Skip to content

Blog

Three hundred questions a day, or eight questions for thirty days?

By Johnny Lagneau · Updated October 21, 2026Originally published October 21, 2026

Translated from French. Read the original

What AI visibility measurements are worth, according to four recent studies and my own numbers.

On 13 and 14 October, I spoke with three people who work on AI visibility every day: an SEO and GEO specialist who has just left a large bank, an independent GEO consultant, and a GEO consulting agency in Brussels.

Three conversations, three profiles, and I heard the same question three times, in different words: how many questions should you track, and how often, for the measurement to be of any use?

I run VectorGap. I measure what the AIs answer about a brand, city by city and language by language. So I have a stake in this question, and I would rather say so right away.

I spent the week reading the studies and redoing my calculations. Here is what I found. The sources are at the bottom, so you can check every number.

Two schools, and a disagreement about the number of questions

The first school sells volume: dozens, often hundreds of questions per brand, asked every day to four, eight or thirteen AI engines. The entry offers I noted start at 25 or 50 questions. "Enterprise" offers go up to several hundred.

The second school, more recent, says 5 to 10 questions per site, no more. On 15 October, a consultant published on X the process of a French SEO agency with more than 400 clients: five to ten questions per site, and beyond ten, the work becomes unmanageable. His idea: behind each question, the AI runs about ten hidden searches, and those searches are what you have to work on.

I think the two schools are arguing about the wrong number. The number of questions matters less than the precision of each measurement.

What an AI answer is worth, measured

When I ask ChatGPT the same question twice, I do not get the same answer twice. Everyone has noticed that. Researchers have measured it.

SparkToro and Gumshoe (study published in January 2026: 2,961 prompts, 12 brand recommendation questions, 60 to 100 runs per question and per engine): ChatGPT and Google AI return the same list of brands less than once in 100. The same list in the same order, less than once in 1,000.

EWR Digital (12,778 answers across five engines, from January to June 2026) looked at the sources cited: the same question, to the same engine, returns the same set of sources in less than 2.5% of cases. For ChatGPT, 0.06%.

The same study gives the margin of error of a presence measurement: with a single run, about ± 33 points. With seven or eight runs, about ± 10.

And the University of St. Gallen (April 2026) did the calculation over time: with twenty-eight days of daily measurement of the same questions, the margin drops to about ± 6.5 points.

I take one thing from these three studies: a single capture is worth nothing. "I asked ChatGPT and we weren't there" is not a measurement, it is a draw.

And precision comes from repeating the same questions over time. The big tools know this, since they ask their questions again every day.

So where is the difference between three hundred questions and eight? In what you can read. Three hundred questions tracked every day give three hundred precise curves, and the client will only ever read the score that sums them up. Eight questions tracked every day give eight curves I can read one by one, with the sources and the competitors behind each one.

I do not claim to measure more finely than a tool with three hundred questions. I claim to measure what a business owner can read and act on, in their city and their language, and to tell them the limit of the measurement.

The calculation nobody does in front of the client

Take a common offer: 150 questions, 8 engines, every day. That makes 36,000 answers a month for a single brand. At nine sources cited per answer on average (that is what the studies measure), that makes about 324,000 citations a month.

What do you do with 36,000 rows? Nobody reads them. The tool aggregates them into a score, and the score moves every day, because the AI itself moves every day.

Sistrix measured it (82,619 prompts, six countries including France, seventeen weeks): from one week to the next, 74% of the sources cited by ChatGPT are new. On Google AI Mode, 56%. On Google's AI Overviews, only 5%: there, it is almost frozen.

So a daily measurement of ChatGPT mostly sees the randomness of the draw, and a daily measurement of an AI Overview mostly sees the same thing as the day before. In both cases, the day is the wrong unit. The right unit is the window: two to four weeks of the same questions, and I look at what holds.

(I am adding this because it is the question I get every time: no, answers are not "locked for three days". Ahrefs checked on 43,000 keywords: no link between how popular a question is and how stable its answer is. What is stable is the meaning of the answer, not its surface.)

Why agencies measure wide, and why I understand them

I am not putting marketing agencies on trial. I meet many of them, and what I see is caution: an agency measures wide because it is afraid of missing a piece of data its client might have seen. Belt and braces.

The problem is that the AI will have changed that piece of data by the next day. St. Gallen measures nearly two thirds of different sources from one day to the next on the same prompt. The agency cannot miss what no longer exists the next day.

The real risk for the agency is elsewhere. Showing the client a "drop" on a Tuesday that is only a draw, acting on it, seeing nothing move, and losing the client. What protects the agency is knowing the noise of its measurement and only talking about the movements that exceed it.

On the client side, studies on decision-making have said the same thing for a long time (a review of 57 studies, published in 2023): beyond a certain volume, more information degrades the precision and consistency of decisions, and lengthens the time it takes to make them. A business owner acts on three things a month. They need those three.

