How AIs interpret a brand: what 43,287 answers from 7 AIs show
By Johnny Lagneau · Updated November 16, 2026Originally published November 16, 2026
Translated from French. Read the original
On 14 November, I asked seven AIs (ChatGPT, Claude, Gemini, Grok, Mistral, Perplexity and Muse Spark) to describe Cebi, an automotive supplier based in Steinsel, Luxembourg.
Three of them turned it into an Italian company: Claude, Grok and Muse Spark.
Without web search, 31% of the answers in English place its headquarters in Italy (13 of 42). Mistral describes it as French, then confuses it with a Turkish maker of aluminium parts.
Cebi was founded in Luxembourg in 1976 (it turns 50 this year, by the way).
Picture a buyer who is looking for a Luxembourg supplier and reads that answer. They move on to the next one, without calling you.
I wanted to know whether Cebi was an isolated case. I went back to all my measurements from September to November 2026: 911 audits and 43,287 answers.
In brief
- Without web search, 26% of the answers that describe a company contain at least one factual error. With web search, 12%.
- An AI describes a company even when it does not know it. For Cebi, only 7 answers out of 96 say they do not know it.
- Across 47 brands, the gap between French and Dutch goes from 5.9 points without web search to 10.6 points with it.
- Two cities with the same language share only 2% of the companies suggested. The ones that travel from one city to another are mostly the big groups.
- The AI you ask weighs as much as the brand: 27% of the citation gaps each. The wording of the question, 10%.
- On the same day, 17.5% of the answers change when I ask the same question again. Seven AIs in two runs bring the standard deviation of a measurement down to 2.8 points.
Three words, to avoid misunderstandings
At VectorGap, I measure three different things.
- Presence: when a buyer looks for a supplier without naming the brand, does the AI suggest it?
- Perception: when the buyer names the brand, what does the AI say about it? What it does, where it comes from, what it is worth, with the errors along the way. This is how the AI interprets the brand.
- Preference: when the buyer compares several brands, which one does the AI choose? For what reasons?
A brand can be absent, wrongly described, or set aside at the moment of choice. These are three different problems. They call for three different actions.
How I measured
A measurement covers one brand, in one language and one place, for one type of buyer. For each measurement, I ask each AI 8 buyer questions.
I question seven AIs, all built on large language models (LLMs): ChatGPT, Claude, Gemini, Grok, Mistral, Perplexity and Muse Spark (until 7 October, DeepSeek held Muse Spark's place).
I measure in two modes. Without web search, the AI answers from what it learned during training. With web search enabled, it searches the web before answering. I question Perplexity with web search only.
I go through each model's API. I write each question the way a buyer would ask it, in their language, with their need and their market. For example: Quels sont les meilleurs courtiers en assurance à Bruxelles ?
(Which are the best insurance brokers in Brussels?). The measured brand is never named in the Presence questions, to avoid rigging the answers.
The numbers come from two sources.
| All my measurements | Test on six industrial companies | |
|---|---|---|
| Dates | 1 September to 12 November 2026 | 14 November 2026 |
| Audits | 911 | 48 |
| Answers | 43,287 | 1,083 Perception answers checked |
| Brands | 130, plus 21 profession panels (one profession in one city, with no target brand) | 6 companies from Belgium and Luxembourg |
| Languages | French (379 audits), English (315), Dutch (177), German (36) | two languages per company |
For the test on the six industrial companies, I compared each Perception answer with the company's website. When the website does not allow me to check a statement, I do not count it as an error.
I give the key percentages with their 95% confidence interval.
These numbers are a sample of my measurements. The local studies I publish on Brussels, Luxembourg, Paris and Montréal point the same way.
Does the AI tell the truth about a company?
The AIs get it wrong twice as often without web search as with it.
Across the six industrial companies, 26% of the answers without web search contain at least one factual error (132 of 502; interval 23 to 30%). With web search, 12% (72 of 581; 10 to 15%). The gap is clear (z-test, p < 0.001).
On equal terms, without Perplexity, which only answers in the second mode, the rate with web search is 14%.
Six industrial companies: the answers that contain an error
6 companies, 2 languages each, 6 to 7 AIs, 7 questions that name the company. Measured on 14 November 2026.
