Match Common Failures to Minimal Fixes

Prerequisites: Lectures 2, 3, 4, 6, 7, 8, 9, 10, 11, 12 and 13.

Before this lecture, you should be able to split an answer into the four rooms, recognise category drift, keep one property consistent, weigh the source trail, write citeable facts, read location signals and review context, handle stale evidence, run a manual prompt routine, and keep branches distinct. Now we use those skills as a repair table: not “what is wrong?” only, but “where should I touch first?”

One owner once showed me a printed ChatGPT answer with three circles in red pen. The answer named the hotel correctly, placed it in the right town, and then called it “a spa-style boutique property with private parking.” The property had neither a spa nor private parking. It did have a small garden, a breakfast room with pale tiles, and one old booking description that used the word “wellness” because someone had translated “relax” too generously. The parking mistake came from a nearby sister property.

That is the point of this lecture. When an AI answer feels wrong, the tempting reaction is to rewrite everything: homepage, About page, service page, listing text, maybe even the whole English version. This is usually too much. It is like repainting every door in the hotel because one guest opened the laundry cupboard. A better habit is to classify the failure, then choose the smallest correction target that makes the wrong sentence harder to produce.

Name the failure before touching the page

A failure pattern is a recurring AI error such as wrong category, old address or invented amenity. I want you to hear the word “recurring” carefully. One strange answer may be noise. A repeated wrongness across several prompts, languages or similar guest questions deserves a name.

The manual prompt routine from Lecture 12 gives you the raw material. You ask plain questions, copy answers, mark the four rooms, and look for the same weak shape appearing again. “Boutique hotel” appears when the property calls itself a guesthouse. “Near Verona” appears when the town is actually a smaller village outside the city. “Private parking” appears because another branch has it. These are not all the same kind of problem. They live in different rooms, and they need different repairs.

A correction target is the page, listing or profile where a specific evidence repair should happen first. First does not mean only. It means the most direct place to reduce confusion. If the category is wrong, the first target may be the About page and structured listings. If the location is vague, the target may be the location page, map listing, or landmark wording. If an old amenity keeps returning, the target may be a stale directory or a change notice on the owned site.

A practical rule: do not repair the whole property picture while you are still unsure which room failed. If the name room is solid and the promise room is inflated, do not spend the afternoon adjusting the canonical property name. If the place room is weak, do not start by polishing breakfast copy. Work where the failure lives.

Match category and place failures to their evidence

Category drift from Lecture 3 often begins politely. A guesthouse becomes a small hotel. A B&B becomes a boutique stay. An agriturismo becomes a rural resort. The answer may still feel flattering, which makes the owner hesitate. But category is not decoration. It sets expectations before the guest reads amenities.

In a composite scenario, a family-run guesthouse uses “boutique atmosphere” in several English snippets because the rooms were renovated with good taste. A directory lists it under hotel. ChatGPT answers: “a boutique hotel suitable for a refined lakeside stay.” The identity room is correct. The place room is mostly correct. The category room is drifting upward, and the promise room is starting to borrow from that upward label.

The correction target is not a new paragraph saying, “We are authentic and charming.” That may make the drift worse. The target is category evidence: headings, first paragraphs, listing fields where available, booking category labels, and the opening line of the About page. The page needs a sentence that a tired system can repeat without polishing it into grandeur: “Casa Rosa is an eight-room family-run guesthouse in the village of ___.” If the property also uses “boutique” because the owner likes the word, place it carefully as atmosphere, not type.

Place failures need the same discipline, but the repair is usually about scale. “Near Verona” may help a foreign traveller orient themselves, but it becomes misleading if the property is in a village where arrival, transport and evening plans are different. “On Lake Garda” may be acceptable in tourism shorthand and still wrong for a property five kilometres inland.

The correction target is usually a location page or map-related evidence. The page should state the exact town, address, neighbourhood or hamlet, and the practical relationship to larger landmarks. A good location sentence has two hands: one hand holds the precise place, the other points toward the better-known place. “The guesthouse is in ___, a village outside Verona, rather than in Verona city centre.” That sentence removes a tempting mistake without pretending travellers never search by Verona.

English evidence often widens the place for international guests, while Italian evidence may give the smaller locality naturally. If the English page says “near Verona” five times and names the village once in the footer, ChatGPT has been handed a megaphone and a whisper. Do not be surprised when it repeats the megaphone.

Treat invented amenities and old facts as promise-room repairs

Invented amenities are some of the most damaging errors because they sound operational. Spa, private parking, airport transfer, lake-view rooms, pet-friendly policy, elevator, restaurant, pool, shuttle. A guest may make a decision from one of these words. The promise room becomes a contract in the guest’s head, even if the model did not mean it that way.

The first question is not “why did ChatGPT invent this?” Ask a rougher, more useful question: which public phrase made this amenity easy to guess? “Relaxation area” may become spa. “Convenient arrival by car” may become parking. “Dinner recommendations available” may become restaurant. Reviews can also bend the promise room. If several guests write “we parked easily nearby,” the answer may smooth that into “parking available.”

