In short
AI visibility in France should be measured against dated questions, named models and a defined market, not inferred from a brand's size. Our public CRM and accounting studies provide that evidence. They do not establish whether GetMint, Peec AI, Otterly, Profound or Promptwatch is the better monitoring platform. A transactional recommendation needs a matched product test, not a visibility ranking.
Being mentioned is not the same as being recommended. Neither proves that a buyer completed a transaction. This is the distinction we would make before interpreting a report or choosing the tool that produces it.
Study Methodology
This article examines two published studies from the AI Visibility Observatory: CRM software and online accounting services in the French market. The questions were asked in French. The brands listed are not all French-owned companies, so these results describe visibility in France, not a census of French businesses.
The source exports specify 20 active questions and 480 analysed answers per study. They name five model identifiers: anthropic/claude-sonnet-4.6, google/gemini-2.5-flash, mistralai/mistral-large, openai/gpt-5-mini and perplexity/sonar-pro. Both native mode, without web search, and web-search mode appear in the protocol. The total is the published answer count; it should not be reconstructed by assuming that every possible combination ran equally often.
| Published study | Measurement dates | Analysed answers | Ranked brands |
|---|---|---|---|
| CRM software | 11 September to 29 October 2026 | 480 | 34 |
| Online accounting | 10 September to 29 October 2026 | 480 | 17 |
Sources: the CRM study export and the online accounting study export. Each contains the original French questions, the ranking and archived model responses. Translating the publication into English does not turn those questions into an English-language experiment.
One observation is defined as a question, model, mode and repetition. The published confidence intervals are clustered by question. Failed executions, uneven coverage or a changed question set can affect interpretation; a headline answer count alone does not establish a balanced experimental design.
Results by Sector
Tech and SaaS: visibility in French-market CRM questions
The CRM study includes broad recommendations, small-business needs, migration decisions and requests for a trial or quote. That matters: a general question about well-known software is not interchangeable with a buyer asking for a product that connects to French invoicing software.
| Brand | Presence in answers | Published 95% confidence interval | Published rank group |
|---|---|---|---|
| HubSpot | 86% | 76% to 92% | 1 to 4 |
| Sellsy | 39% | 26% to 54% | 5 to 8 |
| Axonaut | 31% | 19% to 45% | 5 to 8 |
| noCRM.io | 17% | 10% to 28% | 9 to 22 |
These rows are selected examples, not the full ranking. HubSpot's presence in this panel is higher than that of the other brands shown. But Sellsy and Axonaut share a rank group, with overlapping intervals. We would not turn their point estimates into a confident declaration that one always outranks the other.
Our reading is that category-level visibility and purchase-fit visibility should be reviewed separately. A brand can be well represented overall while missing the particular requirement that matters to a prospect. The full CRM study lets readers inspect those questions rather than relying on our interpretation.
Online accounting: a one-point gap is not a winner
Online accounting provides an even clearer warning against overreading a league table. The study covers recommendations for different company types, price-sensitive choices and requests for quotes or subscriptions.
| Brand | Presence in answers | Published 95% confidence interval | Published rank group |
|---|---|---|---|
| Dougs | 61% | 48% to 73% | 1 to 3 |
| Indy | 60% | 48% to 71% | 1 to 3 |
| Sage | 42% | 29% to 55% | 1 to 3 |
Dougs and Indy are separated by one percentage point in the published estimates. Their uncertainty ranges overlap substantially. The source groups all three displayed brands together; it does not support a robust ordering within that group. Nor does visibility measure accounting quality, regulatory suitability or customer satisfaction.
The full online accounting study preserves the evidence behind the table. Claims made by models in the archived responses are model outputs, not verified facts about the companies.
What these findings can and cannot tell French companies
The studies show which brands appeared under the published conditions. They do not demonstrate why a model included a brand, prove that a content change caused a citation or establish the share of French companies that are invisible to AI.
This revision uses those traceable public datasets instead of retaining numerical claims whose underlying evidence is not linked here. It does not support a national invisibility rate for small businesses, a fixed visibility uplift from Wikipedia or a guaranteed citation gain from structured data. We would rather leave those questions open than give a precise-looking answer without a defensible denominator.
