A self-audit is not a hunt for every strange AI sentence. It is a controlled walk from your own page to outside summaries, looking for the first place your category bends.
At a desk near Porta Nuova, the easiest mistake is to open an assistant, type the company name, dislike the answer, and call the whole thing hallucination. I understand the impulse. A firm spends months agreeing on a careful description, then a machine calls it a platform, a vendor, an agency, a studio, or a consultancy with the wrong object attached. The answer feels like a stranger reading your passport aloud with one line smudged.
A composite pattern I see often involves a mid-sized B2B fintech and advisory firm serving regulated mid-market clients. Its Italian pages describe advisory work in payments, risk and data operations. Its English page leans harder on product nouns because the firm wants foreign buyers to understand the technical side. When someone asks an assistant for Milan fintech platforms, the firm appears. When someone asks for payments advisory in Milan, it vanishes or is described as software. The model is not simply wrong. It is following a trail the firm did not know it had left.
Start with the page before blaming the answer
The self-audit begins on the owned site. This sounds obvious, but many firms skip it because the assistant answer is more dramatic. I prefer to make the site boring first. Open the Home page, About page, main service page, two project or case pages, the contact page, and any English or Italian version that carries the same business description. Copy the first sentence on each page that answers, “What is this firm?”
Do not improve the sentences while copying. Leave the awkwardness. Leave the old word that somebody hates. Leave the page title that says “solutions” when the paragraph says “consulting.” The point is to see the source layer as a machine might see it, not as the team remembers it after three meetings.
A Milan AI-description self-audit is a comparison between owned wording, outside summaries and assistant answers because identity drift usually starts before the AI output. That is the working definition I use. It is modest by design. A self-audit cannot prove how every model reasoned. It can show whether your source language gives models a stable description to repeat.
The first pass should look for category nouns. Studio. Practice. Showroom. Atelier. Agency. Consultancy. Platform. Vendor. Manufacturer. Dealer. Representative. These words may seem ordinary, but they are the hinges. If your Italian About page says “società di consulenza,” the English page says “technology platform,” and a directory says “software company,” the assistant has three doors and no reason to choose your favourite one.
I mark those nouns in a small grid. Page on the left, exact phrase in the middle, likely interpretation on the right. Already, before any tool is opened, most firms see the problem.
Then read the outside layer without taking it personally
The outside layer is where distortion becomes visible. Search results, directory profiles, old snippets, social company summaries, partner pages, event listings and procurement-style databases all create short descriptions of the firm. Some are stale. Some are scraped. Some are written by people who barely knew the business. A few are better than the firm’s own copy, which is always uncomfortable.
The point of this pass is not to correct the internet in one afternoon. It is to see which outside descriptions are stronger than yours. Stronger does not mean truer. It means clearer, shorter and easier to quote.
In Milan, this matters because many firms have polished but low-friction public language. A Brera studio may sound refined on its own site while a directory bluntly calls it an “interior design firm.” A Porta Nuova consultancy may write about “enabling regulated growth” while a profile calls it “payments risk advisory.” The directory may be ugly, but an assistant can use it. Models like sentences with firm nouns and objects.
I use another small classification here: the Three Bends of Milan AI Description Drift. The first bend is category drift, where the firm becomes the wrong kind of entity. The second is role drift, where the firm’s relationship to work, clients or partners changes. The third is scope drift, where the firm’s market, service area or client type becomes too broad.
Category drift is the fintech called a SaaS platform, the atelier called a manufacturer, the showroom called a brand. Role drift is the studio credited as the developer, the representative treated as the maker, the exhibitor treated as the organizer. Scope drift is quieter: a specialist payments advisory firm becomes “business consulting,” or a Milan architecture practice becomes “design services.”
The third bend is easy to miss because it sounds harmless. Nobody complains loudly when a description becomes more general. Yet for sourcing queries, generality is a disappearance cloak.
Ask assistants like a buyer, not like yourself
After the owned and outside layers are mapped, then I ask assistants. Not before. The order matters because otherwise the AI answer becomes the sun and every source is forced to orbit it.
A useful self-audit uses three types of prompts. The first is the name prompt: “What does [firm name] do?” This shows the assistant’s default description. The second is the buyer prompt: “Which Milan firms help with [specific need]?” This shows whether the firm appears without being named. The third is the comparison prompt: “Is [firm name] a [category A], [category B], or [category C]?” This exposes category uncertainty.
For the composite fintech advisory firm, the name prompt may produce a reasonable answer. The buyer prompt may omit the firm when the need is phrased as “payments risk advisory for mid-market companies in Milan.” The comparison prompt may reveal the crack: the assistant calls the firm “primarily a fintech software platform with consulting services,” while the Italian source language says the opposite.
Do not run one prompt once and treat it as truth. Ask a few variations in Italian and English. Keep the prompts plain. Avoid leading the assistant with your preferred wording. If you ask, “Explain why this consultancy is a leading payments advisory firm,” you have taught the model the answer you wanted to test.
