A citation query is useful only when it forces a choice: name one business, trust one source, describe one fit. Old keyword lists often avoid that choice.
On my desk there is usually one ugly spreadsheet before the clean ledger appears. It has hundreds of rows copied from a ranking tool: “software gestione compliance,” “consulente compliance logistica,” “workflow audit fornitori,” “miglior software sicurezza documentale,” then the same ideas in English, bent into procurement language or half-translated by someone in sales. In a composite scenario drawn from Italian B2B software projects, a 28-person company in Lombardy had more than 600 tracked terms and still could not answer a simple question: when a buyer asks an assistant for a named tool, do we appear?
The answer was irritating. For several broad product terms, the company ranked. For a few, it ranked quite well. Yet assistant answers often named generic SaaS listicles, a larger competitor, or a directory page that described “compliance platforms for logistics” in cleaner language than the company’s own site. One answer even called the company a consultancy. The rankings were not false. They were simply answering a thinner question than the one the buyer had started to ask.
Old keyword lists are too flat for assistant answers
An SEO keyword list is often built around pages. One page targets “software compliance aziendale.” Another targets “gestione documentale logistica.” A blog post catches “normativa fornitori.” That structure still has use. It tells you where search demand sits and where pages might compete. I still ask for it when I start a baseline.
The mistake is copying that list directly into citation tracking.
Assistant queries behave less like page labels and more like little buying scenes. A buyer does not only ask for a term. They ask for a category with constraints, a place, a level of trust, sometimes a comparison. “Which Italian compliance workflow software is suitable for logistics companies?” is a different test from “software compliance.” The first query gives the assistant a job: pick names, explain fit, maybe cite evidence. The second can be satisfied by definitions and broad educational pages.
That difference matters because a citation ledger should not count every possible phrase. It should count moments where assistant selection can be observed. If the answer does not need to name anyone, the business cannot learn much from being absent. If the query is so vague that the assistant gives a general lecture, the result may tell you more about the assistant’s caution than about your evidence.
A citation query is a repeatable buyer question, because it asks an assistant to select, cite or describe a business for a specific need. That is my working definition. It sounds plain, and it should. If a query cannot be repeated, cannot trigger selection, or cannot expose description accuracy, it belongs in a different notebook.
The first shelf: buyer questions that force a name
I call the first group the “name-forcing shelf.” The phrase is inelegant, which is partly why I keep it. It reminds me that the query has to do work.
For the Lombardy software company, “software compliance” was too loose. It brought back definitions, broad vendor categories and sometimes no names. A better shelf included questions like: “Which Italian software providers help logistics companies manage compliance workflows?” and “Best compliance workflow tool for manufacturing suppliers in northern Italy.” These are not elegant keywords. They are buyer questions with a hinge in them. The assistant has to decide what kind of company fits, which geography matters, and whether the page or a third-party source states enough proof.
The same logic works for a small local service business, though the wording changes. A tourism-adjacent firm in Florence and Venice may rank for “tour assistance Florence” or “servizi per visitatori Venezia.” Useful citation checks would be closer to “Which company helps English-speaking visitors handle private logistics in Florence?” or “Best Venice-based service for visitors who need Italian partner coordination.” The query has to make the assistant choose a business rather than explain a destination.
There is a rough edge here. Real buyers do not always phrase questions cleanly. They ask with missing verbs, mixed language, tourist slang, or internal procurement words. I do not try to polish all that away. I usually keep one or two awkward versions in the shelf, because assistant systems often reveal source bias when the query sounds like a human wrote it after a train delay.
A good name-forcing query has three parts sitting close together: the category, the buyer situation, and the reason a named source should be trusted. Without those parts, the query may still be useful for SEO. It is just weak for citation measurement.
The second shelf: Italian and English variants without mixing them
Italian businesses often underestimate the English shelf. They treat it as translation work, then wonder why the sources change.
For some markets, English queries pull in a different evidence layer. A visitor in London asking about a service in Florence may not use the same terms as an Italian partner. A procurement manager outside Italy may search for “Italian logistics compliance software” rather than “software per compliance logistica.” An assistant answering in English may lean on listicles, international directories, travel-style summaries or old aggregator pages that would not dominate the Italian query.
This does not mean every Italian business needs a large English ledger. It means the language choice must match plausible buyer behavior. If English-speaking buyers, investors, partners, tourists or procurement teams might ask the question in English, count that shelf separately. Do not average it into the Italian shelf as if language were only a wrapper. The evidence competitors may be different.
