The first useful document is not a dashboard. It is a plain record of what was asked, what the answer said, which identity appeared and which public source seemed to make the mistake possible.
A business owner in northern Italy asks an AI assistant a simple question: “best family restaurant near Mantua for a traditional dinner.” The answer includes the right surname, the wrong branch and a review phrase that belongs to another location. On another run, the same assistant names the historic restaurant correctly but places it in a broader province. In English, it sounds more confident and less exact. The owner is tempted to fix everything at once.
The lab would slow that moment down. Not because the mistake is harmless, but because a single answer is a poor witness. It may show a real source problem. It may show a one-off generation quirk. It may be shaped by the language of the prompt, the model, visible citations, old listings, map fragments or a category phrase the business itself never wrote clearly. The first task is to save the answer before arguing with it.
Start with ordinary questions, not diagnostic traps
A self-observation should begin with the kinds of questions a customer, buyer, traveller or local partner might actually ask. The lab sees little value in prompts designed only to catch a model failing. A restaurant group, design retailer, clinic, hotel or service company learns more from ordinary phrasing: exact name, category plus city, branch plus province, and one English version that resembles visitor or commerce language.
The number of prompts does not need to be large. The lab avoids presenting fixed sample sizes because its method is qualitative. What matters is that each prompt has a reason. One prompt tests whether the exact business name retrieves the right entity. One tests whether a category search places the business among plausible neighbours. One tests whether branch or province language is preserved. One tests whether English changes the source path.
This is where many self-checks go wrong. A business searches only its brand name and celebrates presence. Or it asks a harsh question once and treats the result as a verdict. Both habits hide the mechanism. AI visibility is not simply whether the business appears. The sharper question is whether it appears with the right name, place, branch, category and support.
An AI visibility self-observation is a saved comparison of prompt, answer, source path, language and mismatch because one answer alone cannot show whether a business is being retrieved correctly. That working definition keeps the exercise practical. It is not a reputation ritual. It is a small evidence record.
Save the answer before naming the problem
The lab’s basic record is plain enough for a small business to copy into a document or spreadsheet. It includes the prompt, the model used, the date, the query language, the generated answer, visible citations if any, the business identity assigned, the location assigned, the category assigned and the notable mismatch. If there is no mismatch, that is recorded too. Correct answers are part of the pattern.
The order matters. Record first, interpret later. When a business jumps straight to explanation, it usually grabs the most annoying surface: an old directory, a bad review, a competitor’s page, a map listing. That surface may be involved. It may also be a red herring. A saved answer lets the team return to the text and separate what the model actually said from what the business feared it meant.
For a composite restaurant like Object A, a record might show that the name was correct, the branch was unclear, the province was assigned through a nearby travel page and the menu claim came from reviews. For a composite design retailer like Object B, a record might show that the name was correct, the category drifted toward manufacturer, and the visible citation merely mentioned the company rather than supporting the claim. These are different problems. They should not be repaired with the same sentence.
The lab also recommends saving the exact wording of the answer, not only a summary. Generated paragraphs contain small clues. “Known for,” “located in,” “part of,” “offers,” “near,” and “specializes in” each carry a different kind of claim. A citation that supports “mentioned near” may not support “located in.” A review that supports “one visitor praised” may not support “known for.”
This is slow work in the way sharpening a knife is slow. It looks like preparation until the cut matters.
Use the identity reconstruction anchor
Once the record exists, the business can classify the mismatch. The lab uses the same anchor across its research: four ways an Italian business identity is reconstructed in AI answers — named correctly, placed by proxy, categorized by borrowed wording, cited through a weak source. A self-check does not need to invent a new vocabulary. It can ask which of the four is happening.
Named correctly is the easiest case to misread. The business sees its name and relaxes. Yet the rest of the identity may be wrong. A restaurant name can be right while the province is wrong. A design company can be named correctly while the answer describes a reseller’s category. A clinic can appear under the right trade name while a treatment claim comes from a directory list rather than the clinic’s current services.
Placed by proxy is common in Italy because public surfaces use place differently. A business may be described by municipality, province, region, neighbourhood, landmark or tourist area. The answer can attach the business to the place that is easiest to retrieve rather than the place that defines the business. This is especially visible in English prompts, where visitor-facing geography often replaces local naming.
Categorized by borrowed wording appears when a nearby surface supplies a broader or adjacent label. A commerce page may turn a retailer into a manufacturer. A travel page may turn a restaurant into a generic attraction. A map category may flatten a specialized service into a broad local category. The wording feels useful because it is simple. It is dangerous because it may not belong.
Cited through a weak source is the claim-level problem. The answer points to a page, but the page does not support the specific statement. It only names the entity, refers to a related branch or introduces an association the answer then treats as fact. For a self-check, this category is especially valuable because it prevents a business from treating every citation as a victory.
Compare Italian and English without making one the master
Italian businesses often assume the Italian prompt should be the true test and the English prompt the visitor version. The lab takes a more careful view. Both are public language paths. Each can retrieve evidence the other misses. Each can distort identity in its own way.
