AI search in Luxembourg: your language picks the winner

Learn more about AI implementation in Luxembourg and the wider GEO playbook for Luxembourg firms.

In Luxembourg a buyer can ask the same question three ways: "Quelle fiduciaire pour ma SARL ?", "Welche Treuhand in Luxemburg?", "Which accountant in Luxembourg?" Most firms assume an AI assistant answers all three the same way, from the same idea of who is good. Our measurement says it does not, and the difference is large enough to decide whether your firm is in the answer at all.

I run the AI-visibility work at 20 More, and this is the finding I now open every client conversation with, because it changes where the effort goes. Below: what we measured, what the gap looks like, what we think drives it, and a language-by-language plan you can run yourself.

What did we measure, and what did we not?

The data comes from Edition 1 of our AI Visibility Benchmark, measured on 8 July 2026 (7 answers re-measured on 28 July 2026):

• 405 AI answers to 80 questions asking for a fiduciary or accounting firm in Luxembourg
• 2 engines: ChatGPT (through the OpenAI API with its native web search) and Perplexity Sonar
• 3 languages: French, German and English
• 65 tracked firms: 44 independent firms and 21 large or network firms

Three limits matter before you read any number. It is one measurement date, so it says nothing about how answers move over weeks. It uses the APIs, not the consumer apps, which add their own prompts, memory and ranking, so the figures are a close proxy rather than the product. And it covers one sector, fiduciary and accounting services. We also sell AI-visibility audits, which is why the method and the prompt set are published for anyone to challenge.

How big is the language gap?

This is the share of answers, per language, that named at least one tracked firm in each group:

Read it as three different markets:

• In French, independents compete. Six answers in ten named an independent firm, and 30 different independents surfaced at least once. The field is wide.
• In English, the networks take the answer. Three answers in four named a large or network firm, while only 14 independents appeared at all. An independent firm whose site and listings exist mainly in English is competing in the language where it is weakest.
• In German, most of the field is empty. German answers were the least likely to name a firm from either group: 40% named an independent, 16% a network firm. For a firm with a German-speaking clientele, that is less competition, not less demand.

The engines also behaved differently. ChatGPT's 195 answers named an independent in 52% of cases and read 5.3 sources per answer; Perplexity's 210 answers named an independent in 46% and read 8.2 sources. A firm cited by more sources has more ways into a Perplexity answer.

Why do the answers diverge by language?

The benchmark records what the engines answered, not why, so treat this section as our reading of the evidence rather than a proven cause.

The most likely mechanism is retrieval. Both engines search the web before answering, and a French question retrieves French pages. The third-party sources the engines drew on most across all 405 answers were:

Several of these are French-language directories and guides that list local firms by name. When the question is in English, the web the engine retrieves is different: international rankings, English-language comparison pages and the large networks' own English content. The firm is the same; the evidence the engine finds about it is not.

From the client side, this matches what I see when we run a Snapshot. A firm's English page is often a translated summary, its French page the real one, and its German page missing. The engines reward exactly that imbalance: strong where the content is full, absent where it is thin.

What should a Luxembourg firm do about it?

Treat each language as its own channel, with its own measurement.

1. Measure per language, not per engine. Run the same ten buyer questions in French, German and English and record two things separately: whether you are named in the answer, and whether your website is cited as a source. Named is "does the engine recommend you?"; cited is "does the engine read you?". A firm can lead on one and be absent from the other, which is why our benchmark never adds them together.

2. Give every language a complete page. A German page that is a two-line summary of the French one gives a German-language answer nothing to quote. The pages that earn citations answer a buyer's question directly in the first lines, name the service and the client type, and state facts an engine can reuse: who you serve, where, and what it costs.

3. Fix the directories before the website. More than a quarter of answers drew on editus.lu, and Kompass appeared in 12%. An outdated or incomplete listing there travels straight into AI answers. Check that your name matches your RCS registration, that your services are listed in each language you work in, and that the listing links to the right language version of your site.

4. Pick your battles by language. If your clients are mostly French-speaking, the French answer is where independents already win, and the work is to be one of the 30 names rather than absent. If you serve German-speaking clients, the German field is thin: you are competing with fewer firms. If your growth market is English-speaking expatriates and international companies, expect the networks to dominate and compete on specific questions (a sector, a company type, a fee model) rather than on "best accountant in Luxembourg".

5. Re-measure. One date proves nothing about stability. Run the same question set every month and watch the trend, not the snapshot.

For a firm that wants the measurement done properly, our AI Search service publishes its prices and scope, starting with the free Snapshot.

A question set you can run this week

These are the kinds of questions we use, one per language, per buyer situation. Replace the service and the client type with your own:

Ask each question in a fresh session, save the full answer and the sources listed, and mark per answer: named (yes/no), cited (yes/no), and which competitors appeared. Twenty minutes per language gives you a baseline no dashboard provides.

Will the gap hold outside fiduciary services?

We do not know yet, and I would rather say so than extrapolate. The benchmark measured one sector. Our working hypothesis is that the pattern holds wherever a sector has strong local French-language directories and a group of large international players with strong English content: legal services, real estate, recruitment and IT services all look similar on paper. That is a hypothesis to test, not a finding, and Edition 2 of the benchmark repeats the same versioned prompt set so the numbers stay comparable over time.

If you want to know before then, measure your own sector with the question set above. It costs an afternoon.

Frequently asked questions

Does the language of a question change AI recommendations?

Yes, in our measurement it changed them substantially. Across 405 answers from ChatGPT and Perplexity on 8 July 2026, French questions named an independent Luxembourg fiduciary in 61% of answers, English questions in 34%, and English answers named a large network firm in 74% of cases.

Which sources do AI engines use for Luxembourg firms?

In our fiduciary benchmark the most-cited third-party sources were editus.lu (27% of answers), financialservices.lu (23%), guichet.public.lu (17%), fiduciaire-expert.lu (16%), atlasfiduciaire.lu (14%) and lu.kompass.com (12%). Keeping your listings there accurate, in every language you serve, is the cheapest visibility work available.

Is ChatGPT different from perplexity for Luxembourg searches?

Somewhat. In our data ChatGPT (via the API with web search) named an independent firm in 52% of answers and read 5.3 sources per answer; Perplexity Sonar named one in 46% and read 8.2 sources. Perplexity draws on more sources, so being present on more reputable sites matters more there.

Can an agency guarantee a place in AI answers?

No. No agency can guarantee placement in AI-generated answers, and you should be wary of anyone who promises it. What can be done is measuring where you appear, per language and per engine, fixing the sources the engines read, and tracking whether the share of answers that name you moves.