AI Visibility : Three Questions to Ask Before Believing a Quantified Result

AI Visibility : Three Questions to Ask Before Believing a Quantified Result
3 min read

AI visibility providers announce percentage increases, but a business leader rarely has the means to verify them. Three questions, however, allow for sorting things out: were the measured questions defined before the work began, were control questions tracked, and is the calculation the same before and after?

The same provider displays the same client case at +24.7% or +30.4%, depending on the chosen formula. It also promises +30% traffic in 90 days by adding organic traffic, i.e., from search results, and traffic coming from AIs. Taken in isolation, none of these figures says what the work actually produced in the responses of ChatGPT, Perplexity, or Google.

The difficulty lies in the very nature of the subject: AI responses change without anyone touching them. In September 2026, Orbit Media asked the same 72 B2B buyer questions twice, two days apart. Between the two surveys, Perplexity kept only 65% of the pages it cited. An increase can therefore reflect only this movement.

Were the measured questions defined before the work?

A result is only valid if the list of questions was written and dated before the first survey. A list chosen after the fact indeed retains only those that have improved.

Out of a hundred questions a buyer might ask, a few always go up, by chance or thanks to the work. One only needs to show those. Therefore, one must ask for the complete list, its date, and how it was compiled. In our projects, the questions start from what the client’s buyers ask, based on sales calls, support, and Search Console, and are frozen in the document signed before the baseline measurement.

Were control questions tracked?

Control questions are measured but never worked on. They separate the effect of the work from the inherent movements of the AIs: if they rise as much as the others, the work had nothing to do with it.

A search engine changes its model, a competitor publishes a comparison, a sector directory is updated, and the responses evolve. Without a control, this movement passes for a result. In our projects, each corpus, i.e., the set of tracked questions, includes controls chosen before the baseline measurement and never worked on until the final review.

Is the calculation the same before and after?

An increase can only be read between two surveys conducted in the same way: same questions, same search engines, same number of queries, and same unit of account.

Adding traffic illustrates this: a good season in Google’s results is enough to make it rise, without one more citation in ChatGPT. The number of queries counts just as much. On September 24, 2026, on behalf of a French contact center software publisher, we asked ChatGPT 10 buyer questions, 5 times each, on the same day. The brand appeared in 1 out of 50 responses: a single survey could have concluded either its presence or its absence.

Conversely, a comparison made under the same conditions is read without debate. For this same client, the share of voice in Perplexity, i.e., the place occupied by the brand in the measured responses, went from 1.7% to 5.0%, on the same corpus of 20 questions, asked 3 times per survey. The figure remains modest, but it is reproducible.

Before signing, three elements must be requested in writing: the dated list of questions, the controls, and the calculation formula. A provider that measures seriously already has them on hand.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x