A blue link asks the buyer to decide what matters. An assistant answer has already made part of that decision, then wraps it in a sentence that may travel further than the link.
In a composite scenario from tourism-adjacent work, a 14-person Italian service business had offices in Florence and Venice. The local pages were not weak. They appeared for several city queries, especially in Italian. A founder showed me the rankings with the cautious pride of someone who had paid attention for years. Then we asked an assistant for English-speaking help in Florence, with a branch-specific need. The answer cited an aggregator, named a competitor, and blended the Venice office into the Florence description. The company was visible in search and blurred in the answer.
That blur is where the question “cosa sono citazioni IA” becomes practical. An AI citation is not just a new kind of backlink, and it is not the same as a search position with a prettier wrapper. It is a source relationship inside a generated answer. The assistant chooses a name, attaches or implies a source, and writes a description that may be correct, partial or quietly wrong.
A ranking is outside the answer
On a search results page, the ranking sits outside the buyer’s final interpretation. It is a position on a list. The search engine says, roughly, “these pages may answer you; here is an order.” A human still has to click, compare, distrust, skim, return, and maybe search again.
Inside an assistant answer, the list has already been digested. The user sees fewer names, fewer sources and more confident prose. This gives the answer a calmer surface. It also hides several choices. Which businesses were considered but not named? Which source supplied the description? Which detail came from the business site, and which came from a directory or review page? Was the cited source actually about the branch the buyer asked for?
Ranking, in that setting, becomes one ingredient. It may help a page get found. It may suggest popularity or relevance. It may feed retrieval. Yet the answer still has to turn documents into a sentence. That sentence is where many Italian businesses discover they are being represented by someone else’s wording.
The difference feels small until a buyer acts on it. A local page in Florence can rank; an assistant can still recommend a competitor because the competitor’s source trail says “English-speaking visitors, Florence office, private itinerary planning, licensed local partners” in a cleaner way. The first page had position. The second had answer-ready evidence.
What an AI citation actually is
An AI citation is a source reference inside a generated answer, because the assistant uses that source to support a named claim, description or recommendation. That is the working definition I use with clients. It keeps the focus on three elements at once: source, claim and business name.
If only the source is present, you may have a citation without useful visibility. If only the name is present, you may have selection without a source you can inspect. If the name and source appear together but the description is wrong, you have a citation problem disguised as a win. The neat little link is not the whole event.
I call this the answer triad: name, source, description. Every AI visibility check for a business should separate those three. Did the assistant name the business? Which source did it use or cite? How did it describe the offer, location and fit? A ranking report cannot answer those questions by itself. It was built for a different surface.
In the Florence and Venice example, the business sometimes appeared as a name but the source attached to the answer was an aggregator page with stale branch wording. One run put a Venice-only service into the Florence answer. Another named the company correctly but described it as a tour operator, which was too narrow. The ranking report was calm. The answer triad was messy.
The link may be less important than the sentence
SEO people, myself included, are trained to look for the link. Where is it? Which URL? Which domain? Do we own it? Does it pass traffic? Those questions still matter. Yet in assistant answers the sentence can become the more damaging object.
A buyer may read the description and never click. If the answer says the business is a consultancy, many software buyers will leave it aside. If it says the Florence branch handles only groups when private work is the real strength, the wrong clients arrive or the right ones never do. If it says “near Venice” for a Florence query, the branch has been smeared across the map.
This is why I do not count AI citations as simple mentions. I count the wording. A cited source that produces a false description is a cracked tile in the floor. People may walk over it for months before someone notices why they keep tripping.
There is a quiet discipline here. Write down the exact phrase the assistant used. Do not paraphrase it into something nicer. Keep the ugly noun if it used one. “Agency,” “platform,” “consultancy,” “tour company,” “software provider,” “directory,” “local operator” — these nouns are classification decisions. When they are wrong, the citation is pointing to the wrong shelf.
How ranking enters the generated answer
It is tempting to ask where the ranking “goes” inside the answer, as if the assistant swallowed the search result page and rearranged it. The mechanism is less tidy. Different assistant products and retrieval systems behave differently, and I do not pretend to see inside each one. In practical checks, though, I see four places where old ranking evidence may still matter.
First, ranking can affect discoverability. Pages that are easy to find in search are often easier for tools and source indexes to encounter. Second, ranking pages often have old SEO work behind them: internal links, headings, service pages, structured titles, local clues. Some of that helps extraction. Third, ranked pages may be repeated by other sites. Directories and aggregators often copy or compress business descriptions from visible pages. Fourth, the user’s query may resemble the old keyword closely enough that search-like retrieval remains influential.
Yet none of these guarantees a citation. A page can be discovered and then ignored. It can be read and then mistrusted. It can supply a fact while an aggregator supplies the link. It can rank in Italian while the English answer leans on another source. Ranking enters the answer like a passport stamp, not like a reserved seat.
For the tourism business, the pages that ranked best were city landing pages written for search. They had opening paragraphs, branch addresses, service descriptions and seasonal language. The assistant, however, often preferred aggregator pages because they combined category, location, review snippets and comparison context in a single source. The aggregator did not know the business better. It had packaged the evidence in a way the answer could reuse.
The three citation states I count
When a team first watches assistant answers, everything looks like a yes-or-no question. Are we cited? Are we not cited? That is too crude. I usually classify each run into three citation states.
The first is silent ranking. The page ranks in search, but the business is not named in the assistant answer. This is the most common shock. The second is borrowed citation. The business is named, but the answer leans on a directory, aggregator, article or review page rather than the business’s own page. This can be useful, although it leaves someone else controlling the wording. The third is owned citation. The business is named and the cited or clearly used source is its own page, with a description close enough to reality.
These states are not moral grades. Borrowed citation can be better than absence. Owned citation can still be wrong if the page is confusing. Silent ranking can be acceptable for a query that is not commercially important. The value of the classification is that it prevents a team from celebrating the wrong thing.
In the composite tourism case, several English queries produced borrowed citations. The business appeared only through aggregator language, and the branch details were not reliable. That told us the market had enough evidence to know the business existed, but not enough clean business-owned evidence to describe it safely. The next work was not “get more rankings.” It was to make the branch pages better evidence sources.
What to inspect after the answer appears
When an assistant cites a source, open it like a patient person, not like a hunter looking for one magic line. Ask what the page makes easy. Does it state the business category? Does it separate Florence from Venice? Does it connect each branch to the correct service boundary? Does it contain proof that can survive compression? Does it use the same business name as the site, directory profiles and review sources?
Then compare that source with your own page. Many teams find something humiliatingly simple. The aggregator has the clearer category. The review site has the clearer location. A partner page explains the service boundary better than the service page. The assistant has not betrayed the business; it has followed the tidier crumb trail.
This is also why a citation check should be done in both Italian and English when the buyer market uses both. English queries in tourism-adjacent sectors are not decorative. They can bring different intent and different sources. An Italian page may rank while an English assistant answer listens to an aggregator because the English source is the only one that states what the buyer asked.
Once you see citation as name, source and description, the work becomes less mystical. You stop asking only where you rank. You ask which sentence carries you into the answer, who wrote it, and whether you can bear to have buyers repeat it.
The Citation Ledger
Query shelf: “cosa sono citazioni IA.” Ranking residue: a blue link shows that the page can be found for the old query. Citation hinge: the answer triad — name, source and description — decides whether visibility survives inside generated prose. Next count: record whether each run is silent ranking, borrowed citation or owned citation, then copy the exact description used.