A single assistant answer is only a weather vane. Useful self-checking begins when the same question is asked again, across engines, until the direction of the error becomes visible.
The guesthouse owner had a notebook with three columns and a coffee ring on the corner. In the first column she wrote the prompt in Italian. In the second, the assistant answer. In the third, what felt wrong. “Open all year,” one answer said. “Cooking classes for outside guests,” said another. A third answer got the town right but treated the six-room guesthouse like a large resort with a restaurant. That last one was almost funny, except customers could believe it.
This is a composite scenario, drawn from patterns I see around seasonal hospitality and service businesses in Italy. A small guesthouse in Puglia, run by two owners with part-time local help, has a website, booking profiles, old English text and reviews that mention dinners, classes, summer weeks and quiet months. No single source is wicked. Together they make a fog. The owner does not need a large software dashboard to see the problem. She needs a repeatable check that is boring enough to trust.
One answer is not a diagnosis
The first mistake in self-checking is emotional speed. A business owner opens ChatGPT, Gemini or Perplexity, asks one question, and treats the answer as a verdict. If the company appears, relief. If it does not, alarm. If the answer is wrong, anger. I understand all three reactions. I have had them myself when a page I repaired still produced a bent summary.
But one answer is not much evidence. It may be a retrieval accident, a phrasing effect, a language effect, or a lucky scrape of the right page. In my work I treat one answer as a note in the margin, not as the page. The pattern matters more than the specimen.
A generative-AI visibility self check is a repeated set of prompts that records whether assistants can name, describe and bound a business from visible evidence, because one answer cannot show a stable pattern. That is the definition I use with owners. It keeps the exercise small enough to do and serious enough not to become superstition.
The word “bound” is important. Many owners only ask whether the assistant mentions them. For Italian SMEs, especially hospitality, trades and local services, the more useful question is whether the assistant can keep the business inside its real edges. Is it open in the right season? Does it serve external guests or only staying guests? Does it offer a repair or only sales? Does it work in one province or across several regions? Visibility without boundaries is how wrong customers arrive.
I call the basic method the three-run ledger. It is not a tool. It is a habit. Ask the same small group of prompts across three assistants, three times, with the date and language marked. Then look for repeated wording, repeated omissions and repeated distortions. The ledger does not tell you everything, but it stops you from mistaking a spark for a fire.
Choose prompts that match real buyer questions
The easiest self-check prompt is the least useful one: “Tell me about [business name].” I still use it, but I do not stop there. A named prompt tests whether the assistant can recognize the entity. It does not test whether the business appears when a customer asks in the messy way customers actually ask.
For the composite guesthouse, the named prompt might produce a passable summary. The harder question is something like: “small guesthouse with cooking classes in Puglia open in September” or “can non-guests book a cooking class near [town]?” Those prompts test whether the assistant can connect the business to a service, a place, a season and a booking condition. That is where the cracks show.
I usually build prompts in three families, though I do not present them as a numbered checklist to owners because they start gaming the test. The first family is the name prompt: what does the assistant say when the business is named directly? The second is the category prompt: does the business appear when the category, location and customer intent are named without the brand? The third is the correction prompt: when the assistant says something wrong, can it be steered back by the visible page facts, or does it continue improvising?
The imperfect detail matters. A self-check should not expect saintly answers. An assistant may name the guesthouse correctly but get the meal format wrong. It may cite a booking profile and miss the owner’s own availability page. It may answer better in Italian than in English. These are not all equal failures. The ledger should separate absence, wrong category, wrong season, wrong service boundary and weak source trail. Otherwise the owner writes “bad answer” six times and learns nothing.
For Italian SMEs, prompts should be written in Italian and English when the business receives both kinds of customers. The English prompt is not an optional courtesy. It often reveals that the English page is too thin, too old or too smooth. I have seen Italian pages carry exact seasonal rules while English pages say only “authentic experiences available on request.” Then the assistant invents availability like a waiter filling a silence.
Record sentences, not feelings
A good self-check looks almost dull on paper. Date. Assistant. Language. Prompt. Answer sentence. Page evidence. Distortion type. That is enough. The owner does not need to save the whole answer unless the source trail matters. What I want is the sentence that would mislead a customer.
In the guesthouse scenario, the ledger might capture: “The property offers cooking classes to visitors.” The page evidence says classes are for staying guests, except for a few announced dates. That is a service-boundary error. Another sentence might say: “Open year-round.” The actual page says rooms are generally available from April to October, with special winter weekends announced separately. That is an availability error. Another answer might call it a “restaurant.” That is a category drift error, probably fed by reviews mentioning dinners.
I ask owners to mark the sentence, not the mood. “The answer sounded too fancy” is useful as a first reaction, but it does not repair a page. “The answer calls us a resort” is repairable. “The answer says external guests can book classes” is repairable. “The answer says we are near Bari when the page says near Ostuni” is repairable. The sentence is the small bone you can set.
