dismiss a defamation suit over AI-generated statements. The opinion calls the area a new frontier for defamation law, and then declines to explore it: the court states that the motion "does not force the Court to take a deep dive into those novel issues" and resolves it on established defamation caselaw under a pleading standard that asks only whether recovery is conceivable. Google assumed for the purposes of the motion that an AI output can be defamatory, and the court recorded that it was taking no position on that question. So the widely repeated claim that a court has held AI output to be a publication, or to be capable of defamation, is not what happened.

The counterweight is Walters v. OpenAI, where a Georgia court granted OpenAI summary judgment, finding no defamatory meaning, no fault even at the negligence level, and no damages. Taken together the record is: one settlement with no ruling, one procedural remand, one summary judgment for the AI company, and one case that survived an early motion. No United States court has held an AI developer liable for defamation. Whether Section 230 immunity extends to generative output has not been decided either.

The regulatory picture is similarly narrower than the headlines. The Federal Trade Commission published a proposed policy statement on accuracy in AI systems on 7 July 2026, with comments closing on 31 July. It is proposed rather than final, it concerns companies steering outputs toward undisclosed objectives, and a footnote excludes ordinary hallucinations explicitly, saying the Commission does not believe they raise issues under Section 5 on their own. It is aimed at model developers on behalf of the people using them, and Section 5 gives a private party no right to sue. If your business is the subject of a false answer, this statement offers you nothing directly.

None of the above is legal advice. It is a summary of what specific courts and one agency did, on specific dates. Whether any of it applies to a particular situation is a question for a lawyer, and businesses facing serious, repeated, documented harm have generally consulted defamation counsel early.

The Escalation Ladder

Work these in order. The early tiers are cheap and low-yield; the later ones are slower and carry more weight. Most published advice stops at the feedback button and the source fix, which is why so many of those forum threads end in frustration.

Tier 2, in practice

This is where most of the real progress happens, because a grounded answer is downstream of a page. Identify the page, which is easiest on the assistants that show citations inline, then get that publisher to correct or remove it. Once they have, use Google's refresh tool to push the index rather than waiting for an organic recrawl. Note the asymmetry that makes it useful here: it works only on pages you do not own, and only once the page has actually changed. For other engines, IndexNow notifies several at once, though Google is not among them and no engine guarantees indexing. Our breakdown of where AI citations come from helps with the tracing step.

Tier 3, and the Wikipedia trap

The entity record is what the machines read when they need to know who you are, so a wrong statement there propagates. Correcting it is not the same as building it, and the rules are strict. Wikipedia's conflict-of-interest guideline strongly discourages editing an article about your own company directly; the sanctioned route is a talk-page edit request. Paid editing must be disclosed, and the guideline is blunt that there is no confidentiality for the client. On Wikidata the mechanic that matters is retiring the wrong statement by setting it to deprecated rank, not simply adding a correct one beside it and hoping. If you have no entity record at all, that is a different job, and our entity and knowledge-graph guide covers building one.

Why Fixing the Source May Still Not Be Enough

Almost every guide to this problem ends at "fix the sources." It is the right instinct and it is where your leverage is. It is also not a guarantee, and the reason is uncomfortable.

If a majority of correct answers are not actually supported by the pages cited beside them, then the relationship between a source and a sentence is looser than the interface suggests. Correcting the page you were shown may not change the claim, because the claim may not have come from that page. Google contests the analysis, and the methodology argument is worth taking seriously. The practical takeaway survives it either way: treat source remediation as the highest-leverage move available, not as a switch.

Two adjacent beliefs are worth retiring while we are here. Structured data is not a correction mechanism: Google's own documentation states there are no additional requirements to appear in AI Overviews or AI Mode and no special schema you need to add. And publishing a large volume of corrective content is now an explicit risk rather than a clever tactic, since Google's spam policies were amended in 2026 to name attempting to manipulate generative AI responses. One clear, well-sourced, dated correction page is defensible. Fifty variations of it are not.

What Nobody Actually Knows

Plenty of confident numbers circulate on this topic. Having gone looking for the studies behind them, most trace to vendor posts with no published methodology. These are the honest gaps:

• 01How long a correction takes. No platform publishes a figure, and no credible independent study exists. Anyone quoting you a specific window is guessing.
• 02Whether in-product feedback ever fixes a specific answer. No platform commits to it and none publishes outcome data.
• 03Where corrective content becomes manipulation. The spam policy names the behaviour but sets no threshold and gives no examples.
• 04Whether entity work fixes wrong facts. The published research measures whether entity signals raise mention volume, not whether they correct false assertions.
• 05How to timestamp a private AI conversation. Archiving services capture public URLs. A chat session has none, and there is no established, tested method here.

The reason to say all this out loud is that the alternative is worse. A plan built on a made-up timeline fails on a schedule you cannot explain to anyone.

How Ranqo Helps

Ranqo does not correct anything, and no tool can. What it does is turn the anecdote into a record. Every tracked answer is stored with its full text, the sources the engine cited, the specific sentences that read negative, the platform, and the date. Run the same prompt set on a schedule and you get the thing every tier of the ladder needs and most businesses lack: the share of runs in which the false claim appears, dated, per engine, before and after you act.

That is the difference between telling a publisher, a platform, or a lawyer "an AI said something bad about us once" and showing them that it said it in most runs across two engines for six weeks, citing these three pages. It is also how you find out whether the fix worked, which given everything above is not a question you can answer by looking.