Measuring AI Visibility in Switzerland

Two numbers describe ChatGPT's position in the same month. One says 77.92%, the other says 53.9%. Both come from named measurement companies. Both are correct. They are measuring opposite directions of traffic, and an agency that puts them in the same slide has told you nothing except that it did not read the definitions.

Switzerland makes this worse than most markets, because until recently country-level assistant data barely existed and Swiss GEO plans were built on worldwide averages. That data now exists, and it shows the Swiss mix is measurably different from the global one in a specific direction. Which is useful, and which also creates a fresh opportunity to misread things.

This article covers the Swiss assistant figures and what they actually measure, the three separate Swiss adoption surveys and why merging them destroys the information, what the peer-reviewed GEO evidence supports, why a single citation score is not a performance indicator, and the reporting structure that follows from all of that.

The Swiss assistant mix, and what the number means

Statcounter publishes an AI chatbot share by country. For Switzerland in July 2026 the split ran ChatGPT 73.31%, Google Gemini 8.94%, Claude 6.69%, Perplexity 5.82%, Microsoft Copilot 5.10%, DeepSeek 0.09%.

Worldwide in the same month: ChatGPT 77.92%, Gemini 9.9%, Perplexity 5.88%, Claude 3.19%, Copilot 3.07%, DeepSeek 0.03%.

Read the two lists against each other. Claude sits third in Switzerland and fourth globally, with more than double the Swiss share. Copilot runs at 5.10% here against 3.07% worldwide. ChatGPT runs lower here than globally. Gemini and Perplexity are close enough to be unremarkable. So a plan built on worldwide averages systematically underweights two of the five engines that matter in this market, and overweights the largest one.

Now the definition, because everything depends on it. This measures the share of AI chatbot referrals to websites, meaning which assistants send visitors onward to sites, observed across Statcounter's measurement network. It is not a count of how many people use each assistant. It is not a census. And it has moved substantially month to month through 2026, which is a reason to re-baseline rather than to treat one month as settled.

That definition makes it the right metric for a specific question and the wrong one for another. If you want to know which assistants can plausibly send you traffic, referral share is exactly the thing to look at. If you want to know which assistants Swiss consumers open in the morning, it is not.

The other family of numbers, and why the contradiction is fake

A separate set of figures measures web visits to the assistant products themselves. Similarweb data circulated via Momentic put ChatGPT at 53.9% of worldwide assistant web-visit share in May 2026, down from 54.5% in April and 79.0% a year earlier, with Gemini at 27.9% and Claude at 9.2%. A different Similarweb tracker read ChatGPT 64.5% and Gemini 21.5% in early January 2026, against ChatGPT 86.7% and Gemini 5.7% twelve months previously.

Set those against the referral numbers and they look irreconcilable. ChatGPT at 77.92% or at 53.9%. Gemini at 9.9% or at 27.9%. Claude at 3.19% or at 9.2%.

They are not in conflict. One family counts traffic flowing out of assistants to websites. The other counts traffic flowing into assistants from users. Gemini has a large user base and sends comparatively few onward referrals, which is what you would expect from a product embedded in Google's own surfaces. The web-visit family also excludes native app usage, and Momentic's own commentary notes that app use adds somewhere between 15% and 60% on top of web visits depending on the platform.

So both families are valid and neither substitutes for the other. We use the referral family for Swiss planning, because it has Swiss country data and because it answers the traffic question directly. We do not average the two, and we do not put them on the same axis.

Three Swiss adoption surveys, three different questions

Swiss AI adoption gets quoted anywhere between the high thirties and the high seventies. The range is not a sign that the research is bad. It is a sign that the surveys asked different questions of different populations, and that most people quoting them did not check which.

Comparis and Innofact surveyed 1,035 Swiss adults in March 2026 and published in April 2026. Result: 76.1% use AI chatbots at least occasionally, up from 62.4% in 2025 and 49.7% in 2024. Separately, 41.6% reported using AI for web search instead of a search engine. The same survey found a regional gap worth noting: 81.6% adoption in Romandie against 74.1% in German-speaking Switzerland. Ticino was not published separately, so we do not interpolate a figure for it.

The University of Zürich WIP-CH 2025 study, fieldwork 2 June to 27 August 2025, 1,078 respondents, published November 2025. Result: 73% have ever used generative AI, up from 54% in 2024 and 37% in 2023. But only 47% use it at least monthly, with 21% weekly and 10% daily.

On the enterprise side, Deloitte's AI ROI research, fieldwork 15 August to 5 September 2025, reported 53% of its 99 Swiss respondents using strategic AI tools against a 48% average across the full 1,854-respondent sample.

Now look at what happens if you merge them. Occasional use, ever-used, monthly use and daily use are four thresholds. Individuals and organisations are two populations. A national consumer survey and an enterprise sample of 99 are two evidence bases. Averaging any of these produces a number that answers no question anybody asked.

There is a smaller cautionary detail inside one of them. The Zürich report gives 10% for daily use, while its own executive summary reports 17%, because the summary folds daily together with several-times-daily. Both figures are defensible once the grouping is stated. Neither is usable without it. If a single study's own summary can shift a headline number by seven points through grouping alone, an agency quoting one figure with no definition attached should not be trusted with the rest of the report.

What the peer-reviewed GEO evidence supports

There is one study that anchors this field, and it is worth quoting in its own numbers rather than in the ranges that circulate in vendor material.

