A Google AI summary told searchers that Anna Mae's Bakery and Restaurant in Millbank, Ontario, would shut for good on July 31. The business is not closing. Its owner spent the following days correcting the record, one phone call and one walk-in customer at a time.

The error surfaced over the weekend of July 18 and 19, 2026. According to CBC News, someone searched for Anna Mae's Bakery and the Google AI overview appeared with a statement that the Millbank business, just northwest of Kitchener, was officially closing on July 31. The summary was wrong. Owner Amanda Herrfort was on her way to a wedding when a screenshot of the search result landed on her phone.

"That's when I panicked," Herrfort told CBC News on Thursday, July 23, 2026.

What had happened, according to CBC News, was a case of mistaken identity at scale: Google had confused the Ontario restaurant with an Illinois bakery carrying the same name that is in fact closing. Two businesses, one name, one generated sentence, and a closing date attached to the wrong province.

A rumour with a search engine behind it

The correction problem was not theoretical. Herrfort launched an online campaign to dispel the rumour, and customers still emailed, called and walked through the door to ask whether the reports were true.

"We have a lot of people coming in talking to our employees saying: 'It says that you are [closing].' And then the employees would say: 'Well, no, we're not,'" Herrfort said, according to CBC News.

For a seasonal, tourism-dependent operation in southern Ontario, the timing of a phantom closure notice carries obvious commercial weight. Anna Mae's is described by CBC News as a popular destination for both tourists and locals. A visitor planning a drive from Toronto or Kitchener who reads that a restaurant shuts at the end of the month does not typically call ahead to verify. They pick somewhere else.

The mechanism of the fix was, in its way, as notable as the error. Herrfort eventually resolved the problem by typing a request into the Google AI search bar itself.

"I said: 'I would like this fixed please.' And then it wrote back saying: 'Oh yes, I can see how this could be detrimental, and we will get this fixed as soon as we can and I will make sure the correct information is given going forward.' It's like, OK, I hope this works," Herrfort said, according to CBC News.

That exchange describes a conversational interface responding with an apology and an assurance. It does not describe a documented remediation channel with a ticket number, a service level, or an appeal. No formal reporting path for factual errors in AI overviews is identified in the CBC account.

What Google told CBC

A Google spokesperson told CBC News on Thursday that the "vast majority of AI overviews are high quality and meet our high bar for helpfulness and accuracy."

The spokesperson added: "Just like other features in search, issues can arise when there is an absence of high quality information on the web on a particular topic and we use these examples to improve AI overviews broadly."

Google said the accuracy of AI overviews is similar to other search features and that mistakes can happen given the billions of searches processed daily, according to CBC News. The company said it takes action when necessary and learns from those mistakes. In small type beneath its AI summaries, Google warns that "AI responses may include mistakes."

The framing is familiar. It locates the cause in the web rather than in the summarisation layer, and it measures error against volume rather than against consequence. For a single restaurant in a village of a few hundred people, aggregate quality rates are not the operative statistic.

'Real consequences for real people'

Christine Logel, a professor of social development studies at the University of Waterloo who recently completed research on large language model companionship, offered a blunter reading to CBC News. Logel said that despite the training and research invested in systems such as Google's AI overview, they were released before being fully tested.

"That means things like this [mistake with Anna Mae's Bakery] are going to happen," Logel told CBC News.

She went further on the pattern: "I think people are learning the hard way ... you see this still happening in science and in law where there's hallucinated cases that never actually existed. It is kind of a buyer beware thing."

And on the distribution problem: "When they're integrated so much through the internet and in so many different products, when they make mistakes, those have real consequences for real people."

Herrfort, for her part, was taking it in stride by the time CBC News spoke to her. "Obviously, it wasn't a great thing that happened ... people do make - or not people, I guess it wasn't people make mistakes, I guess AI makes mistakes," she said, breaking into laughter.

The error record is not new

CBC News noted that AI overviews rolled out in 2024, and that early mistakes, including a suggestion to use glue to keep pizza toppings in place, generated headlines. That episode was part of a broader pattern of factual accuracy problems documented at launch, followed weeks later by Google's public response outlining refined triggering and stricter safeguards for sensitive topics.

The failure mode in the Anna Mae's case is a specific one: entity resolution. Two businesses share a name, one is closing, and the generated summary merges them. Similar confusions have surfaced before. In April 2025, an AI overview stated that SEO professional Lily Ray was nine years old, attributing details about her dog to Ray herself. Practitioners have also reported incorrect business phone numbers appearing in AI overviews even when correct numbers were present in Google Business Profile, a mismatch between Google's own structured business data and its generated text.

Survey data puts numbers on user exposure. Research from NP Digital found that 25% of users encounter AI overview errors, with inaccurate answers accounting for 51% of reported issues and outdated information a further 21%. A later study from the same firm found that 47.1% of marketers encounter AI inaccuracies several times a week, while 36.5% said hallucinated or incorrect AI-generated content had already been published publicly.

CBC News also cited a report released last year by the BBC examining AI chatbots from Google, Microsoft, OpenAI and Perplexity, which found "significant inaccuracies" in their summaries of news stories, though it did not examine Google AI overviews specifically.

Why a village bakery sits inside an advertising story

The Millbank incident touches infrastructure that marketers manage directly. Google Business Profile has become the primary data layer feeding Gemini, Search and Maps, a shift documented in April 2026 through a combination of Google's own community newsletter and API changes analysed by local SEO practitioners. What was once a directory listing now functions closer to a structured identity record, and the quality of that record carries direct consequences for local visibility.

