A wrong AI Overview is not a wrong search result. It is a sentence Google writes itself, at the top of the page, and people read it without opening a single source. When that sentence is about your company and it is false, you no longer have a ranking problem.
An AI Overview with wrong information is a Google-generated summary that attributes to a company, product or fact something the cited sources do not say, or that actually belongs to someone else. There are three origins: outdated data in the sources Google uses, mistaken identity between similar companies, or a summary that draws a conclusion from the texts that the texts do not contain. The first two can be fixed at source. The third cannot, and it changes the kind of work you need to do.
Below you will find how often it really happens, the procedure we follow when a client reports a false answer about themselves, the reporting channels that work and the ones that do not exist, and what changed in Europe after 28 May 2026.
How often AI Overviews get it wrong
The most solid data available comes from a study published by Oumi on 14 April 2026, picked up by the New York Times. The method: 4,326 factual queries taken from OpenAI's SimpleQA dataset, run through Google and assessed on two model generations, Gemini 2 in late 2025 and Gemini 3 in early 2026. The measurement was on the US market, in English.
The interesting result is not the error rate. It is the gap between two measures that are usually confused.
| Measure (Oumi study, April 2026, US sample) | Value |
|---|---|
| AI Overviews containing the correct answer | 91% |
| AI Overviews that are correct and fully supported by the cited sources | 39% |
| Individual claims not supported by the cited sources | 33% |
| Estimated false or misleading answers per day (New York Times, based on around 5 trillion searches a year) | ~616 million |
| Users who open a source when an AI summary is shown (Pew Research Center, July 2025, US sample) | 8% vs 15% |
Ninety-one per cent correct answers and thirty-nine per cent verifiable answers are two numbers about the same object. The first says the model usually knows the answer. The second says that in around half of cases it writes it while leaning on sources that do not say it. For a company the second one matters, because the invented sentence about you is born in the gap between the two.
Google disputed the study. Its spokesperson Ned Adriance, quoted by TechRepublic, said that "this study has serious holes" and "doesn't reflect what people are actually searching on Google". The objection is reasonable: SimpleQA is a dataset of hard questions, not a sample of real traffic. Still, the order of magnitude has not been refuted with alternative data.
Your exposure is higher than you think
The largest single-country study we work with comes from Italy: SEOZoom published it on 2 December 2025, covering more than 30 million keywords and around 168 million domains visible on Google Italy. The numbers that change the risk calculation:
- 54.15% of Italian searches show an AI Overview box;
- 175,000 domains are selected as sources, 0.1% of visible sites;
- around 40% of cited sources do not appear in the organic top 10 for that query;
- health accounts for 40.3% of appearances, finance for 20.2%;
- Wikipedia appears as a source 864,517 times, YouTube 875,914.
That 40% is the point almost nobody comments on. Google builds the answer from pages that are not in the top positions too, so holding the top 10 is not enough to control what is said about you. The same pattern is easy to see on Google UK. For "what does b2b mean" the AI Overview opens with a definition cited to the US Chamber of Commerce and a YouTube video, while the organic top 10 is made up of Indeed UK, Reddit, Investopedia, Xero, Wikipedia and a handful of UK and SaaS blogs: the first source in the box is not in the organic results at all (DataForSEO, Google UK, September 2026). We cover this in detail in the guide to AI Overviews ranking factors and in the study on the impact of AI Overviews on SEO.
First-hand data from our own projects: on visilay.com, in the quarter from 20 May to 18 August 2026 and on Italian traffic only, the two Italian equivalents of "what does B2B mean" produced 59 impressions at an average position of 10.2 and zero clicks, and 54 impressions at an average position of 9.7, again with zero clicks. Being on page one for a definition question is worth less today than it used to be, and that is exactly the category of query where the AI box appears most often.
Five reasons Google writes false things about your company
Before reporting anything you need to work out which of the five cases you are dealing with, because the fix is completely different. This table is the part of the work that usually gets skipped, and without it you end up sending feedback that achieves nothing.
| Cause | How to recognise it | Where to fix it |
|---|---|---|
| Your own site still says the old thing | The wrong claim appears, word for word, on one of your pages that has not been updated in years | On the site: correct it, then request a crawl from Search Console |
| Google Business Profile out of sync | It concerns opening hours, address, category, phone number | In the profile, field by field |
| An outdated third-party source | The wrong sentence sits on a directory, a press round-up, an old press release | Ask the publisher of that page for a correction |
| Mistaken identity | The summary mixes you up with a company with a similar name, often in another sector | Organization structured data, explicit disambiguation in your copy |
| Unsupported inference | The cited sources are correct but do not contain the claim | Cannot be fixed at source: it needs reporting, and this is the case where the legal question matters |
The first four cases are routine data hygiene. The fifth is the one no SEO checklist solves, and it is also the most frequent in the reports we receive, because it is the one the company cannot explain.
