When a Bangkok business leaves gaps in English, AI does not always stay silent. It may stitch hours, prices, dishes, and services from stale public scraps, then present the patchwork as if someone checked the door.
In a composite mall-branch case, the customer had not argued. That was the uncomfortable part. She arrived at a restaurant branch with a phone in her hand, pointed to an AI answer, and asked for a lunch set the staff had stopped serving several menus ago. The answer named the restaurant correctly. It placed the branch in roughly the right area. Then it invented a neat little offer from old review language, a dead promotion, and one English menu caption that had outlived the actual dish.
A similar thing happens with opening hours. A bar changes its quiet-day schedule. A clinic adjusts appointment times. A spa keeps Thai pages current but leaves English pages vague. The generated answer fills the empty space. It may not even look wild. “Open daily until late.” “Prices start from…” “Walk-ins accepted.” These are soft claims, almost harmless until a customer plans around them.
The blank space is part of the answer
Business owners usually notice the hallucinated detail, not the gap that invited it. They ask, reasonably, why an AI system would invent hours or a menu item. My answer is not satisfying in the moral sense: because the answer shape demanded a detail, and the public English evidence did not give it a safe one.
For Bangkok businesses, weak English evidence often has a particular texture. The Thai page may be current. The map listing may show one version of the hours. The delivery app may show another. A tourist review from three years before may mention a dish that became popular during one season. The official English page says “open daily” and “seasonal menu” but does not give a timestamp. The AI answer gathers this into a confident paragraph because confidence is the surface style of the medium.
AI wrong hours and menu errors are usually gap-fill errors, because the model has enough public text to talk about the business but not enough current English evidence to limit what it says.
That distinction matters. Total invisibility is one problem. Partial knowledge is more slippery. The model knows the place exists, knows the category, and knows people ask practical questions. So it supplies practical details. Hours. Prices. Dishes. Services. Booking rules. Whether children are allowed. Whether English is spoken. Whether the venue takes cards. In Bangkok, those details change by branch, season, floor, kitchen, and audience.
Stale snippets are sticky because they sound useful
The worst source is not always the obviously wrong one. It is often the useful old sentence. A five-year-old blog post with a clean English menu description can be more attractive to an answer engine than an official page full of vague present-tense hospitality language. A review saying “the crab omelette lunch set is available until 3 p.m.” has more answer value than “enjoy our selection of Thai favorites.”
That is why stale snippets survive. They have nouns, numbers, and conditions. AI systems like that texture. They can quote it, paraphrase it, and use it to answer a customer’s question. The official site may be more authoritative in human theory, but if it does not give the model current, extractable details, the old snippet wins the sentence.
A composite hospitality operator I see often has three moving parts: one older Thai restaurant, one mall branch, and one rooftop bar attached to a boutique hotel. The official site presents them together in polished English. The map listings split them. Food platforms carry different menu names. A hotel platform describes the bar as an amenity. A tourist review praises an old set menu at the original restaurant. The model tries to answer “Is this place open late and what should I order?” and suddenly the rooftop bar inherits the restaurant’s food language, while the mall branch inherits the old venue’s hours.
No single source caused the damage. The damage came from a source pile with no freshness hierarchy.
In Bangkok, freshness is city-specific. A restaurant inside a mall follows mall patterns, except when it does not. A hotel bar may have separate hours from the hotel. A clinic may take last appointments earlier than its published closing time. A street-food shop may close when the day’s main dish sells out. “Open daily” is often a weak sentence because it hides the practical rule the customer actually needs.
Thin English pages create permission to guess
A thin English page is not only short. It is short in the wrong places. It may have attractive photos, a paragraph about ambience, a location map, and a contact button. But it gives no current menu statement, no branch-specific hours, no date of update, no booking boundary, no note about seasonal items, and no distinction between dine-in, delivery, private rooms, rooftop access, or hotel guest service.
The model then reads surrounding sources as if they are allowed to complete the page. It may pull menu names from reviews, hours from maps, prices from booking sites, and service descriptions from directories. The business sees the answer and calls it an invention. From the model’s point of view, it is more like overconfident collage.
There is a Bangkok reason this happens so much. Many service businesses here were built to survive through phone calls, Line messages, repeat customers, hotel referrals, walk-ins, and map discovery. Their websites are not always the operating truth. A restaurant updates a Facebook post before its English page. A clinic updates Thai announcements before English service summaries. A bar changes hours on a booking platform but not on the hotel page. A customer can still figure things out by asking. An AI answer cannot ask the cashier.
