A baseline is not a dashboard with impressive furniture. It is a repeatable little table that stops one strange assistant answer from becoming the whole story.
A composite scenario from regional service-page work began with a safety-training and compliance advisory firm between Modena and Bologna. It served small manufacturers, workshops and logistics depots, and its site ranked for several course and workplace-safety phrases. The owner had a tidy ranking export, a half-finished English page, and one assistant answer that made the whole room tense. Asked for firms that could help an Emilia-Romagna workshop with mandatory safety training and document follow-up, the assistant cited a trade-directory page, named a larger competitor, and described the local firm as a “document consultant.” The firm was not entirely absent, but it appeared only after the query included its brand name.
The temptation in that room was to rewrite everything. New service pages, cleaner category copy, new English wording, directory cleanup, all at once. I understand the impulse. When an answer leaves you out, it feels personal, almost like hearing your name omitted from a guest list while standing by the door. But the first job is measurement. Before changing pages, you need to know how often the omission happens, on which queries, in which language, and through which cited sources.
A baseline is a small repeatable truth
A citation baseline is a first measurement of how often assistants name, cite and describe a business for selected queries, because single screenshots cannot show a pattern. That is the definition I use. The words “selected queries” are doing heavy lifting. A baseline is never the whole market. It is a shelf of questions chosen well enough to reveal whether the business is visible in answer form.
I call the first version the name-source-description ledger. It has no glamour. Query. Language. Assistant used. Date. Business named or not. Other businesses named. Cited sources. Own site cited or not. Description accurate or not. Wrong noun. Missing proof. Next page to inspect.
This plain table does something useful to the room. It slows the panic down. One run may be strange. Five runs across related queries may show a pattern. Twenty runs may show that Italian service queries behave differently from English buyer queries. The table does not remove uncertainty, but it gives uncertainty a chair and a pencil.
The important rule is to build the baseline before major edits. If you rewrite first, you lose the old ground. You can still improve the page, of course, but you will not know whether the assistant changed because of your edits, a source refresh, a query variation, or ordinary answer drift.
Choose queries that buyers would actually ask
A baseline built from old SEO keywords alone will usually undercount the problem. Traditional keywords are often clipped: “corsi sicurezza Modena,” “consulenza sicurezza lavoro,” “documenti sicurezza azienda.” Buyers using assistants ask in fuller shapes. They ask for recommendations, comparisons, suitability and local fit. They may include “for a workshop,” “near Bologna,” “with training and document support,” “not only an online course,” or “for a small logistics depot.”
For the composite advisory firm, I would not start with a hundred keywords. I would start with perhaps fifteen queries. Some would be Italian category queries. Some would be buyer-problem queries. A few would compare training, consultancy and document follow-up, because the company was being narrowed into the wrong noun. A few would use English, since international managers sometimes ask in English even when the work happens in Italian.
The roughness matters. Real buyer questions are not tidy keyword strings. They contain doubts, place names, constraints and clumsy phrasing. A baseline made only of perfect terms becomes a polished brass instrument that nobody in the street can play.
I usually divide the first query shelf into four groups: category, need, comparison and local fit. Category asks what the business is. Need asks what problem it solves. Comparison asks which options belong together. Local fit asks whether geography changes the answer. This “four-shelf baseline” is not a law. It is a way to avoid measuring only the easiest corner.
Record the answer, not your memory of it
The worst baseline is the one reconstructed after the call from memory. Someone says, “It mentioned us once,” or “It always cites that directory,” or “The answer was mostly right.” Mostly right is not a measurement. It is a mood.
Copy the exact names. Copy the cited source titles or URLs where available. Copy the sentence that describes the business. If the assistant gives no citation, record that too. If the same answer names the company but calls it a document consultant, keep the wrong word. If it cites a directory that lists an old service category, write that down. The unattractive details are the useful ones.
In the advisory-firm scenario, one answer named the business only after a follow-up and then described it as helping companies “prepare safety paperwork.” That was not false, exactly, but it shrank training, inspection preparation and ongoing support into paperwork. Another answer cited a trade directory whose page title sounded relevant, while the body barely separated training from general consulting. Without copying the wording, the team might have called that a partial win. With the wording in the ledger, it became a source gap.
There is also a date problem. Assistant outputs can change. Source indexes change. Pages change. The same query may not return the same answer later. A baseline does not pretend to freeze the world. It records enough context so a later check can be compared without theatre.
Keep ranking in a separate column
The baseline should include ranking position, but it should not let ranking position run the meeting. I keep it in a separate column because it answers a separate question. Ranking asks whether the page appears in search. Citation asks whether the business survives the assistant’s selection and wording.
This separation prevents two common mistakes. The first mistake is comfort: “We rank well, so the AI issue cannot be serious.” The second is overreaction: “We were not cited, so our SEO must be broken.” Neither follows. A business may have strong search visibility and weak answer evidence. Another may have modest rankings but enough directory and page clarity to be named.
The four-shelf baseline makes this visible. A page might rank well for category queries but be absent in comparison queries. It might be cited in Italian but misdescribed in English. It might appear through an aggregator for local fit queries while its own page is skipped. Those differences point to different fixes.
For the safety-training firm, ranking was strongest on course terms. Citation weakness appeared when the query asked for a named recommendation by business situation. That distinction mattered. The page was not failing to be found. It was failing to become the safe named answer for a specific buyer need.
Read missing competitors as evidence
A baseline should not only count your own business. It should count the businesses that appear instead. This is where a small table becomes sharper than a complaint. The competitor named by the assistant is a piece of evidence. The source cited for that competitor is another piece. Together they show what the assistant found easier to trust.
Sometimes the competitor has a clearer category page. Sometimes it appears in a list that matches the query language. Sometimes it has a branch page with a cleaner local sentence. Sometimes the cited source is poor but consistent, which is a strange advantage. Consistency can beat richness when the answer needs one safe line.
I do not like turning every competitor mention into a drama. Some queries simply belong to another firm. Some listicles are lazy. Some assistant answers are odd. But when the same outside source appears across several queries, the baseline is telling you something. It may be an evidence competitor, not merely a traffic competitor.
For Italian businesses, this often includes aggregators and directories. They may be thin from a sales point of view, yet strong from a citation point of view because they organize names, categories, locations and snippets in regular patterns. A business page that tells a richer story can still lose if the basic evidence is scattered.
Decide the first fix from the pattern
After the first baseline, the question is not “how do we improve AI visibility?” That question is too large. Ask a smaller one: which repeated failure deserves the first fix?
If the business is never named, inspect category and source discoverability. If it is named but cited through an aggregator, compare the aggregator’s wording with the business page. If it is cited but misdescribed, fix the sentences that define service, audience and boundary. If Italian queries perform well and English queries fail, do not merely translate; inspect whether the English page answers the English buyer’s actual phrasing.
This order protects the team from cosmetic rewrites. I have seen pages made prettier while the wrong noun remained. I have seen headings improved while the location proof stayed buried. I have seen English pages expanded with confident language that still never said the service covered the province named in the query. The baseline is there to keep the work honest.
A first citation baseline is modest by design. It will not tell you everything. It should tell you enough to stop arguing from a single answer and start counting a pattern. That is already a large step for teams used to treating ranking position as the final number.
The Citation Ledger
Query shelf: “misurare citazioni IA azienda.” Ranking residue: old reports show which pages rank, but not whether assistants name the business. Citation hinge: a baseline counts the answer triad across a chosen query shelf before edits change the ground. Next count: repeat the same queries, record names, cited sources, description accuracy and missing competitors, then choose one page fix from the pattern.