It Knew the Name.

It Answered for the Wrong One.

Google’s own Knowledge Panel help lists “mixed with another of the same or similar name” as a standing support case. This year two reported AI Overview errors showed the same observable shape, one business’s information attached to another with a similar name: a software company’s founder reporting six months of another company’s customers sent to his support address, and an Ontario bakery said to be closing because a bakery in Illinois is. Google’s explanation to CBC of how such issues can arise: “an absence of high quality information on the web.” The name was found. The business was not.

By William Douglas Baker

Founding Editor

Over a weekend in July, someone searched for a bakery. Anna Mae’s Bakery and Restaurant sits in Millbank, Ontario, northwest of Kitchener, and the search returned an AI Overview saying the business was officially closing on July 31. The owner, Amanda Herrfort, learned of it from a screenshot sent to her on the way to a wedding. “That’s when I panicked,” she told CBC News. There is an Anna Mae’s Bakery in Illinois, and that one is closing. The AI Overview named the Ontario bakery and gave it a closing date that belongs to the Illinois one. She launched an online campaign to say otherwise. Customers still emailed, called, and walked in. “We have a lot of people coming in talking to our employees,” she said, to tell them what the machine had told them.

Six months earlier, the founder of Glama, an AI workspace company, had written up the same failure from the other side. “Six months ago, I started receiving urgent support tickets demanding subscription cancellations and refunds.” The tone was hostile, and the customers were not his: “I couldn’t find these people in Stripe.” Eventually he pieced it together. “They were looking for an entirely different company” — “one with a bright pink and yellow brand (we’re dark green). The names are similar, but the products aren’t even close.” The cause, once a persistent customer sent a screenshot, was not on his site. “It was Google’s AI Overview. Not our website.” Google, he wrote, was presenting his support address as the contact for the other company. The toll: “Hundreds of frustrated customers over six months.” Then the detail that turns a mishap into a mechanism: “it works both ways.” Search for his company and Google sometimes returned the other one’s details. With a VPN he could reproduce both errors “(consistently) in some territories, but not others.”

Google has a page that names the shape, though it sits on a different surface: its Knowledge Panel help, not the AI Overview. That help lists the situations that warrant contacting support, and one of them reads, in full: “The content of the knowledge panel is mixed with another of the same or similar name.” It sits beside “The knowledge panel is showing the wrong websites or profiles.” A separate item covers a panel already verified by someone else. Google first tells the claimant to check with colleagues or others who may have claimed it on the entity’s behalf. If the owner still cannot be identified, Google’s Account Recovery process asks permission to share the claimant’s email with the panel’s current owners or managers; Google then gives them three business days to approve access, and without a response the panel becomes claimable again. That is an ownership-recovery process, not Google’s documented remedy for separating two entities with similar names. Same-name mixing is an anticipated support case, filed by the platform itself, for its panels. The two accounts above show the same kind of output failure in AI Overviews: information belonging to one entity appeared in an answer about another. They do not establish that Knowledge Panels and AI Overviews share the same underlying process. The record reads the resemblance as entity confusion; Google has not said so.

An absence, in Google’s words

Asked about the bakery, a Google spokesperson told CBC News that the “vast majority of AI overviews are high quality and meet our high bar for helpfulness and accuracy,” and then offered its explanation for how such issues can arise: “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’s own account of AI Overview errors, from May 2024, is general in the same way: when they get it wrong, “it’s usually for other reasons: misinterpreting queries, misinterpreting a nuance of language on the web, or not having a lot of great information available.” Read that as written. A business CBC describes as “a popular spot for both tourists and locals” does not lack standing. What Google points to is the web: an absence of high-quality information on a topic. Not an absence of authority. An absence of information a machine can use. That is the diagnosis this record has filed for forty-six editions. Google did not file it, and did not say which of its general reasons applied here; the condition it described was information, not standing. The distinction is the record’s, and the record says so.

The record has a name for what the outputs show. Authority Misclassification — the first clinical characteristic of Digital Derangement Syndrome™: the system cannot confidently associate the entity with its own domain, and so associates it with something adjacent. Usually the failure is a wrong category or a wrong tier. This one presents as a wrong-entity association: a closing date belonging to one bakery appeared in an answer about another, while one software company’s support address appeared in answers about another company. The outputs show attributes crossing entity boundaries. The public record does not reveal where inside Google’s systems that association failed. Edition No. 029 filed credentials that were real and never assembled. This is the mirror image: attributes assembled, confidently, onto the wrong entity. The record does not claim that more signal would have prevented either case; what it notes is what Google’s explanation points to.

Neither owner did anything wrong. Neither had a way to know until a customer told them: one from a screenshot on the way to a wedding, the other from six months of refund demands for a product he does not sell. Both then did the only thing on offer, and asked the machine to stop.

The machine knew the name. Knowing the name is not knowing you.