- Cebi57.1% → 21.6%
- Company A31.7% → 10.2%
- Company B23.8% → 5.2%
- Company C20.2% → 6.2%
- Company D14.3% → 10.2%
- Company E10.7% → 21.1%
Answers with at least one factual error (%)
- Without web search
- With web search
View the numbers
| Without web search | With web search | |
|---|---|---|
| Cebi | 57.1% (48 of 84) | 21.6% (21 of 97) |
| Company A | 31.7% (26 of 82) | 10.2% (10 of 98) |
| Company B | 23.8% (20 of 84) | 5.2% (5 of 97) |
| Company C | 20.2% (17 of 84) | 6.2% (6 of 96) |
| Company D | 14.3% (12 of 84) | 10.2% (10 of 98) |
| Company E | 10.7% (9 of 84) | 21.1% (20 of 95) |
| Six industrial companies: the answers that contain an error | 26.3% (132 of 502) | 12.4% (72 of 581) |
Cebi is the most striking case. Without the web, 57% of its answers contain an error (48 of 84; 47 to 67%). With the web, 22% (21 of 97; 15 to 31%).
These errors concern the country of the headquarters, the plants, the headcount and the history of the group.
Web search reduces errors. It does not remove them. At Company E, they even go up: 11% without the web (9 of 84), 21% with it (20 of 95; p = 0.06). With the web, Claude, Gemini and Grok repeat the former name of one of its shareholders (11 answers).
Does an AI describe a brand even when it does not know it?
For Cebi, almost always.
Without web search, I received 96 answers about Cebi. Only 7 say they do not know the company (4 to 14%).
All 7 come from Claude. In 4 of them, Claude describes Cebi anyway: Cebi est, à ma connaissance, le nom d'un groupe italien
(Cebi is, to my knowledge, the name of an Italian group).
ChatGPT reacts differently. In 5 answers, it asks which Cebi I mean. It makes no errors (0 of 14).
Mistral asks no questions. 12 of its 14 answers contain an error.
So an error about a brand often comes from a lack of information. The AI fills in with what it associates with the name.
Does the language change what the AI answers?
The language of the question changes the answer, especially when the AI searches the web.
Cebi is suggested in 16% of the answers in English with the web (18 of 111, two runs; 11 to 24%), against 4% in French (4 of 109; 1 to 9%). The gap is clear (p = 0.002).
I measured 47 brands in French and in Dutch.
Without the web, the citation rate is almost the same in both languages: 20.4% and 19.9%. The correlation between the two languages is 0.94 (0.89 to 0.97).
With the web, the rate stays at 20.4% in French and falls to 14.0% in Dutch. The correlation drops to 0.77 (0.62 to 0.87).
Brand by brand, the gap between the two languages widens with the web. The average gap goes from 5.9 to 10.6 points. 23 brands out of 47 have a gap of at least 10 points with the web, against 12 without.
Each brand, in French and in Dutch
47 brands measured in both languages, from 1 September to 12 November 2026. Same order in both panels.
