Why do AI assistants recommend businesses that don't exist?
Because when an assistant answers from memory rather than by searching, it produces the most plausible answer, and the most plausible answer can be a business that has never existed where it says it does.
A real case, from the same Zürich study in July 2026. Before collecting any assistant answers, I ran a map scan of central Zürich for one keyword, "bjj": 49 measuring points across a five-mile radius, listing every martial arts business the map returned, down to entries with no reviews at all. It found 87.
Then I asked the assistants where to train, with web search switched off, so they had to answer from what they already knew. Two large international martial arts franchise names came up repeatedly, described as having Zürich branches. Neither appears anywhere among the 87 businesses the scan found. The assistants were confidently recommending gyms that do not exist in this city.
With search switched on, this largely stopped. When an assistant goes and looks, it finds the businesses that are actually there.
Why it happens, in plain words. An assistant answering from memory has absorbed a great deal of text about martial arts gyms in general. Big franchises appear in that text far more often than any single independent academy. So when it reaches for a name to attach to "Zürich", the franchise is the likelier guess. It is completing a pattern, and the pattern points at the franchise.
Why this matters to you. Many products built on these assistants have no web search at all, so the from-memory answer is the only answer some of your customers get. Even where search is available, advice-shaped questions ("where should a complete beginner start?") are often answered from memory anyway. And what an assistant already believes shapes which names it picks when it does search, so the memory sits underneath the search rather than beside it. Every invented competitor in an answer is a place on the list you did not get.
Nobody can edit an assistant's memory. What you can change is what it finds when it looks, and what it finds, over time, becomes what the next version remembers.