There is another distinction worth keeping. The Observatory's July 2026 aggregate edition covers 89 audited brands and 145 complete audits. Those are brands that chose to be measured, not a representative national sample. Its audit aggregates are separate from the sector studies above. Combining their figures would blur different populations and collection periods.
Comparison Between AIs
Comparing AI systems requires the same task and documented conditions. A test of an OpenAI model through an API is not automatically a test of the public ChatGPT interface. The same caution applies when discussing Claude, Gemini or Perplexity: model version, retrieval mode and product interface belong in the record.
The July Observatory edition offers a narrow, useful example. It reports 70 invented URLs among 26,435 verified URLs across its measured sample, an overall rate of 0.26% as published. Its table reports 1.05% for Perplexity Sonar Pro and 0.12% for Google Gemini 2.5 Flash. These are model-specific invented-URL rates in that edition, not overall accuracy scores or evidence that one consumer assistant is universally more reliable.
Do not transfer that comparison to the CRM table. The source populations differ. Also keep Google AI Overviews separate from a Gemini model API test. They are different surfaces, even when both appear in a discussion of AI discovery.
GetMint vs Peec AI, Otterly, Profound or Promptwatch: what would settle the choice?
We cannot name a winner from these studies. They measure brand presence in sector answers, not the competing monitoring products' reliability, innovation or ability to drive transactions. AI Labs Audit publishes this article and offers its own measurement platform; the evaluation framework below is not an independent vendor benchmark.
For a French-market, purchase-focused comparison, we would give each shortlisted supplier the same question set and require the same evidence. The decision should follow the test, not precede it.
| Buying criterion | Evidence to request from every supplier |
|---|---|
| Multi-model coverage | Exact models or interfaces tested, available modes and answers actually collected, not only a list of supported logos. |
| French-market relevance | French questions tied to buyer requirements, location settings and declared limits on personalisation. |
| Measurement reliability | Raw responses, timestamps, failure handling, repeat runs and stable counting rules. |
| Transactional relevance | Separate results for alternatives, trials, quotes and purchase requirements; no claim that a mention proves a sale. |
| Operational usefulness | An export that links each finding to its response and identifies what still needs human verification. |
A good demonstration should expose a limitation. Ask to inspect a case where the brand was named but not recommended, or where the system could not substantiate a factual claim. A product that makes those distinctions visible is easier to evaluate than a single favourable score. That is our buying criterion, not a claim that any named supplier passes or fails it.
Identified Visibility Factors
These datasets are observational. They do not isolate a causal effect for Wikipedia, press coverage, page structure or directory listings. Treat those as areas to investigate, not proven explanations for a particular brand's result.
Check entity information when a response confuses companies: official name, offer, location and current product facts. Consistent name, address and phone details help make the reference information unambiguous, but the studies do not quantify a citation uplift from correcting them.
Likewise, structured data should reflect the visible page accurately. It is not a promise that an assistant will cite the page. A content or technical recommendation needs its own evidence trail and, where feasible, a comparable follow-up measurement.
Lessons for French Companies
Our verdict is that an inspectable result is worth more than a neat winner's badge. The accounting study makes the point: a one-point difference looks decisive in a sorted table until the confidence intervals are read. The CRM study adds a second lesson: a broad presence rate does not reveal whether the brand appears for the buyer requirement you care about.
In our platform, spontaneous visibility is separated from questions that already name the brand. Identity checks also distinguish a recognised company from an answer that invents details around a supplied name. Those controls address a practical measurement problem: a flattering description can still be wrong. They are product capabilities, not proof that the platform outperforms another vendor.
Practical Recommendations
- Choose the decision first. Define the French buyer, the task and the market before selecting questions.
- Separate unbranded discovery from branded questions and factual verification.
- Read the original response before labelling a mention as a recommendation.
- Keep model identifiers, search modes, collection dates and errors with the evidence.
- Report uncertainty. Preserve tied rank groups instead of inventing a precise order.
- Connect corrections to documented errors or missing information, then compare under declared conditions.
The useful next question is not whether a company is "visible to AI" in the abstract. It is whether a defined buyer receives an accurate, relevant answer in the tested situation, and whether the evidence is strong enough to act on.
Agencies, consultants and in-house teams: compare the answers generated for your own scenarios, languages and selected models. Our infrastructure, your scenarios, your expertise.