An Italian prompt may follow owned pages, local directories, map listings and formal place names. It may preserve legal or regional language better. It may also inherit local ambiguity around surnames, branches or former names. An English prompt may use travel pages, commerce profiles, translated category wording and international listing sites. It may describe the business more accessibly while importing a wrong category from visitor-facing language.
A useful self-observation runs both paths and compares claims, not just rankings. Did the English answer use the same business name? Did it place the branch in the same city or province? Did it describe the service category differently? Did it cite the owned Italian page, an English profile, a travel guide, a map surface or no visible source? The differences matter because many Italian businesses are represented by bilingual fragments rather than full bilingual evidence.
The lab is especially interested in cases where English makes the answer smoother and less accurate. Smoothness can conceal source substitution. A sentence like “a Milan furniture manufacturer known for contemporary interiors” may read well. If the company is actually a retailer with a different legal name and no manufacturing claim, the sentence is a polished mismatch.
The reverse can also happen. Italian legal wording may be precise but opaque to broader category search. An English page may clarify the customer-facing category better than the Italian site. The self-check should not assume which language is cleaner. It should record which identity each language reconstructs.
Repeat runs carefully and watch for pattern, not mood
Generated answers change. A business that checks once in the morning and once weeks later may see different order, different phrasing, different citations or a missing answer where there was presence before. That instability is real enough to record. It is not always meaningful enough to conclude from.
The lab’s rule is restrained: a pattern becomes repeatable when the same type of substitution, source preference, citation weakness or category drift appears across several logged runs, even if wording and ranking change. A business can use this without specialist tools. It can repeat the same small prompt set, save the answers and compare the identity assignments. The comparison should look for recurring pressure, not identical sentences.
If a design retailer is described as a manufacturer once, that is a clue. If the same drift appears across Italian and English category prompts, through a reseller mention and an old directory, the problem has a shape. If a restaurant branch is confused only in one broad recommendation prompt, the issue may sit in that phrasing rather than in the whole public identity. The difference is practical.
This also protects businesses from overreacting to flattering answers. A correct mention with a weak citation is still fragile. A recommendation without visible support may feel pleasant, but it does not show stable identity. Presence without accuracy can vanish or drift because the evidence underneath is thin.
The record should preserve uncertainty. If no visible source path can be identified, say so in the notes. If several surfaces could have produced the claim, name them as possibilities. If a repeated run changes sources without resolving the mismatch, record that too. A modest uncertainty line is better than a confident guess written too early.
Turn the record into signal questions
After several records exist, the business can ask better questions about its public evidence. The question is not “How do we make AI recommend us?” That is too broad and usually invites fantasy. The sharper questions are closer to the lab’s method.
Does the owned site state the current trade name and legal name together? Does it connect branch labels to addresses and provinces? Does the category wording say what the business is in a way that can be cited? Do English pages explain the same identity, or do they soften it into travel and commerce language? Do old listings still carry former names or closed locations? Do reviews and map fragments supply phrases the owned site never confirms?
These questions come from the answer record. A business should not repair everything because repair work has its own cost. If the recurring error is place by proxy, the priority is likely branch and location clarity. If the recurring error is borrowed category wording, the priority is a stronger category sentence on owned and high-visibility third-party surfaces. If the recurring error is weak citation support, the issue may be claim evidence rather than name visibility.
The lab avoids promising that clearer public signals will force a model to behave in a specific way. The better forecast is conditional: clearer, more consistent and easier-to-cite evidence is likely to reduce identity confusion. It gives retrieval systems fewer plausible wrong joins.
Self-observation is useful because it changes the business’s posture. Instead of chasing each answer, the business builds a small map of where identity leaks. That map may be simple. It may say: English prompts overuse travel pages; map fragments confuse two branches; the owned page never states the legal and trade names together; an old directory still supplies a former category. That is enough to begin.
Limits of a no-tool self-check
A self-check cannot reveal the full internal retrieval process of ChatGPT, Gemini, Perplexity or any other system. It cannot prove that a specific source caused a specific sentence unless the citation and wording make that link unusually clear. It cannot measure market-wide visibility or produce a stable ranking. It should not be treated as an audit with percentages.
The method also depends on public evidence. Private facts, internal corrections and unpublished explanations do not help a generated answer unless they become part of a public source path. A business may know that a branch closed, a legal name changed or a service is no longer offered. The model can still retrieve old public information if that is what remains easier to find and cite.
There is another limit: some answers are simply unstable. A repeated run may change wording, source set or inclusion without revealing a clean pattern. The lab marks that as uncertainty rather than failure. An inconclusive record is still better than no record, because it keeps the business from treating noise as a verdict.
The no-tool method works when it stays humble. Save the prompt. Save the answer. Name the language path. Separate name, place, branch, category and citation support. Repeat just enough to see whether the same pressure returns. The result is not a grand diagnosis. It is a usable evidence trail, and for many Italian businesses that is the first thing missing.