The first two hours: document before you react
AI Overviews are not stable. The same query, repeated the next day from another device, can return a different box or no box at all. If you do not capture the evidence straight away, in two weeks you will have nothing to show either Google or a lawyer.
- A full-page screenshot with the query visible in the search bar, plus date and time.
- The exact query, written as it was typed, and the variant that triggered it if you tried more than one.
- The full list of sources Google shows in the box, with their URLs.
- For each source, a check of whether the disputed claim is there or not. This is the step that tells you which of the five rows in the table you are in.
- Repeat the search in an incognito window and from a second device, to see whether the box is served to everyone or is a one-off.
Step 4 is worth more than the other four put together. If the claim does not appear in any of the cited sources, you are not arguing about someone else's content that Google has indexed: you are arguing about a text Google wrote. Until June 2026 that was a technical distinction; since then it has become a legal one.
The reporting channels that actually exist
To put it plainly: there is no "correct this AI Overview" form. There are three real channels, with decreasing effectiveness and very different timescales.
Feedback inside the box. The three-dot menu on the AI Overview, then thumbs down and "Report a problem". It is instant, gets no reply and guarantees nothing. It is still worth doing: it is the record that shows you reported the issue before taking other routes.
Legal removals. Google's page on removals for legal reasons lists defamation among the accepted grounds, alongside copyright, personal data and court orders. AI features are not explicitly named among the products covered, which leaves room for interpretation that so far has worked in Google's favour.
Fixing it at source. In most cases this is the only route that produces a measurable change, and it is the work of the next section.
Google Search Central's official guide to AI features describes how the box is built: retrieval-augmented generation anchored to the classic ranking systems, plus query fan-out, the parallel queries the model generates on its own to gather more material. The same guide contains no correction or rectification mechanism. Google gives you a single lever: change what the system reads.
Fix the cause, not the symptom
Order matters, because each change needs a fresh crawl to take effect and the crawls do not all arrive together.
Make the fact explicit, not something to be inferred. If Google gets your sector wrong, somewhere on your site the definition is implicit. One flat sentence on an indexable page ("Visilay is a search marketing agency based in Como") is worth more than three paragraphs that hint at it. The reverse also holds: pages that contradict each other are ideal material for a wrong summary, which is why duplicate or near-duplicate content does damage beyond rankings.
Sort out your Organization structured data. Legal name, trading name, address, sector, sameAs pointing to your official profiles. In cases of mistaken identity it is the change with the best return on effort: you give the system a formal way to tell you apart from the company it confuses you with. The guide to Schema structured data covers the markup in detail.
Align your Google Business Profile. Opening hours, primary category, address, phone number. For errors on operational data the problem almost always starts there: see the Google Business Profile guide.
Search for the wrong sentence, not for your name. Search Google for the incorrect claim in quotation marks. If you find its source on a directory, a trade portal or an old press round-up, you have someone to ask for a correction, and fixing that page is often worth more than ten reports to Google.
Strengthen the trust signals on the pages that describe you. Who signs them, with what credentials, with what verifiable data. This is E-E-A-T territory, and it counts double here, because in the Italian study the system selected 175,000 domains out of 168 million.
Request a crawl, then wait. URL Inspection in Google Search Console for every page you changed. In the projects we manage, the box usually changes between two and six weeks after the fix, not in two days.
The Munich ruling: what changed for companies in Europe
On 28 May 2026 the Regional Court of Munich I (Landgericht München I) decided case 26 O 869/26, brought against Google by the publisher Verlagshaus24 and one of its subsidiaries. AI Overviews had linked the two companies to unfair practices and "subscription traps" which, the judges found, did not even appear on the sites shown as sources.
Google argued that it is a search engine displaying third-party content. The court rejected the argument with reasoning worth reading in full, available together with the text of the decision on the Transparency Coalition website: since the defendant's artificial intelligence compiles the information independently and summarises it in a text, this is an independent representation for which Google is responsible. Whoever introduces AI and offers it to users is also responsible for its output.