Without web search
average gap between the languages: 5.9 points
- Brand 162.5% → 54.8%
- Brand 212.5% → 20.8%
- Brand 318.8% → 18.8%
- Brand 437.5% → 25.0%
- Brand 58.3% → 10.0%
- Brand 60.0% → 0.0%
- Brand 760.9% → 72.5%
- Brand 88.7% → 8.5%
- Brand 90.0% → 0.0%
- Brand 1037.6% → 44.6%
- Brand 110.0% → 0.0%
- Brand 1223.9% → 39.5%
- Brand 130.0% → 0.0%
- Brand 140.0% → 0.0%
- Brand 152.1% → 0.0%
- Brand 160.0% → 0.0%
- Brand 170.0% → 0.0%
- Brand 180.0% → 0.0%
- Brand 194.5% → 0.0%
- Brand 202.2% → 2.1%
- Brand 2183.3% → 100.0%
- Brand 2238.6% → 17.8%
- Brand 230.0% → 4.0%
- Brand 2427.1% → 19.0%
- Brand 258.5% → 6.8%
- Brand 2611.6% → 17.4%
- Brand 2733.3% → 46.5%
- Brand 2814.6% → 12.5%
- Brand 2920.8% → 38.1%
- Brand 3016.3% → 0.0%
- Brand 314.3% → 0.0%
- Brand 320.0% → 11.9%
- Brand 3318.8% → 4.2%
- Brand 3438.1% → 35.4%
- Brand 3527.4% → 29.4%
- Brand 364.4% → 2.1%
- Brand 374.3% → 0.0%
- Brand 3866.7% → 63.8%
- Brand 3910.4% → 14.6%
- Brand 403.2% → 1.4%
- Brand 4125.0% → 0.0%
- Brand 4267.4% → 70.1%
- Brand 4314.6% → 12.5%
- Brand 4459.1% → 61.7%
- Brand 4576.7% → 64.9%
- Brand 463.3% → 5.3%
- Brand 472.1% → 0.0%
Answers that cite the brand (%)
With web search
average gap between the languages: 10.6 points
- Brand 19.3% → 30.8%
- Brand 27.1% → 25.0%
- Brand 328.8% → 45.1%
- Brand 439.3% → 51.8%
- Brand 53.9% → 14.3%
- Brand 60.0% → 7.1%
- Brand 771.4% → 75.0%
- Brand 83.8% → 7.3%
- Brand 95.8% → 8.2%
- Brand 103.7% → 6.0%
- Brand 114.1% → 5.7%
- Brand 1222.5% → 23.2%
- Brand 130.0% → 0.0%
- Brand 140.0% → 0.0%
- Brand 150.0% → 0.0%
- Brand 160.0% → 0.0%
- Brand 170.0% → 0.0%
- Brand 180.0% → 0.0%
- Brand 190.0% → 0.0%
- Brand 200.0% → 0.0%
- Brand 2182.1% → 82.0%
- Brand 221.9% → 0.0%
- Brand 234.3% → 2.0%
- Brand 243.6% → 0.0%
- Brand 253.9% → 0.0%
- Brand 2618.5% → 14.3%
- Brand 2731.5% → 25.0%
- Brand 289.4% → 0.0%
- Brand 2932.7% → 23.1%
- Brand 3010.7% → 0.0%
- Brand 3111.8% → 0.0%
- Brand 3221.2% → 9.1%
- Brand 3314.0% → 1.8%
- Brand 3422.2% → 9.1%
- Brand 3523.2% → 9.3%
- Brand 3632.1% → 17.0%
- Brand 3716.1% → 0.0%
- Brand 3838.0% → 21.4%
- Brand 3937.5% → 20.0%
- Brand 4032.7% → 14.3%
- Brand 4126.8% → 7.4%
- Brand 4235.7% → 16.0%
- Brand 4334.0% → 11.3%
- Brand 4446.0% → 19.2%
- Brand 4560.2% → 29.6%
- Brand 4652.4% → 12.5%
- Brand 4757.1% → 16.4%
Answers that cite the brand (%)
- In French
- In Dutch
View the numbers
Without web search
| In French | In Dutch | |
|---|---|---|
| Brand 1 | 62.5% | 54.8% |
| Brand 2 | 12.5% | 20.8% |
| Brand 3 | 18.8% | 18.8% |
| Brand 4 | 37.5% | 25.0% |
| Brand 5 | 8.3% | 10.0% |
| Brand 6 | 0.0% | 0.0% |
| Brand 7 | 60.9% | 72.5% |
| Brand 8 | 8.7% | 8.5% |
| Brand 9 | 0.0% | 0.0% |
| Brand 10 | 37.6% | 44.6% |
| Brand 11 | 0.0% | 0.0% |