In a teaching example, imagine a small guesthouse stopped offering airport transfers after a staffing change. The official service page was updated, but one partner page still says “transfer on request.” ChatGPT repeats “airport transfers may be available.” This is partly stale evidence, partly promise-room uncertainty. The correction target is the old partner page if it can be edited, plus a clear service boundary on the owned page: “Airport transfers are not offered; guests can use local taxi services from the station and airport.”

For invented amenities, the smallest fix is often a negative fact. Hospitality owners dislike negative facts because they sound unwelcoming. I understand that. But a calm boundary is better than a disappointed guest at reception. “No private parking is available on site; paid public parking is five minutes away” is not hostile. It is useful. Parking nearby is not the same as property parking. A restaurant nearby is not the same as an in-house restaurant. A beach nearby is not the same as a private beach.

Old facts require a date, not only deletion. If a property moved, renamed, renovated, stopped transfers, changed breakfast rules, or altered parking, ChatGPT may preserve the old version unless the current evidence is easy to read. A change notice gives the current version a recency signal: “From March 2025, breakfast is served only by advance request.” Deleting the old sentence removes one clue; a dated notice adds a clearer present tense.

Separate lookalikes and branches at the boundary

Entity confusion from Lecture 7 and multi-location confusion from Lecture 13 require a different repair instinct. If ChatGPT borrows a pool from the hotel next door, the correction target is not only your own pool sentence. The target is the boundary between the two properties.

A disambiguation page can help when two unrelated properties share a similar name, town, landmark or review context. It should state who the property is, where it is, what it offers, and what it is not. But the page should not sound like an accusation against the neighbour. It should read like a careful label on a luggage tag: “This guesthouse is located at ___ in ___ and has eight rooms; it is not the similarly named hotel in ___ and does not offer a pool.”

For related branches, the correction target is often a branch page or a brand comparison section. If the road branch has parking and the centre branch does not, both branch pages should say so in their own words. The brand page can then point guests to the difference instead of blending it: “Parking differs by property.” This prevents the wrong fact from sliding across the group.

In a composite scenario, two nearby properties share branding and a booking engine. One is in the historic centre; the other is outside town near the lake road. ChatGPT says the historic-centre property has private parking and lake-road access. The failure pattern is branch crossover. The correction target is not the homepage mood copy. It is the branch evidence: address, access, parking, landmarks, and a short comparison sentence that names both branches.

Reviews make this messy. A guest may mention both properties in one paragraph: “We stayed in the centre and visited their sister hotel near the lake for dinner.” That review is honest but slippery. You cannot rewrite guest speech. You can give official pages cleaner separation so the review does not carry the whole explanation.

Choose the smallest repair and test it again

The smallest repair is not always the easiest edit. It is the edit closest to the failure’s source. If the map category is wrong, rewriting the About page helps only partly. If the About page is vague, correcting a distant directory may not carry enough weight. If reviews repeat a phrase, you cannot edit them, but you can add a citeable fact that gives the repeated claim a boundary.

Think of the source trail as a corridor with several doors: owned page, directory listing, review profile, map listing, booking page, local tourism page. The correction target is the first door that actually opens onto the failed sentence. Sometimes you can open it. Sometimes you can only put a clearer sign on your own door and ask the other door’s owner to fix theirs.

This is why the answer log matters. It stops you from repairing by mood. If a failure appears in the category room across three prompts, you mark wrong category and repair category evidence. If the failure appears only when asking about a sister property, you mark branch crossover. If the failure appears in English but not Italian, you inspect translation drift. The pattern tells you where to act.

A good minimal fix has three qualities. It is specific enough to repeat, placed where public evidence can find it, and narrow enough that it does not create a new exaggeration. “We are a lovely place for everyone” fails all three. “The guesthouse has eight rooms, breakfast by request, and no private parking on site” is plain, perhaps too plain for a brochure opening, but excellent as evidence.

After the fix, do not declare victory. Run the same audit prompts later and compare. If the wrong sentence weakens, you may be on the right path. If it persists, look for a stronger or older source. If a new wrongness appears, classify it before touching another page. The habit is simple: small named failures, small targeted repairs, patient checking.

What to remember

A failure pattern is a recurring AI error such as wrong category, old address or invented amenity. Name the pattern before rewriting pages.

A correction target is the page, listing or profile where a specific evidence repair should happen first. The first target should match the room that failed.

Wrong category, invented amenity, vague place, similar-property confusion, branch crossover and old facts are different problems. They may appear in one sentence, but they do not share one repair.

Four rooms of Italian hospitality visibility are the name room, the category room, the place room and the promise room, because ChatGPT must recognise who the property is, what kind of property it is, where it belongs and which promises public evidence can support.

The smallest useful repair is often a clear sentence, a corrected listing field, a dated change note or a branch-specific boundary.