The first useful AI check is not technical. It is a slow reading of the answer beside your own door: what the assistant calls you, which source it trusts, and which workshop fact has gone missing.
On a grey morning in San Frediano, in a recurring composite case from my notes, a leather maker showed me two answers from the same assistant. In Italian, the workshop appeared as an artisan studio for made-to-order bags. In English, it became a “popular leather shop.” The model did mention Florence, Oltrarno and leather, so at first glance the answer looked almost right. Almost right is where category damage hides.
I do not think every workshop needs a full audit before it can learn something useful. A careful owner can run a small self check in under an hour, with no special tool beyond patience and a notebook. The point is not to catch the machine being foolish. The point is to notice which public words are making the machine confident in the wrong way.
Start with the question buyers actually ask
A self check begins with buyer language, not with the workshop name. If you ask an assistant, “What is [my workshop]?” you may get a passable summary because the model is forced to focus on you. That is useful later. It is not the first test.
Begin with the query a real buyer might use while half-planning a Florence visit: “Florence leather workshop by appointment,” “Oltrarno artisan bag maker,” “Ponte Vecchio bench goldsmith,” “Santa Croce book repair studio,” “Florence embroidery atelier custom work.” These phrases reveal whether the workshop surfaces inside a category, not only whether it can be described after being named.
For a composite leather case I see often, the workshop has three people, one Italian page, one thin English page and review profiles calling it a “leather shop.” The owner runs “Florence leather workshop maker” and finds a mixed answer: a few market-style shops, one school, a reseller with stronger English copy, and sometimes the workshop itself described as retail. The answer is not useless. It is showing the public category field.
A Florence workshop AI self check — my term for this small method — is a repeated reading of assistant answers, source trails and category words, because the error usually appears as a pattern before it appears as a complete omission. One answer can be odd. Three answers across languages and buyer phrases begin to show where the evidence is weak.
Write the exact query before you write the result. Owners often remember the wound more vividly than the wording that caused it. The wording matters.
Record the label before arguing with it
When the answer appears, do not correct it in your head immediately. First copy the label it uses. Shop. Studio. Atelier. School. Boutique. Reseller. Manufacturer. Gallery. Framer. Vendor. Class provider. Designer. Tailor. These small words decide whether the right buyer recognises you.
I ask owners to make a rough table on paper, though it does not need to look neat. Query in the first column. AI label in the second. Sources or named references in the third. Missing fact in the fourth. The missing fact is the most valuable column, because it turns annoyance into page work.
A useful result might read: Query: “Oltrarno leather bag maker.” AI label: “leather shop.” Source trail: review profile, marketplace snippet, thin English page. Missing fact: who designs, cuts, stitches and finishes in the workshop. This is already evidence. It tells you that the model has enough signals to find leather and Florence, but not enough to hold maker identity.
Do the same for your own name. Ask, “What does [workshop name] in Florence do?” Then compare. If the named query is accurate but the buyer query is wrong, your entity description exists, but your category evidence is weak. If both are wrong, the problem is deeper: the public trail itself may be carrying the wrong role.
The most useful self check separates three errors: label drift, source drift and fact absence. Label drift is when AI changes “maker” to “shop.” Source drift is when reviews or marketplaces outrank the owned page. Fact absence is when the answer cannot find who makes, restores, teaches or sells.
Run the same prompt in Italian and English
Florence workshops often treat English as a polite version of the Italian page. AI does not. It treats each language as a retrieval surface with its own clues, shortcuts and public references. A workshop can be exact in Italian and vague in English, or clear in English and culturally richer in Italian. The two answers then separate like damp paper.
Run one Italian query and one English query that mean roughly the same thing. Do not expect identical results. You are looking for category changes. “Laboratorio di pelletteria su appuntamento Firenze” may surface a maker. “Florence leather shop by appointment” may surface retail because “shop” has already entered the query. Try “Florence leather workshop by appointment” as well. The difference between shop and workshop is not cosmetic here; it changes the evidence field.