| Brand 12 | 23.9% | 39.5% |
| Brand 13 | 0.0% | 0.0% |
| Brand 14 | 0.0% | 0.0% |
| Brand 15 | 2.1% | 0.0% |
| Brand 16 | 0.0% | 0.0% |
| Brand 17 | 0.0% | 0.0% |
| Brand 18 | 0.0% | 0.0% |
| Brand 19 | 4.5% | 0.0% |
| Brand 20 | 2.2% | 2.1% |
| Brand 21 | 83.3% | 100.0% |
| Brand 22 | 38.6% | 17.8% |
| Brand 23 | 0.0% | 4.0% |
| Brand 24 | 27.1% | 19.0% |
| Brand 25 | 8.5% | 6.8% |
| Brand 26 | 11.6% | 17.4% |
| Brand 27 | 33.3% | 46.5% |
| Brand 28 | 14.6% | 12.5% |
| Brand 29 | 20.8% | 38.1% |
| Brand 30 | 16.3% | 0.0% |
| Brand 31 | 4.3% | 0.0% |
| Brand 32 | 0.0% | 11.9% |
| Brand 33 | 18.8% | 4.2% |
| Brand 34 | 38.1% | 35.4% |
| Brand 35 | 27.4% | 29.4% |
| Brand 36 | 4.4% | 2.1% |
| Brand 37 | 4.3% | 0.0% |
| Brand 38 | 66.7% | 63.8% |
| Brand 39 | 10.4% | 14.6% |
| Brand 40 | 3.2% | 1.4% |
| Brand 41 | 25.0% | 0.0% |
| Brand 42 | 67.4% | 70.1% |
| Brand 43 | 14.6% | 12.5% |
| Brand 44 | 59.1% | 61.7% |
| Brand 45 | 76.7% | 64.9% |
| Brand 46 | 3.3% | 5.3% |
| Brand 47 | 2.1% | 0.0% |
With web search
| In French | In Dutch | |
|---|---|---|
| Brand 1 | 9.3% | 30.8% |
| Brand 2 | 7.1% | 25.0% |
| Brand 3 | 28.8% | 45.1% |
| Brand 4 | 39.3% | 51.8% |
| Brand 5 | 3.9% | 14.3% |
| Brand 6 | 0.0% | 7.1% |
| Brand 7 | 71.4% | 75.0% |
| Brand 8 | 3.8% | 7.3% |
| Brand 9 | 5.8% | 8.2% |
| Brand 10 | 3.7% | 6.0% |
| Brand 11 | 4.1% | 5.7% |
| Brand 12 | 22.5% | 23.2% |
| Brand 13 | 0.0% | 0.0% |
| Brand 14 | 0.0% | 0.0% |
| Brand 15 | 0.0% | 0.0% |
| Brand 16 | 0.0% | 0.0% |
| Brand 17 | 0.0% | 0.0% |
| Brand 18 | 0.0% | 0.0% |
| Brand 19 | 0.0% | 0.0% |
| Brand 20 | 0.0% | 0.0% |
| Brand 21 | 82.1% | 82.0% |
| Brand 22 | 1.9% | 0.0% |
| Brand 23 | 4.3% | 2.0% |
| Brand 24 | 3.6% | 0.0% |
| Brand 25 | 3.9% | 0.0% |
| Brand 26 | 18.5% | 14.3% |
| Brand 27 | 31.5% | 25.0% |
| Brand 28 | 9.4% | 0.0% |
| Brand 29 | 32.7% | 23.1% |
| Brand 30 | 10.7% | 0.0% |
| Brand 31 | 11.8% | 0.0% |
| Brand 32 | 21.2% | 9.1% |
| Brand 33 | 14.0% | 1.8% |
| Brand 34 | 22.2% | 9.1% |
| Brand 35 | 23.2% | 9.3% |
| Brand 36 | 32.1% | 17.0% |
| Brand 37 | 16.1% | 0.0% |
| Brand 38 | 38.0% | 21.4% |
| Brand 39 | 37.5% | 20.0% |
| Brand 40 | 32.7% | 14.3% |
| Brand 41 | 26.8% | 7.4% |
| Brand 42 | 35.7% | 16.0% |
| Brand 43 | 34.0% | 11.3% |
| Brand 44 | 46.0% | 19.2% |
| Brand 45 | 60.2% | 29.6% |
| Brand 46 | 52.4% | 12.5% |
| Brand 47 | 57.1% | 16.4% |
The distribution of the gaps shows it too. Its standard deviation goes from 8.8 points without the web to 13.4 points with it. Its mean slides from -0.5 to -6.4 points, to the detriment of Dutch (paired t-test, p < 0.01).
With the web, the gap between the languages widens
47 brands. Bars: the measured gaps. Curves: normal distribution with the same mean and the same standard deviation.
