2026 Barometer: Sector Leaders' Presence in AI Answers

In short

Across nine French sectors, the observed presence of the leading brand ranges from 37% for organic-search agencies to 86% for CRM software. Each study analyses 480 answers. This barometer compares sector leaders, not industries' total citation volumes or the concentration of mentions between brands. Confidence intervals overlap substantially, so the table is descriptive rather than an established statistical ranking. The translated editions report the same French evidence.

Leading-brand presence is not an industry citation ranking

Which industry appears most often in AI answers? Our data cannot answer that question as a measure of overall volume. They answer a narrower one: how often does the first-ranked brand in each study appear in the answers collected for that study? The denominator is the set of answers in the relevant scan, not every conversation people have with assistants. An 86% presence rate therefore describes a bounded corpus. It does not measure an industry's share of AI usage or the number of people interested in that market.

There is another distinction worth keeping. A leader's frequency is not the concentration of mentions across competing brands. Several names can occur in one answer. Without examining their full distribution, we cannot tell whether a few brands account for most mentions or whether many competitors share them. We have chosen to compare an available, documented measure without asking it to support a different claim. That is less dramatic than a league table of industries. It is also a more defensible way to use these studies.

Nine French studies with a shared measurement framework

The nine studies use data collected from 9 to 11 September 2026. Publication and collection dates must remain separate: a new presentation does not turn September answers into a November measurement. Each study uses 20 French-language questions and 480 analysed answers from five models. The question sets cover areas including recommendations, comparisons, information and transactional intent, but the actual wording is specific to each market. These are not identical questions with different industry labels pasted into them.

The protocol has eight model-and-mode combinations, with three repetitions per question. Claude Sonnet 4.6, Mistral Large and GPT-5 mini are queried both without web search and with web search. Gemini 2.5 Flash is queried without web search; Perplexity Sonar Pro uses web search. Describing this as five models in two modes would incorrectly suggest ten combinations. The common framework supports a descriptive comparison, not the removal of every difference between sectors. Question selection still defines the scope of what was observed in each study.

A descriptive table with uncertainty beside each estimate

Presence rates below are rounded to whole percentages. The 95% confidence limits reproduce the published snapshot values to two decimal places. Descending order is a reading aid. Each sector link leads to its full study, with the protocol, published brands and answers needed to inspect the result rather than relying on a headline number alone.

Sector	Leading brand	Rounded presence	95% CI	Ranked brands
CRM software	HubSpot	86%	75.58% - 92.47%	34
Cabin luggage	Samsonite	70%	56.57% - 81.31%	26
Dog health insurance	SantéVet	63%	49.01% - 74.97%	22
Online accounting	Dougs	61%	47.63% - 73.31%	17
Mattresses	Emma	56%	41.27% - 69.82%	18
Invoicing software	Sage	54%	42.13% - 65.74%	28
Car insurance	Direct Assurance	50%	36.73% - 64.04%	22
Electric bikes	Decathlon	49%	35.90% - 62.55%	37
Organic-search agencies	Eskimoz	37%	23.81% - 53.09%	21

Several studies group brands into tied ranks. Leading brand here means the first published row, not a statistically unique winner.

The intervals use the Wilson method with adjustment for clustering by question, recorded as wilson_clustered in the data. Repeated answers to one question are not treated as entirely independent observations. Having 480 answers therefore does not guarantee a tiny error margin. Keeping the limits next to the estimate is deliberate: a reader should not have to reach the final disclaimer to discover that a seemingly precise percentage carries substantial uncertainty.

CRM, luggage and dog insurance: what the observed values show

HubSpot is the leading brand in the CRM study, with 86% presence and an interval from 75.58% to 92.47%. This is the highest point estimate in the table. Samsonite reaches 70% in cabin luggage, with an interval from 56.57% to 81.31%. Those two intervals overlap. We therefore do not describe the displayed order as statistically demonstrated superiority between industries, much less as evidence that assistants generally prefer one kind of business. These are measurements of particular brands responding to particular question sets.

For dog health insurance, SantéVet records 63%, with an interval from 49.01% to 74.97%. Online accounting records 61%, led by Dougs, with limits from 47.63% to 73.31%. Different companies and different questions sit behind those figures. They can help a team find the study relevant to its own market. They cannot establish that an insurer would be more likely to appear than an accounting firm in some universal comparison. There is no shared universal question set behind such a claim.

Mattresses and invoicing: two points do not create two groups

Emma appears in 56% of answers in the mattress study; Sage appears in 54% in the invoicing study. Their respective intervals, 41.27% to 69.82% and 42.13% to 65.74%, overlap substantially. Calling one market fragmented and the other concentrated would create a distinction that this indicator cannot support. We have removed that interpretation. The presence rate of the first-ranked brand is not enough to establish either market structure, regardless of how convenient those labels might be for a narrative.

The same caution applies to car insurance at 50% and electric bikes at 49%. Their limits are displayed in the table. A difference between point estimates is not, by itself, proof of a lasting difference. Conversely, overlapping intervals do not prove that two industries are identical. This barometer supplies no comparative test that settles that question. Its job is to show observations and their uncertainty, without inventing categories or assigning more meaning to small gaps than the analysis provides.

A leader at 37% does not make a sector invisible

The organic-search agency study ranks 21 brands. Eskimoz has 37% presence, with an interval from 23.81% to 53.09%. The snapshot also lists 50 names below the ranking's publication threshold. That list is capped, so the number should not be presented as an exhaustive inventory of every other brand that appeared. The result describes the first-listed actor's presence in the collected answers. It does not establish the absence of agencies from AI answers, or how often people ask about agencies overall.

For an agency, the useful next step is to read that study's questions and responses, then examine its own brand's results. An overall presence rate cannot reveal exactly which customer need a company is missing from. That requires looking at individual answers, the alternatives named and the facts attached to each name. Mention, recommendation and factual accuracy are different readings. None should be inferred automatically from another. A brand can be present while being described incorrectly; the presence percentage alone does not resolve that issue.

Why the toys study stays separate

The children's toys study uses data collected on 29 October 2026. Lego reaches 63.75%, rounded to 64%, with an interval from 46.50% to 78.06%. It analyses 80 answers to 20 questions, from four models, exclusively in web mode. Its panel includes Gemini 3.7 Flash rather than Gemini 2.5 Flash, and does not include Mistral Large. This is not simply a smaller version of the other nine scans. Its search conditions and model selection are materially different.

Different sample sizes alone do not invalidate a comparison. Here the panel, Gemini version, search mode and collection date also differ. We keep toys outside the common table to avoid attributing to an industry something that might depend on the protocol. The study remains useful as a separate observation with its own limits. Combining it with the others would require an aligned protocol or an analysis that explicitly handles these differences, rather than sorting all the percentages into a single list.

What we will not explain without evidence

Why does one leader appear more frequently in its scan than another does elsewhere? This table establishes no causal mechanism. It does not measure the effect of documentation quality, accessible evidence, brand awareness or what a model may have memorised. Those ideas can become research hypotheses. They are not findings from this comparison. We therefore do not advise copying a page format with a promise of reproducing the CRM rate. The observations do not provide the intervention evidence needed for that promise.

Our verdict is straightforward: this barometer is an entry point into nine studies, not a map of AI's favourite industries. Choose the relevant market, check the panel and collection date, then read the answers behind the percentage. Translating the article does not change the population studied: all seven editions describe the same French data, not seven national surveys. Any later measurement must document its own scope before claiming improvement. That leaves readers with a usable benchmark and a clear boundary around what it can actually tell them.