Google was barred from republishing the statements and ordered to pay 80% of the legal costs. It has announced an appeal, so the ruling is not final, and it is a German decision with no binding force in the UK. It still changes two practical things for a company operating in Europe: there is now a precedent in which the AI box was treated as the operator's own content, and the evidence gathered at step 4 of the procedure above (the claim is in none of the cited sources) is exactly the fact on which that judge based the decision.
Health and finance: where errors cost the most
In the Italian data, 40.3% of AI Overviews concern health. It is no coincidence that this is also where Google stepped in first: in January 2026, after an investigation by the Guardian, it removed the box from some health questions, for example on reference ranges for liver function tests, where the summary returned numbers without distinguishing by age, sex or ethnicity. Euronews found, however, that minor variants of the same question still generated the summary.
We checked this on Google UK. The query "liver function test normal range" still opens with an AI Overview that lists a single set of adult ranges taken from Mayo Clinic, with albumin in g/dL and bilirubin in mg/dL, while the NHS Specialist Pharmacy Service page and Geeky Medics, both in the organic top 10 below it, use the g/L and µmol/L units printed on UK lab reports. No split by sex or age (DataForSEO, Google UK, September 2026).
For a clinic, a professional practice or a financial firm the maths is simple: health and finance are the sectors where the box appears most often, and Google's policing consists of spot removals on individual queries. Monitoring stops being optional.
Three moves that make things worse
- Publishing positive content to bury the false claim. It does not work the way it did with ten blue links: the model summarises, it does not rank. You add noise and, at best, nothing changes.
- Blocking snippets out of spite. The Google Search Central guide ties eligibility for AI features to eligibility for snippets: with
nosnippetyou drop out of the cited sources, but the box can keep existing, built on third-party sources. You lose your voice and keep the damage. - Stopping at the first query that looks right. The box changes by query, not by topic. Check the variants too, including those with your name misspelt, because they are often the ones that carry the error longest.
Spot it before a customer tells you
A list of twenty to thirty queries is enough for most companies: the brand name on its own, the name plus "reviews", "legit", "scam", "contact", "who owns", plus the two or three product categories you are described with and comparisons with your direct competitors. Check monthly, archive a screenshot each time. AI visibility monitoring tools automate the collection, but the reading stays manual: no tool knows which claim about you is false.
It is worth extending the check beyond Google. The same wrong sources feed ChatGPT's answers and those of the other generative engines, and the approach is the same one described in the guide to answer engine optimisation and the one on generative engine optimisation. If you would rather hand the monitoring and correction work to someone else, it is what we do with our SEO for AI service.
Frequently asked questions
There is no general obligation and no dedicated form. On 28 May 2026, however, the Regional Court of Munich I, in case 26 O 869/26, ruled that the AI summary is Google's own content and barred it from republishing the false statements challenged by Verlagshaus24. The ruling is German, it is not binding in the UK and Google has announced an appeal, so it remains a precedent, not a rule you can already rely on.
In the projects we manage, the box usually changes between two and six weeks after the fix, and only once the edited pages have been crawled again. Requesting URL Inspection in Search Console shortens the first step but does not guarantee the summary will update.
You can drop out of the cited sources, not out of the box. The Google Search Central guide ties eligibility for AI features to eligibility for snippets in Search: blocking snippets takes you out of both. Google remains free to generate a summary about you from third-party sources, so you give up the chance to correct the story from the inside without removing the risk.
In the largest single-country study we use, by SEOZoom on 2 December 2025 across more than 30 million keywords tracked on Google Italy, the figure was 54.15%. The distribution is uneven: health accounts for 40.3% of appearances and finance for 20.2%. On Google UK in September 2026, both the definition queries and the health queries we checked opened with an AI Overview.
It depends on the cause. If the wrong claim exists in a cited source, you fix that source and legal action is not needed. If the claim appears in none of the sources shown and it is damaging, you are in the situation the Munich court ruled on, and it is worth having a lawyer assess it. We are not lawyers and this is not legal advice.
Something the industry rarely says: the data hygiene that prevents a false AI Overview is the same work that gets you into the cited sources. A company name stated unambiguously, facts made explicit rather than left to inference, consistent structured data, third parties brought into line. Whoever does it out of fear of one wrong sentence ends up, as a side effect, meeting the entry requirement for that small pool of cited domains. It is worth doing before you have the problem, not after.