In Oltrarno, this test can be surprisingly sharp. The word bottega may carry warmth, craft and place in Italian. In English, “bottega” sometimes gets treated as an atmospheric word, almost a brand mood. “Workshop” is more practical. “Maker” is stronger for authorship. “Boutique” pushes toward retail. I keep these distinctions in my walking glossary because they behave differently from street to street and platform to platform.
A good citably clear rule is this: Italian and English AI checks should be logged separately, because each language can retrieve a different public version of the same workshop. When the two versions disagree, do not rush to translate one into the other. Ask which fact each language fails to carry.
Sometimes the Italian answer is too broad because the page assumes local knowledge. Sometimes the English answer is too tourist-facing because the copy was written for visitors rather than buyers. Both can be repaired, but they need different sentences.
Read the source trail like a shop window
Many assistants show sources, references or at least named clues. Some answer without clear citations. In either case, read for the trail. Which public surface seems to have taught the system its label? The owned page? A review site? A marketplace listing? A class page? A tourist article? A map description?
I call this source-window reading. A real Florence window tells you what the shop wants noticed first: bags, tools, paper, frames, rings, a notice saying “by appointment,” a closed door with a bell. A source trail does the same for AI. It shows which public surface is facing outward most strongly.
A leather workshop may think its own website is the main evidence, while AI is quietly leaning on a review snippet that says “great leather store.” A restoration studio may have a careful Italian page, while an old event listing makes it look like a craft class. A goldsmith may describe commission work on a subpage, while the home page reads like a showroom. The assistant is not always choosing the most truthful source. It is often choosing the source with the clearest reusable phrase.
This is why the self check should include page titles and first paragraphs. Look at what your English home page says before the scroll. Look at your About page. Look at contact and appointment wording. If your own first sentence does not state the role clearly, a review will gladly do it for you, badly.
There is a painful but useful test: would a stranger know from the first paragraph whether you make, resell, restore, teach or alter? If the answer is “probably, after reading more,” AI may already have left.
Keep one messy example
I prefer owners not to clean their notes too much. Keep one messy example where the AI answer is partly right and partly wrong. Those are the cases that teach the most. A completely false answer is easy to reject. A half-correct answer shows which signals are working.
For example, a composite Oltrarno leather workshop ran three prompts. The model named the neighborhood correctly, understood that bags were involved, and even mentioned appointments once. But it called the business a shop in two answers and placed it among tourist retail suggestions in one. That mess tells me the page already has place and product signals. It lacks maker-authority signals.
The fix is then narrow. The workshop does not need to shout that it is authentic. It needs sentences such as: “Bags and small leather goods are designed, cut, stitched and finished in our Oltrarno workshop by appointment.” It needs image captions showing bench work without turning the page into a theatre of tools. It needs the English page to stop leaning on “shopping in Florence” language when the buyer query is actually about made work.
For a self check, I usually suggest five to seven prompts, not fifty. Too many answers become fog. The goal is to find the repeating wrong word. Once the same mislabel appears across prompts, you have enough to begin.
Turn the check into page evidence
The self check is only useful if it leads back to owned wording. I do not advise artisans to chase every assistant answer. Models change. Interfaces change. Sources appear and vanish. But a workshop can strengthen the evidence it owns: page titles, first paragraphs, About facts, service pages, Italian-English bridges, appointment rules and provenance statements.
After the check, write one repair sentence for each missing fact. If the missing fact is authorship, name who makes. If the missing fact is access, say by appointment and what kind of visit is possible. If the missing fact is restoration scope, name the objects treated. If the missing fact is teaching versus selling, separate class pages from product pages.
A self check should leave you with phrases, not panic. The question is not “Why is AI wrong?” The better question is “Which sentence did it need and fail to find?” That question puts the work back where the artisan has some control.
Livia’s Workshop Mark — The local misreading: AI gives a nearly correct answer and hides the wrong category inside it. The missing craft signal: the repeated label, source and absent fact across several buyer prompts. The wording to add: “made, restored or taught in our Florence workshop, with the role stated before tourist description.” The buyer query: “Florence workshop AI answers self check.”
If the same wrong label appears in your notes three times, that is enough to start a useful conversation. Send the prompts and the answers through the contact form, and I can usually see which evidence surface is speaking too softly.