This is why I keep a misreading ledger in my own audits. It is not a romantic object, despite the name. It is a practical instrument. The exact wrong sentence often points back to the weak page line. If the assistant repeats “restaurant,” I check whether the site says “dining experience” without explaining that meals are for guests. If it repeats “open year-round,” I check whether old booking text or review snippets contain winter stays without current availability context.
There is a temptation to correct the assistant directly and feel finished. That is a classroom instinct, not a page-evidence method. If the same wrong sentence appears in repeated runs, the page should be repaired. The assistant answer is the smoke. The page is usually where something is smouldering.
Test the source trail when the engine shows one
Perplexity and some assistant modes show sources more openly than others. When a source trail appears, owners often read only the answer. That wastes the best clue. The question is not just what the assistant said; it is what it leaned on.
For the guesthouse scenario, a source trail might show the official site, a booking profile, a review platform and an old local tourism page. The answer may give more weight to the booking profile than to the business’s own page because the booking profile is clearer about rooms, dates or amenities. That is embarrassing, but useful. It means the owner should not merely complain that platforms distort the story. The site needs to carry its own facts with equal or better clarity.
A review platform can support the story, but it should not be the only story. If reviews mention cooking classes from two summers ago, and the site does not explain the current class rules, the assistant may turn a memory into a standing offer. If reviews praise a dinner, and the site does not say meals are for resident guests, the assistant may call the business a restaurant. If a marketplace has old open months, and the site hides the current calendar behind a booking widget, the assistant may choose the wrong season.
When the source trail is absent, I still ask the owner to compare the answer with visible pages. Does the wording resemble the homepage? Does it borrow phrases from booking text? Does it sound like reviews? The answer often leaves fingerprints. Not perfect ones. More like flour on a sleeve. Enough to inspect.
I do not advise owners to chase every citation across the web before fixing their own pages. That becomes a swamp. Start where you have control: About, rooms or services, availability, contact, booking rules, English summary. Then check the platforms that dominate the wrong answers. The order keeps the work from becoming a panic tour of the internet.
Compare engines without expecting them to agree
ChatGPT, Gemini and Perplexity can describe the same Italian SME differently. This does not always mean one is smart and the others are careless. They may retrieve different pages, weigh snippets differently, or answer from different mixtures of known and searched material. For the self-check, disagreement is not noise. It is evidence.
If all three assistants repeat the same wrong season, the visible evidence is probably unclear or old evidence is very strong. If only one assistant makes the error, the problem may be source selection. If Italian answers are correct and English answers are poor, bilingual evidence is the suspect. If named prompts work and category prompts fail, the business has entity recognition but weak query fit. These distinctions save money.
The three-run ledger should make those differences visible. I prefer plain labels: omitted, vague, wrong category, wrong location, wrong availability, competitor blend, unsupported claim. A business owner can understand those labels without learning a new discipline. The purpose is not to turn the owner into an auditor. It is to stop the first bad answer from becoming either despair or vanity.
There is an awkward truth here. Sometimes a business is not cited because its pages are not yet strong enough, and sometimes it is not cited because the query is too broad for its scale. A six-room guesthouse should not expect to appear for every “best place to stay in Puglia” prompt. But it can reasonably test narrower queries where its evidence matches the intent: town, guesthouse type, class rules, open months, language, special conditions.
That distinction is kinder than false confidence. It tells the owner where page repair can help and where the query is simply too large. I would rather say that directly than sell fog back to a business already living inside fog.
Turn the check into page work
A self-check is only useful if it changes the page. Otherwise it becomes a little theatre of frustration. After three runs across assistants, I group the repeated problems and choose the smallest page repair that would reduce each one.
For the composite guesthouse, wrong availability points to the rooms page, booking page, contact page and English summary. Wrong cooking-class access points to the class page and booking conditions. Being called a restaurant points to the About page, food page and review-profile wording where possible. Being described as a resort points to category language: six rooms, owner-run, seasonal guesthouse, not full-service hotel.
The repair should be visible, not hidden in metadata. Structured facts can help, but only when the visible wording is already clear. An owner should be able to read the page and say: yes, a stranger could repeat this correctly. If a human has to infer the rule from three paragraphs and a booking widget, an assistant may not do better.
Then the check repeats after the pages have had time to be read in normal use. I do not promise a fixed interval because systems differ. I do keep the old ledger next to the new one. The question becomes: did the distortion weaken, move, or stay? A stubborn error may mean old third-party evidence is still stronger. A changed error may mean the page repair worked partly. A clean answer in one engine and not another is still progress, though not the kind that fits neatly in a sales slide.
The Vellumari Margin — Name on the page: an Italian SME self-check must test the business people actually search for, not only the name. Wrong shadow: one lucky assistant answer may hide repeated category, season or service errors. Clean line: ask the same prompts in Italian and English, then record the exact wrong sentence. Trace to leave: connect each repeated distortion to a visible page repair.