So how many questions?

No study gives a magic number. The data still let us reason.

Behind a question, ChatGPT runs 2.1 searches on average (Peec, about 5 million searches analyzed in April 2026). You often read ten; it is two. Ten questions about the same profession in the same city therefore share the same searches and the same sources. Measuring them separately adds volume, not information.

The useful number of questions is the number of truly different intentions a buyer can have. For a local business, you can count them on the fingers of one or two hands: who is the best near me, how much does it cost, is this option better than that one.

The St. Gallen researchers work with eight questions per campaign. The agency with 400 clients caps it at ten. I ask eight questions, written for the brand (its profession, its city, its language), and I freeze them so I can ask them again.

The sources concentrate too. Across 1.2 million ChatGPT answers, thirty domains capture 67% of the citations on a topic. The 324,000 citations in the calculation above come down to about thirty names, and I find them with eight questions.

What I see on the same questions, without spending anything more

This is the point the second school forgets. When I ask my eight questions, the AI does not only answer about the brand. It names other brands and other sites. In the same answers, I see who it recommends instead, who it cites as a source, and where nobody from the sector is present.

On 9 October, I asked 8 buyer questions, in French and Dutch, to 13 AI configurations, for 46 Belgian online shops. ChatGPT never names 41% of the shops (19 of 46). 17% (8 of 46) are named everywhere. Instead of the missing ones, it sends people to marketplaces and directories.

That is where the places are to be taken. I did not see it by adding questions; I saw it by reading the answers to the eight.

Then comes the loop: measure, analyze, conclude, act, then run the measurement again on the same questions. The remeasurement only makes sense on the same questions, and it only makes sense if I know the noise.

The dimension almost nobody measures

A buyer in Brussels who searches in Dutch does not get the answer a buyer in Paris gets in French. It seems obvious, and yet almost every measurement I see is "world" or "country", in one language, with no city.

From 24 to 30 September 2026, I measured 40 audits and 2,127 answers. In Brussels, in Dutch, 3.5% of the sources cited come from the Netherlands. In Luxembourg, for an insurance broker in French, 24.8% come from France. In Canada, in English, 14.6% come from the United States.

And for the same Brussels shop, with web search enabled, 50% of answers name it in French (27 of 54) against 25% in Dutch (14 of 56), on the same questions.

The St. Gallen researchers say it themselves in the limits of their paper: their data come from Swiss servers, and their results may not generalize to other languages or other markets. The measurement depends on the place and the language. A "country" average does not exist for any buyer.

(I also think local has an advantage: narrow intentions and fewer sources, so less noise. I say "I think" because I have not measured it yet. It is in progress, see below.)

What I don't know yet

I would rather write down what I am missing.

First, how much an AI answer to a local question moves from one day to the next, when nobody touches anything. The studies above measured it on general questions, in English or German, not by city, not in French or Dutch. I am preparing a thirty-day test: ten Belgian brands, two languages, eight questions, every day, with no action.

I will publish the result, whatever it is. If it shows that the day-to-day level moves a lot, the frequency argument will have to be taken seriously. If it does not move, a weekly measurement will be enough, and will cost seven times less.

Next, the hidden searches the AI runs behind each question. I know how to count them. I do not read their text yet. That is the next step.

And a limit of my own measurement, to be consistent with what I criticize in others: with about a hundred answers per language, a snapshot sees a change of 15 points, not 5. To see 5 points, you need the series over time. That is why I run it.

What I take away

Volume is not the problem, and it is not the answer either. Three hundred questions tracked every day measure a trend; eight questions tracked every day do too. The difference is that the eight can be read, and that in the same answers I see the competitors, the sources and the places to take, in the city and the language where customers buy.

I may be wrong. This is an opinion built on my measurements and on four studies. If you measure differently and it works, I am interested.

You who track brands in the AIs: how many questions do you really track, and how many do you read?

Sources. SparkToro and Gumshoe, 2,961 prompts (via Search Engine Journal, January 2026). EWR Digital, "AI answer variance", 12,778 answers, January to June 2026. Schulte, Bleeker, Kaufmann (University of St. Gallen), "Don't Measure Once: Measuring Visibility in AI Search (GEO)", arXiv 2604.07585, April 2026. Sistrix, 82,619 prompts, December 2025 to April 2026. Ahrefs, 43,000 keywords, AI Overviews.

Peec, about 5 million searches, April 2026. Kevin Indig, 1.2 million ChatGPT answers (via Search Engine Land). Review of 57 studies on information overload and decision-making, Management Review Quarterly, 2023. VectorGap measurements: 24 to 30 September 2026 (40 audits, 2,127 answers); 9 October 2026 (46 Belgian e-commerce sites, FR and NL).