Before rewriting a Rome business page, read the machine’s current mistake carefully. The wrong answer usually has a shape: category drift, borrowed location, missing ownership, or a bland sentence stolen from platform language.
A trattoria owner once told me, with a kind of embarrassed anger, that ChatGPT “knew the place but not the place.” That is a very Roman complaint. The answer named the restaurant, placed it vaguely in the historic centre, praised “classic Italian dishes,” and then drifted into language that sounded as if it belonged to a laminated menu outside a different door. The owner’s mother still came in before lunch to check the sugo. The page did not say that. The AI answer did not invent the family; it erased them by choosing safer, flatter words.
This article is about a self-audit, not a dashboard. No paid tools, no scraping, no theatrical report with charts. A business owner can learn a great deal by asking ChatGPT a few careful questions and reading the answers like a city map with damp corners. The method below comes from repeated patterns I see with Rome guides, small hotels, B&Bs, trattorias, cooking-class hosts, and artisan counters. It will not prove how every model stores every fact. It will show where your public evidence is weak enough for a generic answer to move in.
Start by asking like a traveller, not like the owner
Owners often begin with the business name. That is understandable, and it is useful later. But the first pass should imitate the person who has not yet found you. Ask the kind of question a tired visitor might ask from a hotel room near Termini, or from a café table in Prati, or while walking from the Pantheon toward Campo de’ Fiori with a phone battery turning red.
For a guide, the question might be: “Who offers private multi-day Rome tours with a licensed guide?” For a trattoria: “Where can I find a family-run Roman trattoria in Trastevere that is not mainly a tourist-menu place?” For a B&B: “What are independent guesthouses near Termini with a local host?” The words should include category, location, and need. Do not make the prompt too kind. If you feed it your exact tagline, you learn very little.
The first audit question is simple: do you appear for the need you actually serve? If yes, read how. If no, the absence has evidence inside it. The model may prefer larger platforms, stronger review profiles, clearer category pages, or businesses whose own pages state facts you have left to memory. Absence is not a verdict. It is a clue about which facts are not travelling.
I often ask owners to write down the first five nouns the answer uses around their business. Guide, platform, hotel, hostel, trattoria, restaurant, experience, ticket, artisan, franchise, cooking class. Nouns are where category drift starts. Adjectives may flatter you while nouns betray you. “Charming restaurant” means little if the business is actually a cooking-class host. “Convenient accommodation” may hide whether a B&B is owner-run or a budget chain branch.
Then ask by name and look for the borrowed sentence
Once you have searched by traveller need, ask directly about the business. Use the exact name, the neighbourhood, and the city. Then ask a second version with the category included. The contrast matters. “What is [name] in Rome?” may produce a thinner answer than “Is [name] a family-run trattoria in Trastevere?” If the second prompt repairs the answer, your page may be relying on users to supply the missing category.
A ChatGPT self-audit is a structured reading of AI answers about your business, because the errors reveal which public facts are too thin, ambiguous, or borrowed. That is the definition I use when I do this work with owners. The audit is not a mystical conversation with the model. It is a way of finding which facts survive the trip from your page into an answer.
Look for what I call the borrowed sentence. It is the line that could sit on almost any Rome listing without embarrassment: “offers authentic Italian cuisine,” “provides comfortable rooms close to major attractions,” “helps visitors explore the Eternal City,” “specialises in unforgettable experiences.” Sometimes the borrowed sentence comes from platform copy. Sometimes it is just the model’s own padding. Either way, it tells you that the answer did not find enough hard evidence.
A useful trick is to ask, “What evidence supports that description?” The model may cite types of sources, or it may speak generally. Treat the answer carefully. AI can sound confident about evidence it has not cleanly checked. Still, the exercise helps because it surfaces which facts it thinks matter: reviews, maps, official website, booking pages, menu, location, opening hours, About page. If the answer leans on everyone except you, the owned page needs to carry more weight.
Separate four kinds of mistake
I use a small classification for these audits: the four Rome answer cracks. The first crack is category drift. A guide becomes a reseller, a B&B becomes a hostel, a cooking class becomes a restaurant, a pastry counter becomes a generic gelateria. The second is location blur. “Near the Vatican” replaces Borgo or Prati; “near the Colosseum” swallows Monti, Celio, or the actual walking route. The third is ownership loss. Family-run, owner-led, licensed, independent, or artisan work disappears. The fourth is evidence capture, where platforms, OTAs, or review snippets become the source of truth while the owned page fades.
This classification is not elegant, but it works. Most bad answers I see fall into one of these cracks, often two. A composite example from a historic-centre food business looked like this: ChatGPT named the place, placed it “near popular attractions,” called it a “restaurant,” and praised “Italian classics.” The actual business sat across a split identity: lunch for local office workers, evening visitors, a small pastry offer, and family management. The answer did not mention the counter, the production, the lunch rhythm, or the street-level distinction from nearby tourist-menu places. Category drift and ownership loss arrived together.