Around that layer, Google has been adding AI-mediated surfaces at a steady rate. In July 2025, the company began automatically calling small businesses on behalf of customers to check pricing, wait times and appointment availability. In November 2025, Google executives described how AI Mode generates related queries and searches reviews to produce curated business recommendations, drawing on a product graph containing 50 billion items. The following month, Google retired community Q&A on Business Profiles in favour of an AI-generated answer button, and Gemini began serving local Maps results in visual card format.

Each addition inserts a generated layer between a business and a prospective customer. When that layer states a closing date, the business learns about it the way Herrfort did: secondhand, from someone else's screenshot.

Scale sharpens the asymmetry. AI overviews now exceed 2.5 billion monthly active users and AI Mode exceeds one billion, figures cited in Google's documentation for the new control that lets site owners exit AI Overviews and AI Mode at domain level. That control clears content from AI features within one to two days, though it does not stop model training and it applies to sites an owner controls. It offers nothing to a restaurant whose problem is a statement generated about it rather than content taken from it.

Liability has stopped being hypothetical

Courts have begun testing whether a generated summary is the search engine's own speech. On September 10, 2025, the Frankfurt Regional Court dismissed a preliminary injunction sought by plastic surgeons over an AI overview but confirmed that Google could in principle be liable for objectively incorrect statements, rejecting the argument that AI overviews amount to mere third-party information.

Eight months later the position hardened. On May 28, 2026, the Munich Regional Court I issued a preliminary injunction against Google over false claims its AI overviews generated about two German publishing companies, backed by fines of up to EUR 250,000 per violation and assigning Google 80 percent of costs. The written reasoning, published later, rested on three grounds: that AI overviews speak in their own words, that they contain claims found in no source, and that treating them as ordinary search results would leave victims of machine-generated falsehoods without remedy.

United States litigation has followed a parallel track. Minnesota solar contractor Wolf River Electric sued Google on March 11, 2025 after its AI systems produced claims about a state attorney general lawsuit that never existed, a case reported as one of at least six AI defamation suits filed in the country over two years. In October 2025, activist Robby Starbuck filed a $15 million defamation suit over AI-generated claims.

None of that legal machinery is available in practice to a rural bakery facing a fabricated closing date. The cost of an injunction exceeds the cost of the harm, which is precisely why the remedy in Millbank was a message typed into a search bar.

The measurement backdrop

The reputational question sits alongside a traffic question that has been quantified repeatedly. Ahrefs research published April 17, 2025 found AI overviews reduced organic clicks to top-ranking pages by 34.5% across 300,000 keywords. A February 4, 2026 update from the same firm put the figure at 58% for top-ranking pages, comparing December 2025 data against a December 2023 baseline.

Causal evidence arrived in 2026. A randomized field experiment with 1,065 desktop Chrome users found that when an AI overview appeared, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%, while clicks on sponsored results stayed flat. Counter-evidence exists: clickstream data from Datos covering desktop behaviour showed zero-click searches in the United States falling from 24.5% in December to 22.4% in March 2026, the lowest level in the period measured.

The two bodies of evidence describe different things. Aggregate click share can hold steady while individual queries resolve inside the summary. A search for a specific restaurant name is exactly that kind of query: the answer is short, the intent is factual, and the user has little reason to click through. When the answer is wrong, the correction never reaches the person who read it.

What the incident establishes

Three things are documented by the Millbank case. First, entity confusion between similarly named businesses in different jurisdictions produces confident, dated, factual-sounding claims. Second, the remediation path available to an affected small business is informal, conversational and unverifiable. Third, the burden of monitoring falls entirely on the business, which has no notification when a generated summary describes it incorrectly.

For agencies managing local search presence, the operational implication is a monitoring gap rather than an optimisation opportunity. Structured data hygiene in Business Profile does not prevent a summary from importing facts about a different entity with the same name. Franchise networks, regional chains and businesses with common or duplicated trade names sit at the highest exposure.

Herrfort described what she took from the episode to CBC News: "The best thing to do, if something like that comes up, is to contact the people directly or go to like our actual website or our Facebook page, because nowhere on any of our actual branding was any information [about closing] posted." She added: "don't just read what AI is telling you at the top."

The CBC report was published on July 23, 2026 at 2:21 PM EDT, written by Ieva Lucs, with files from The Canadian Press. As of that reporting, Anna Mae's Bakery and Restaurant remained open.

Timeline

Summary

Who: Amanda Herrfort, owner of Anna Mae's Bakery and Restaurant in Millbank, Ontario, along with Google, a Google spokesperson who responded to CBC News, and Christine Logel, a professor of social development studies at the University of Waterloo.

What: A Google AI overview stated that the Millbank restaurant was officially closing on July 31, after confusing it with an Illinois bakery of the same name that is closing. Customers emailed, called and visited to ask about the closure, and Herrfort resolved the error by typing a correction request into the Google AI search bar rather than through any formal reporting channel.

When: The false summary circulated over the weekend of July 18 and 19, 2026. Herrfort spoke to CBC News on Thursday, July 23, 2026, and the broadcaster published its report the same day at 2:21 PM EDT.

Where: Millbank, Ontario, a community northwest of Kitchener in southern Ontario, with the summary distributed through Google Search results visible to any user querying the business name.

Why: Google attributed the mistake to an absence of high quality web information on the topic and said the vast majority of AI overviews meet its accuracy bar. Logel attributed it to systems released before full testing, arguing that mistakes carry real consequences when generated summaries are embedded across so many products. The incident matters commercially because AI overviews now reach billions of users, because Business Profile data feeds Google's generative surfaces directly, and because businesses receive no notification when those surfaces describe them incorrectly.