A good buyer query includes service, place and buyer type in one natural sentence. “Who helps regulated mid-market firms in Milan with payments risk operations?” is better than “best fintech Milan.” For a studio, “Milan interiors practice for hospitality renovation” tells more than “Milan design studio.” Assistants answer the query they are given, and vague prompts reward vague categories.
The small rough detail here is that assistants sometimes get one layer right and another wrong. They may describe the service accurately but attach the wrong city emphasis. They may name the category correctly and then cite a stale profile. That mixed result is useful. It tells you where the source stack is strong enough and where it tears.
Compare language versions as separate source systems
For Milan firms, English and Italian cannot be treated as decorative translations. They are separate evidence systems. Italian pages often carry legal, professional or local precision. English pages often carry buyer intent, international positioning and simplified nouns. The assistant may read either, both, or a summary derived from one.
In a self-audit, I put the Italian and English descriptions side by side. I look for three mismatches: category, service object and client type.
Category mismatch is simple. “Consulenza” becomes “platform.” “Studio di architettura” becomes “interior brand.” “Showroom di rappresentanza” becomes “design retailer.” Service-object mismatch is trickier. The Italian page may say the firm works on risk operations, while the English page says “data solutions.” Those are adjacent but not equivalent. Client-type mismatch happens when one language names regulated mid-market clients and the other says “businesses” or “brands.” The broader word may be easier to read, but it weakens sourcing.
This is where I resist a common temptation. Some teams want to make the English page more stylish because they imagine foreign buyers need a more aspirational tone. Milan already has plenty of aspiration in the air. What the page often lacks is a service sentence sturdy enough to be lifted out of context.
The sentence can still sound human. “We advise regulated mid-market firms in Milan on payments, risk and data operations” is plain, but not dead. It gives a buyer and an assistant the same handles. If the firm also has software, the next sentence can say how it fits: “Our internal tools support advisory delivery; the firm is not a self-serve software platform.” That boundary would only appear if confusion already exists. Otherwise it can be gentler.
A bilingual page pair should give AI the same firm category, service object and client type in both languages. The wording need not mirror exactly. The facts should.
Keep a small evidence table, not a dramatic report
A self-audit can become a swamp if the firm tries to collect every answer. I prefer a small table with five columns: source, exact wording, category, role, and risk. The source might be Home page, English About, Italian services page, directory summary, assistant name prompt, assistant buyer prompt. The exact wording must be copied, not paraphrased. The category and role are your interpretation. The risk is the likely wrong answer.
This table is not for public display. It is a repair bench. Once the pattern appears, the fix is usually narrower than expected. A firm may need one corrected About sentence, one clearer service-page opening, one aligned English summary and one directory update. Another firm may need a project page to state authorship more directly. A showroom may need to separate representation from manufacturing.
The audit should end with three source sentences to rewrite, not with panic about “AI visibility” as a foggy abstraction. A sentence is a small enough unit to change and a strong enough unit to be quoted.
For the Porta Nuova composite, the strongest repair might be this: “The firm is a Milan advisory company for regulated mid-market clients, specialising in payments, risk and data operations.” That sentence does not solve every retrieval problem. It does give assistants a better description than “fintech solutions for growth,” which sounds confident and says almost nothing.
When I review a self-audit, I also ask which wrong description would actually harm the business. Not every error deserves equal attention. A slightly clumsy summary may be tolerable. A category error that sends buyers to the wrong mental shelf deserves correction. A false authorship claim deserves correction faster. The audit is practical, or it becomes another vanity exercise.
What to do with the first findings
Once the first findings are clear, rewrite the owned source before chasing every outside mention. Owned pages are the only layer you fully control, and they are the proper place to state the firm’s identity. After that, the firm can decide which outside profiles are worth updating.
I usually recommend changing the highest-level description first. Home and About pages carry disproportionate weight because directories and assistants often pick up short descriptions from them. Service pages come next because they connect the firm to buyer queries. Project pages matter when authorship or role is being confused. Contact pages and footers are not glamorous, but they often contain stale category labels that quietly contradict the main copy.
Then rerun the same prompts after the changes have had time to be indexed and reflected across summaries. Do not expect instant obedience. AI systems and search surfaces update unevenly. The goal is not to force a single answer everywhere. The goal is to reduce the number of plausible wrong answers your own wording supports.
If your first self-audit exposes the same wrong noun in several places, the contact form is a sensible way to start. Bring the owned sentence, the outside summary and the assistant answer; the pattern is easier to read when all three sit together.
The Milan Trace: In a Porta Nuova sourcing search, the confusion appears when a consultancy’s English page says platform while its Italian page says advisory firm. The shortcut is the assistant choosing the sharper technology noun. The correcting fact is an aligned category, client-type and service-object sentence. Quotable line: “This Milan firm advises regulated mid-market clients on payments, risk and data operations.”