Number of brands
Dutch minus French gap (percentage points)
- Without web search: standard deviation 8.8 points
- With web search: standard deviation 13.4 points
View the numbers
| Dutch minus French gap (percentage points) | Without web search | With web search |
|---|---|---|
| -60 – -55 | 0 | 0 |
| -55 – -50 | 0 | 0 |
| -50 – -45 | 0 | 0 |
| -45 – -40 | 0 | 1 |
| -40 – -35 | 0 | 1 |
| -35 – -30 | 0 | 1 |
| -30 – -25 | 0 | 1 |
| -25 – -20 | 2 | 1 |
| -20 – -15 | 1 | 7 |
| -15 – -10 | 3 | 6 |
| -10 – -5 | 2 | 3 |
| -5 – 0 | 14 | 6 |
| 0 – 5 | 16 | 14 |
| 5 – 10 | 3 | 1 |
| 10 – 15 | 3 | 2 |
| 15 – 20 | 3 | 2 |
| 20 – 25 | 0 | 1 |
| 25 – 30 | 0 | 0 |
| 30 – 35 | 0 | 0 |
| 35 – 40 | 0 | 0 |
| 40 – 45 | 0 | 0 |
| 45 – 50 | 0 | 0 |
| 50 – 55 | 0 | 0 |
| 55 – 60 | 0 | 0 |
| Without web search | μ = -0.5; σ = 8.8 |
|---|---|
| With web search | μ = -6.4; σ = 13.4 |
This is how I explain it: what the AI has learned about a brand holds for every language, while the pages it finds on the web depend on the language of the search. A brand with no content in Dutch loses citations in Dutch searches.
Does the buyer's place change the answer?
Place changes the list even more than language.
I compared the companies the AIs suggest for the same profession (lawyer, accountant, insurance broker, dentist, real estate agent, marketing or SEO agency) in Brussels, Luxembourg, Paris, Montréal and Québec City, with web search.
In the same city, two languages share on average 21% of the companies (45 comparisons). In the same language, two cities share only 2% (67 comparisons).
Changing city changes almost the whole list
Companies suggested by the AIs, with web search, for the same profession. Brussels, Luxembourg, Paris, Montréal, Québec City.
- Same city, other language
- Same language, other city
View the numbers
| Same city, other language | 20.8% of the companies in common (45 comparisons) |
|---|---|
| Same language, other city | 2.4% of the companies in common (67 comparisons) |
The names that travel from one city to another are mostly those of the big groups: Aon, AXA and Marsh for brokers, Dentons, Linklaters and Norton Rose Fulbright for lawyers.
For local businesses, it is another story: the AIs suggest them in their own city, rarely elsewhere.
So I measure a brand in each city where its buyers are, then in each language of that city.
Do all AIs read a brand the same way?
The AI you ask weighs as much as the brand itself.
On 5,820 Presence answers with web search (51 brands, from 29 September to 12 November), I broke down the variance of the citation rate. The measurement (the brand, its language, its place, its type of buyer) explains 26.7% of it. The AI asked also explains 26.7%. The wording of the question explains 9.8%. The remaining 36.8% mixes chance and the combinations between these factors.
For brands cited between 20% and 80% of the time, the AI's share rises to 37%, against 12% for the question.
Where the citation gaps come from
Share of the variance, with web search. 51 brands, 5,820 answers, from 29 September to 12 November 2026.
- The measurement (brand, language, place, type of buyer)
- The AI asked
- The wording of the question
- The rest (chance and interactions)
View the numbers
| The measurement (brand, language, place, type of buyer) | 26.7% |
|---|---|
| The AI asked | 26.7% |
| The wording of the question | 9.8% |
| The rest (chance and interactions) | 36.8% |
Across the six industrial companies, Gemini suggests the company in 58% of its answers (55 of 95; 48 to 67%). Mistral suggests it in 16% (15 of 96; 10 to 24%).
Claude suggests none of the four least-known companies (0 of 64). Their names do not even appear in its answers. Gemini suggests one of them in all of its own (16 of 16).
Six industrial companies, as seen by each AI
6 industrial companies, 2 languages each, with web search, 12 audits. Measured on 14 November 2026.