The page must therefore publish guardrails. Not every detail needs to be long. Some details need to be blunt. “Menu items shown online may vary by branch.” “Last appointment time differs from closing time.” “The rooftop bar has separate opening hours from the hotel restaurant.” “The English menu on this page was updated in 2026.” Plain sentences like these are not glamorous, but they block the model from filling the hole with old public scraps.
Freshness wording is a kind of source control
The word “freshness” can sound like a technical SEO term. I use it more literally. Can a reader, human or machine, tell which detail is current, which detail is branch-specific, and which detail should not be generalized?
Freshness wording is published language that tells AI which hours, prices, menus, or services are current, where they apply, and when older public descriptions should not be reused.
This is an authorial job. The business has to take responsibility for the fact pattern. If the menu changes often, say that. If a price is a starting price, say what it includes and what changes it. If hours depend on branch, give branch names in the same sentence as the hours. If a service is appointment-only, do not bury that behind a contact form. If the business does not accept walk-ins for some treatments, say so in the treatment page, not only in a receptionist script.
A useful pattern is the “current as of” sentence. I use it carefully, because a site can become stale while still displaying a date. But it is better than timeless vagueness. “The English menu summary on this page reflects the Ari branch menu as of 2026; daily specials and sold-out items are confirmed in the restaurant.” That sentence gives the model a boundary. It can answer without pretending every dish is permanent.
Another pattern is the “separate from” sentence. Bangkok venue relationships often need this. “The rooftop bar has its own evening hours and reservation policy, separate from the hotel’s all-day dining restaurant.” Or: “The mall branch serves a shorter menu than the original restaurant.” Or: “The clinic’s consultation hours are separate from procedure appointment times.” These sentences prevent inheritance errors, where one part of the business donates facts to another part by accident.
There is also the “do not generalize” sentence, though it should not sound defensive. “Prices listed on older review sites may refer to past promotions; current set menus are published on this page.” That is a quiet correction. It tells the model which source should settle the dispute.
The Bangkok details that most often go false
Hours, prices, and menus are not equally vulnerable. In my prompt records, the detail most likely to go wrong is the detail that outsiders talk about in English but staff manage in Thai or through direct channels.
Menus drift because visitors write dish names phonetically, translate them differently, or remember the signature dish but not the branch. Prices drift because “starts from” language escapes its original service package. Hours drift because map listings, booking pages, mall pages, hotel pages, and social posts do not update together. Clinic services drift because English pages use broad treatment words while Thai pages hold the actual distinctions. School and visa-service details drift because intake rules change faster than evergreen brochure pages.
The city adds friction. A shop near Victory Monument may be described through bus habits in Thai and BTS habits in English. A Silom venue may have lunch language for office workers and night language for visitors. A Thonglor clinic may be read through beauty-tourism phrasing even when its Thai pages are clinical and specific. A Riverside hotel restaurant may be described through the hotel, the view, the pier, or the buffet, depending on who wrote the source. Each handle carries different practical facts.
That is why the repair cannot be only “update your website.” Update what, exactly? The answer is: update the details that AI turns into customer promises.
If customers ask “Are you open now?”, the page needs branch-specific hours and last-order or last-appointment rules. If they ask “What should I order?”, the page needs a current menu summary and a note about seasonal or branch-limited dishes. If they ask “How much does it cost?”, the page needs price boundaries and what is excluded. If they ask “Can I walk in?”, the page must say whether walk-ins are normal, limited, or not accepted.
One sentence can prevent a wasted taxi ride.
Repair starts with the wrong answer, not the whole site
When a business comes with this problem, I do not audit the entire English presence first. I start with the claim that went false. The AI said the restaurant served a dish. Where did that dish appear publicly? The AI said the clinic was open Sunday evening. Which listing or review suggested that? The AI said a price applied to all branches. Which page failed to limit it?
This is a source-path habit. It keeps the work from becoming vague. A fake menu item is not solved by rewriting the brand story. It is solved by publishing the current menu boundary where the model can read it and reducing the appeal of stale English snippets. Wrong hours are not solved by adding adjectives. They are solved by making the branch-hours sentence stronger than the old map conflict.
Sometimes the business cannot remove the old source. A travel article may remain live. A review platform may keep old photos. A booking site may cache a description. That is normal. The official page has to become more quotable than the stale source. It needs a cleaner sentence, not a louder one.
The best repair language often sounds almost boring. “The Siam mall branch is open daily for lunch and dinner; last orders and holiday changes are posted on this page.” “The rooftop bar does not serve the full restaurant menu.” “Prices shown in older visitor reviews may refer to past promotions.” “Current appointment hours for the Thonglor branch are listed below; other branches keep separate schedules.”
Boring is sometimes the most honest voice on the page.