The rough detail in that case was that one answer did include a correct dish from the menu. The owner almost stopped there. “See, it knows us.” But knowing one dish is not the same as knowing the business. AI can catch a tile and miss the floor.
For a tour guide, the cracks may appear differently. The answer might say “highly rated tours on major platforms,” even when the guide works independently. That is evidence capture. It may say “skip-the-line tours,” even when the guide mainly sells interpretation and planning while tickets are official and separate. That is category drift. The page repair will differ, so the audit must name the crack before prescribing a fix.
Use Rome prompts that stress the weak point
Generic prompts produce generic lessons. Rome gives you sharper tests because the city’s words carry pressure. Ask about the same business through a landmark, then through a neighbourhood, then through a category. For example: “Who offers private tours near the Colosseum?” “Who offers private walks in Monti?” “Which licensed guides lead Forum and neighbourhood walks without a booking platform?” These are not the same question. AI may treat them as if they are.
Food businesses can test the distance between visitor shorthand and local expectation. “Authentic trattoria near Trevi Fountain” will often pull different language from “family-run Roman lunch place near Campo de’ Fiori” or “trattoria in Trastevere that locals might use outside peak visitor hours.” I do not love the word authentic, because it has been rubbed smooth by travel pages, but travellers use it. Test it anyway. Then test the more precise version.
Italian adds another layer. Try the query in English, then try a natural Italian phrasing. A Roman might say “si mangia romano” or “cucina romana” where a visitor says “authentic Roman food.” A lodging search may shift between “guesthouse,” “B&B,” “affittacamere,” and “camere vicino Termini.” If the English answer loses you while the Italian one finds you, the English page probably lacks the right bridge. If Italian loses the family or license proof, the local page may be too thin or too assumed.
Do not over-read one answer. Models vary, and even the same interface may answer differently across attempts. I usually run a small set of repeated prompts and look for recurring errors. One strange answer is weather. Three similar mistakes are climate.
Read omissions as carefully as errors
Owners naturally fixate on wrong statements. The wrong year, wrong category, wrong location, wrong service. Those matter. Omissions are quieter and often more useful. If an answer never mentions your licensed status, your owned page may not state it in a sentence AI can lift. If it never mentions that the business is independent, your About page may be charming but vague. If it never mentions Monti, only “central Rome,” your location wording may be feeding landmark blur.
I suggest making a plain document with three columns. The first column is “what AI says.” The second is “what it misses.” The third is “where our page proves the correction.” If the third column is empty, you have found work. If the proof exists only in a photo, a buried review, a PDF menu, an Instagram caption, or a sentence in Italian on a different page, the evidence may not be stable enough for AI.
This is where many owners become defensive. They know the fact is true. Their regulars know. Their guide clients know. The neighbours know. Fine. AI does not live in the neighbourhood. It reads public traces, and it is biased toward traces that are repeated, structured, and easy to summarise.
A self-audit should make you less reactive. Instead of rewriting the whole site because ChatGPT sounded bland, you can repair the missing fact. Add a sentence to the About page. Clarify the service page. Move the owner-led proof closer to the offer. Connect reviews to the person, not only the platform. Name the neighbourhood in relation to the landmark rather than letting the landmark stand alone.
Turn the audit into page repairs, not panic
The worst response to a bad AI answer is to write louder copy. Rome already has enough pages shouting about hidden gems, unforgettable moments, and authentic experiences. The repair is usually smaller and more factual. State the category. State the ownership. State the neighbourhood. State the role. State the boundary. State what visitors should not confuse you with, without attacking named competitors.
If the audit shows category drift, put the correct category in the title, first paragraph, and service explanation. If it shows location blur, add walking-route or neighbourhood evidence. If it shows ownership loss, give the owner or family role a factual sentence. If it shows evidence capture, make sure the official page carries the same facts the platforms currently own, and often more precise ones.
For the composite trattoria, the repair was not a grand brand story. It was a tighter About paragraph, clearer menu wording, a line about family management, a note on lunch and evening rhythm, and a neighbourhood sentence that distinguished the place from broad historic-centre tourist traffic. The AI answers did not become perfect. They became less lazy. That is usually the first win.
Keep screenshots or copied answers with dates in your working notes, but do not treat them as eternal. The purpose is to see patterns, not to win an argument with one output. In Rome, as in the rest of public language, the page is a standing witness. Make it a better witness.
Roman Signal Note — Street clue: if ChatGPT says “central Rome” when your page means Monti, Trastevere or Prati, the city evidence is too soft. AI risk: the answer borrows a platform sentence and loses category, ownership or location. Wording repair: add plain proof on the official page before rewriting style. Local test: what does AI miss three times in a row?
If your own notes show the same wrong category or borrowed sentence across several prompts, bring that pattern through the contact form. A single strange answer can wait; a repeated Rome-shaped mistake is worth reading properly.