Answers that suggest the company to the buyer (%)
- With web search
View the numbers
| Answers that suggest the company to the buyer (%) | |
|---|---|
| Gemini | 57.9% (55 of 95) |
| Grok | 49.0% (47 of 96) |
| Perplexity | 45.3% (43 of 95) |
| ChatGPT | 44.1% (41 of 93) |
| Muse Spark | 34.1% (31 of 91) |
| Claude | 21.1% (20 of 95) |
| Mistral | 15.6% (15 of 96) |
The AIs do not suggest the same companies either. For the same question, two different AIs share on average 9% of the companies (20,629 comparisons). The same AI, on the same day, repeats 41% of them (112 comparisons).
Does an AI always give the same answer?
The same question does not always get the same answer.
On 14 November, I asked the same questions twice during the day, for 7 companies. 17.5% of the answers changed (84 of 480; 14 to 21%), for brands cited between 20% and 80% of the time. The time of day matters little: 16% of changes 4 minutes apart, 18% 11 hours apart.
About a third of these changes come from my reading of the answers, not from the AIs. That is true of any tool that counts names in answers.
So the precision of a measurement depends on the number of answers. From one run to the next, its standard deviation is the square root of p ÷ 2n: p is the share of answers that change between two runs, n the number of answers in the measurement.
For the seven AIs together, p is 17.5%. For ChatGPT alone, p is 37% (20 answers of 54). I get:
- ChatGPT alone, 8 questions: standard deviation of 15.2 points, or ±30 points at 95%;
- 7 AIs in one run (56 answers): standard deviation of 4.0 points, or ±7.7 points;
- 7 AIs in two runs (112 answers): standard deviation of 2.8 points, or ±5.5 points.
What a single measurement can give
Brand cited in 40% of the answers. Standard deviations taken from the two runs of 14 November 2026; in brackets, 95% interval.
true value: 40%
Measured citation rate (%)
- ChatGPT alone, 8 questions: standard deviation 15.2 points (±29.8)
- 7 AIs, 1 run (56 answers): standard deviation 4.0 points (±7.7)
- 7 AIs, 2 runs (112 answers): standard deviation 2.8 points (±5.5)
View the numbers
| true value: 40% |
| ChatGPT alone, 8 questions: standard deviation 15.21 points (±29.82) |
| 7 AIs, 1 run (56 answers): standard deviation 3.95 points (±7.75) |
| 7 AIs, 2 runs (112 answers): standard deviation 2.80 points (±5.48) |
A GetMint Research paper reaches the same order of magnitude for ChatGPT: ±28 points for one question measured on a single day, across more than a million answers (Amine Aziz Alaoui, "A Measurement Model for Brand Visibility in ChatGPT Answers", 5 October 2026).
Does an AI write your name the way you do?
The AI does not always use the exact name of the brand.
It sometimes attaches a product, a country or a legal form to the name: "Name S.A.", "Name France", the name followed by a product range. At Company C, recognising these forms raises its Presence in English from 59% to 75% (from 66 to 83 answers of 111).
The name can also be an old one. This company has changed its name. 42% of the answers to "what is it?" still give the old one (11 of 26).
What I take from this for a brand
- Start with Perception. Before trying to be cited, I check what the AIs say about a brand when it is named.
- Publish the basic facts in each language of your buyers. Headquarters, founding date, headcount, products, markets. The errors about Cebi concern exactly these facts.
- Clean up what is out of date. Former name, former shareholders: the AIs repeat them, even with web search.
- Measure in each city and each language of your buyers. Two cities with the same language share only 2% of the companies suggested.
- Measure several AIs, in both modes, in two runs. With 7 AIs and two runs, the standard deviation of a measurement falls to 2.8 points, against 15.2 points for ChatGPT alone.
- Run the same measurement again later. Otherwise, there is no way to know whether an action changed anything.
Places to take
The AIs answer with what they find, city by city and language by language. They get it wrong twice as often without web search.
The big groups hold in every city. For mid-sized and local businesses, which the AIs still know poorly, the places are there to take.
Do you know what the AIs say about your brand in the language and in the city of your buyers?
