Getting cited by ChatGPT: the measurement method

Being well ranked and being cited are two different things

In short: A search engine returns a list of pages.

A search engine returns a list of pages. An answer assistant reads those pages, summarises a few of them and writes an answer. In the first case, your job is to appear in the list. In the second, it is to be named in the sentence, with a source the reader can open.

That shift changes the question to ask. “Am I in the top ten results?” says nothing about “am I in the answer?”. A site can be well ranked and absent from the answers, just as a modest site can be cited on a precise question because one of its pages answers exactly what is asked.

For an SME, the stake is practical. Some of your buyers now ask an assistant to compare suppliers, to summarise a trade, to find someone in a city or a region. If your company never appears in those answers, it leaves the conversation before the visitor even reaches your site. And that absence does not show in your traffic statistics: you cannot count the visitors who never came.

Why a single test proves nothing

In short: An assistant answer depends on a number of parameters that escape observation.

An assistant answer depends on a number of parameters that escape observation. The model and its version, the exact wording of the question, the language, the session, the conversation memory, the date, and the fact that answer engines rely on an index that moves.

In practice: two people on the same team ask the same question at the same time and receive two different answers. One sees the company cited, the other does not. Neither lied, and neither measured anything. That is an anecdote point, not a measurement point.

The consequence is simple: a screenshot cannot serve as a diagnosis, even less as proof of progress. What matters is asking the same list of questions, in the same way, at a regular interval, and comparing the answers with each other. One rule only: you do not change the question because the answer does not please you.

Writing the question panel, once, in black and white

In short: Everything starts with a written list. It must resemble the questions your customers really ask, not your slogans. A good way to build it is to take the questions received in meetings, by email or by phone, and rewrite them as a hurried buyer would ask them.

Everything starts with a written list. It must resemble the questions your customers really ask, not your slogans. A good way to build it is to take the questions received in meetings, by email or by phone, and rewrite them as a hurried buyer would ask them.

A useful panel covers four families:

First, the trade with no city named: “who should a small company trust with its bookkeeping”. Second, the trade with a territory: “which companies automate invoices in Quebec”. Third, the comparison: “web agency or specialist agency, how do you choose”. Fourth, the problem to solve: “how to cut the time spent answering the same customer emails”.

Write about twenty questions, in French and in English if you sell in both languages, then freeze the list in a file. That file becomes your yardstick: it is the one you will ask again every month, identically, even and especially when the result is disappointing.

Two method precautions. First, keep a set of neutral questions, without your company name: asking “what is my company worth” only measures the assistant's politeness. Then, note the creation date of the panel and change it only by adding questions, never by rewriting the old ones, otherwise comparisons from one month to the next no longer mean anything.

Ask, log, date

In short: The repetitive part of the measurement is also the most valuable: asking the whole panel, at a fixed interval, and logging everything.

The repetitive part of the measurement is also the most valuable: asking the whole panel, at a fixed interval, and logging everything. A monthly rhythm is enough to see a trend, provided nothing is skipped.

For each question, four elements are noted: the date, the assistant used and its version, the answer obtained copied word for word, and the sources the assistant cites. That last column is often the most instructive: the cited sources form the list of those who occupy the place you are aiming for, and the gap between their content and yours then becomes obvious.

The reading is done on three levels of presence, from the weakest to the most useful:

First, absent: your company appears nowhere in the answer. Second, mentioned: it is named, with no source to open. Third, cited with a source: the answer points to a page the reader can consult.

The third level is worth far more than the second. A name cited without a source brings no visit, no trust and no proof. A cited page brings in a real reader.

Reading the results without telling yourself stories

In short: The first reflex to forget is the visit counter.

The first reflex to forget is the visit counter. There is nothing to count in an assistant answer, and that is exactly why the measurement has to be done by hand, or tooled in a disciplined way. What you compare is the number of questions in the panel where your company appears, from one reading to the next.

Then you have to separate three causes of absence, because they are not fixed the same way. The first is the absence of content: none of your pages really answers the question asked. The second is the absence of quotable content: your pages answer, but drowned in a commercial pitch, with no short answer, no dated marker, no source to cite. The third is the absence of an entity signal: your company is described differently everywhere, and nobody clearly links it to its trade and its territory.

One last trap, a frequent one: asking the question with the brand name and rejoicing at the answer. A mention obtained because you fed the name measures nothing at all. The questions that matter are the ones where you are not named.

What to fix first, and what a digital employee can carry

In short: Once the diagnosis is set, the fixes are few and often simple.

Once the diagnosis is set, the fixes are few and often simple.

First, write one page per important question. A page that starts with the answer, in a few lines, before explaining it. An assistant more readily cites a text that answers directly and states things clearly.

Then, make the content quotable: dated markers, named sources, precise definitions, frequently asked questions handled separately. Content that makes people want to reuse it gets reused.

Finally, align the entity signals: the same name everywhere, the same description of the trade, a territory written in black and white, and consistent public mentions. That is what allows an assistant to attach you to a question and to a geographic area.

What remains is the most important: starting again. Asking the panel again the following month, comparing, keeping what rises, fixing what stalls. That work is regular, repetitive and demanding in record keeping, exactly the profile of a task to entrust to a digital employee. In L'Escouade IA, that is the role of the first recruit: holding the panel, logging the answers, preparing the comparison and reporting the list of pages to fix. The owner keeps the decision: which questions matter, which pages get written, and what is accepted for publication.

Conclusion

In short: Being cited by ChatGPT is neither a lottery nor a slogan: it is the result of content that answers, coherent entity signals and a repeated measurement.

Being cited by ChatGPT is neither a lottery nor a slogan: it is the result of content that answers, coherent entity signals and a repeated measurement. Start small: a list of twenty questions written today, asked on the same day every month, logged in a file. After two readings, you will know whether your company exists in those answers, and you will be able to prove it. Before that, you will know nothing.

The online diagnostic on digital-vpartners.com shows where to start for your situation, and the squad is then built one mission at a time.