A customer once forwarded us a ChatGPT screenshot with the comment, "this is wrong, my Tuesday hours are 5pm not 6pm — but everything else is right." That single misstatement was small but instructive. It cost the restaurant two reservations that night and surfaced a structural feature of how AI assistants represent your business: they fabricate details under predictable conditions, and the fix is mechanical, not editorial.
The technical name for this is hallucination. The accessible explanation is that LLMs generate text by predicting the most-probable next token based on prior context. When your business has thin grounding signal — for example, your Google profile lacks structured hours, your website's hours page is buried, your TripAdvisor profile lists older hours — the model hits a low-confidence region and falls back on its statistical baseline of "what restaurants like yours typically list as hours." That baseline produces plausible-sounding but invented details.
Where hallucination concentrates
Across our audit base of 200+ businesses, we've cataloged where hallucination shows up most. In rough rank order:
- Hours of operation, especially edge cases (Sunday brunch, holiday hours, late-night). Most common because hours data is updated frequently and aggregator listings drift.
- Specific menu items — invented dishes, often plausibly riffing on the cuisine type but not on the actual menu.
- Reservation policy — "you should book 2 weeks ahead" when reality is walk-in friendly, etc.
- Price range — particularly for restaurants without a clear price-tier signal.
- Specific awards or accolades — sometimes inventing a Michelin mention that doesn't exist, occasionally on the basis of a single old article that misread the situation.
The reservation/menu hallucination is especially expensive
If a customer asks ChatGPT about your restaurant before booking and gets back invented menu items or wrong dietary information, two things happen. First, expectations are mis-set — the diner arrives expecting a dish you don't serve. Second, when the diner reports the issue (often in a review), the negative review compounds back into the AI's grounding for future queries.
This is the mechanism by which a single hallucination becomes a drip of repeated mis-information. Fixing the upstream signal is more durable than addressing each downstream review.
The signal hygiene that fixes it
Three actions cover most cases:
1. Schema.org Restaurant markup on your own site. JSON-LD with hours, menu sections, accepted payment types, address. Schema gives AI a clean, structured parse — far higher confidence than scraping HTML or aggregator data. The grounding pass picks up structured data preferentially.
2. NAP consistency across the web. Name, address, phone — identical to the character on Google, your website, TripAdvisor, OpenTable, Yelp, Booking, your Instagram bio. Even small differences ("123 Main St" vs "123 Main Street") sometimes cause AI to treat two listings as different entities, doubling your fabrication risk.
3. Refresh your Google profile data quarterly. Hours change with seasons. Menu changes with chefs. Photo galleries get stale. A 90-day refresh cadence dramatically reduces hallucination on transient details.
What we don't recommend
Don't argue with the AI in chat. Telling ChatGPT "no, that's wrong" doesn't update anything — your correction lives in the conversation context, not in the model's grounding. The next user asking the same question gets the same wrong answer.
Don't pay for "AI optimization services" that promise to inject your business into AI training data. Training data is set at the model's training time; you can't push into it post-hoc. The vendors selling this are selling SEO under a new label.
How we surface hallucination in your audit
Detailed and Full Audit reports include a "verbatim AI response" section where you see exactly what each platform said about your business. If we detect hallucination — invented hours, fabricated menu items, inaccurate awards — it's flagged in the action list with the specific signal-hygiene fix prioritized.
Most owners are surprised by what they find. The fixes are often 30 minutes of work spread across three signal layers. The reduction in hallucination is measurable on the next audit.