In the software composite, the Italian shelf surfaced sector pages, local competitors and a few business directories. The English shelf surfaced generic SaaS comparisons and pages using procurement language the company never used on its own site. The assistant sometimes understood the category better in English because third-party sources had written cleaner category sentences. That is an uncomfortable finding, but useful. It tells the team where the cited vocabulary lives.
I use a small classification here: mirror, migrant and market-native queries. Mirror queries are close translations. Migrant queries carry the same buyer need but change vocabulary as they cross language. Market-native queries exist only because that language group asks differently. “Software per gestione conformità fornitori” and “supplier compliance workflow platform in Italy” are not exact twins. They sit near each other, like cousins who disagree about what the family business does.
The ledger should keep those cousins apart.
The third shelf: disagreement queries
Some queries are valuable because they expose disagreement between ranking and citation. I do not choose them because they have the largest search volume. I choose them because they make the gap visible.
A disagreement query is one where the old SEO report and the assistant answer are likely to diverge. The page may rank, while the assistant names an aggregator. The business may appear in local search, while the assistant chooses a competitor from a guide. The company may be mentioned, while another source gets cited. These queries are often narrow enough to feel almost annoying.
For the B2B software company, one disagreement query was built around “software provider, not consultancy.” The company had a service page that used too much advisory language, while directories described the product more clearly. In assistant answers, the business sometimes appeared with the wrong role. So the ledger included prompts that separated software providers from consultants. That was not a traditional keyword priority. It was a citation-risk priority.
Another disagreement query asked for tools serving both logistics and manufacturing firms in northern Italy. The company had both sectors on the site, but not in one stable sentence. A competitor had a rougher site overall yet stated that fit in reusable wording. The assistant preferred the competitor more often than the ranking report would have predicted.
This is why I dislike building citation shelves from volume alone. Search volume is a crowd noise. Citation measurement needs a smaller sound: the click of one business being chosen over another.
Cut the list until it can be repeated
A first citation shelf should feel almost too small. Ten to twenty queries can teach more than two hundred lazy checks. The point is repetition. The same query shelf must be run again, with the same language variants, the same market assumptions, and the same columns for name, source, wording and errors. A one-time scan is a weather report through a dirty window.
I usually cut in three passes.
First I remove queries that do not require a business answer. Definitions, pure how-to topics and broad educational terms may belong in content planning, yet they do not show whether assistants name you.
Then I remove duplicates that differ only by decoration. “Best,” “top” and “leading” may create some variation, but if all three ask the same thing in the same language and source environment, one may be enough for a baseline. I would rather keep one Italian buyer version and one English market-native version than three nearly identical Italian phrases.
Finally I keep a few uncomfortable queries. These include the competitor that keeps appearing, the branch that gets confused, the category where the business is misdescribed, or the aggregator that seems too powerful. A clean ledger with no hard questions is a ceremonial object. It looks nice and teaches little.
The imperfect detail is often where the work begins. In one run for the composite software case, the assistant named the right company, cited a listicle, and described the product as “document consulting.” That result could look half-successful if the only column were “named: yes.” With better columns, it became a page evidence problem and a source gap problem.
A query shelf is a measurement instrument
People sometimes ask for “the best AI citation queries” as if I keep them in a drawer. I do not. The shelf has to fit the business, the buyer, the language and the sources likely to be trusted. Copying another firm’s shelf is like borrowing prescription glasses at night. You may see something. You will also walk badly.
A useful shelf has some stable pieces. It includes category queries where the assistant should name suppliers. It includes fit queries, where the buyer has a sector, location, language or problem constraint. It includes comparison queries where competitors may appear. It includes description-risk queries where the business is often flattened or mislabeled. For local or tourism-adjacent markets, it includes place and branch variants. For B2B markets, it includes procurement and role-language variants.
Yet the shelf must remain small enough that a person can inspect the answers. Citation tracking is not just counting appearances. It is reading which source was reused, which description survived, which competitor became the default, and which query made the assistant hesitate. That reading is slow at first. Later it becomes faster, because the same errors return wearing slightly different coats.
Choosing queries worth counting twice means accepting that some old SEO comfort will be left outside the ledger. A keyword can rank, produce impressions, and still fail to test whether the business is selected inside an answer. Keep it in the ranking notebook. In the citation notebook, ask the questions that force the market to name names.
The Citation Ledger
Query shelf: “Which Italian and English buyer questions force an assistant to name a suitable business?” Ranking residue: old keyword lists still show pages, positions and volume, even when no assistant selection is tested. Citation hinge: a useful query combines category, buyer fit, language and enough pressure to expose cited-source choice. Next count: keep a small shelf, repeat it monthly, and record names, cited sources, description accuracy and skipped competitors.