Aggarwal and colleagues presented work at ACM SIGKDD in 2024 built on a benchmark of 10,000 queries. It reported visibility increases of over 40% across various queries, with the strongest methods reaching 41% on a position-adjusted word count measure and 28% on subjective impression. Adding statistics, citations and quotations helped. Keyword stuffing did not.

We flag that precision deliberately. Across four research passes, one model reported this same study as showing 30% to 40%, and added a Perplexity-specific figure of up to 37% that no other pass reproduced and that we could not trace to the paper. When a source degrades a paper's own numbers into a vaguer range and then adds a specific figure the paper does not contain, that is a signal about the source rather than about the research.

The qualifications matter as much as the headline, and they come from later academic work. Rewriting content purely to win citations can degrade retrieval, so optimising the citation stage at the expense of the retrieval stage can lose you visibility overall. GEO is also multistage, covering retrieval, context allocation, mention, citation, prominence and fidelity. Improving one stage does not demonstrate improvement in another, which means a report showing more mentions is not automatically a report showing better outcomes.

Why a single citation score is not a performance indicator

This is the part that separates measurement from theatre.

A review of roughly 45 GEO studies published between 2023 and 2026 found low source overlap between runs, substantial run-to-run variability, and persistent fidelity gaps. It also found that topical relevance and position within the retrieved context were the most reproducible levers, while generic heuristics transferred poorly between engines. Separate academic work notes that visibility depends on the engine, the date, the location, the exact query phrasing and whether search was activated for that response, so a point estimate is not a stable performance indicator.

Translate that into agency practice. If an agency sends you one number per engine per month, they have run each prompt roughly once and presented a sample as a measurement. The number will move next month whether or not anything they did worked, and neither of you will be able to tell which.

The alternative is not complicated, it is just more work. Run each tracked prompt multiple times per cycle, per engine, per locale. Report the distribution rather than a value. State the run count. Flag when a change sits inside normal run-to-run variation rather than presenting noise as progress. That last discipline is the one that costs an agency short-term credit and earns long-term trust, which is presumably why it is rare.

Metric on the report	What must be stated with it	The failure it prevents
Share of voice	Engine, locale, date window, run count, and a range rather than a single value	Presenting a single run as a stable measurement, then explaining away next month's movement
Citation rate	The prompt set size, how prompts were selected, and whether the set changed since last cycle	Improving the number by quietly editing the prompt set rather than the content
Source-origin ratio	Country classification method, and which language the prompts were in	Mixing German-language and English-language runs, which have different source pools
Assistant platform weighting	Whether the weighting came from referral share or web-visit share, and whether it is Swiss or worldwide	Using worldwide averages in a market where Claude and Copilot both run materially higher
Adoption context	Survey name, sample size, fieldwork dates, and the exact threshold measured	Quoting 76.1% and 47% as if they answer the same question about the same population
Change since last cycle	Whether the change exceeds observed run-to-run variation	Claiming credit for noise, which works until the noise moves the other way

The Swiss figures that do not exist

Four independent research passes, and four gaps that came back empty every time. They are worth naming because they are the gaps most likely to be filled with invention.

Nobody has published a Swiss AI Overview trigger rate. There is one figure circulating from a marketing blog covering 374 French-language retail-banking keywords in Romandie, and that is one vertical, one language region and one interested party, so it cannot carry a national claim. Availability is documented, since Google launched AI Overviews in Switzerland on 25 March 2025 in four languages to signed-in users aged 18 and over. Availability and prevalence are different measurements, and the gap between those two words is where most agency claims in this category fall apart.

There is no credible size figure for the Swiss GEO or AEO services market. There is no published measurement of how often German-domain sources displace Swiss ones in AI answers. And there is no independently audited citation-uplift case study from any named Swiss agency, a gap three of our four research passes reported explicitly.

That last one deserves a sentence more, because it cuts both ways. It means nobody in this market can show you audited proof, including us. It also means an agency presenting a Swiss citation-uplift case study should be asked who audited it, over what period, on which engines, with how many runs.

What a defensible Swiss baseline looks like

Pulling it together, a first cycle that can actually be built on has a specific shape.

Separate tracked prompt sets for de-CH, fr-CH and it-CH where Ticino matters, plus an English set for the international business audience, since 23.6% of the Swiss working population uses English at work. Each prompt run several times per engine per cycle, with the count recorded. Engines weighted to the Swiss referral mix rather than the worldwide one, which in practice means Claude and Copilot get real attention rather than a footnote. Every cited source classified by country of origin, per language, kept separate. And every figure in the report carrying its engine, locale, run count and date.

None of that guarantees a citation. It does mean that when something changes, you can tell whether it changed, which is the precondition for everything else. The Swiss market currently has no agency publishing audited citation uplift, so measurement discipline is the only differentiator available that a buyer can verify for themselves.

Tessar Napitupulu covers AI citation measurement, run-to-run variability and the difference between visibility metrics and vanity metrics in Cited or Silent, available as a free gated edition, with retailer editions on Amazon, Google Play and Apple Books. The measurement approach also underpins our AI Citation Rate Report 2026.

---

Frequently Asked Questions

Which AI assistants matter most in Switzerland?

Different ones than the worldwide averages suggest. Statcounter's July 2026 figures for share of AI chatbot referrals to websites give Switzerland ChatGPT 73.31%, Google Gemini 8.94%, Claude 6.69%, Perplexity 5.82%, Microsoft Copilot 5.10% and DeepSeek 0.09%. Worldwide the same month ran ChatGPT 77.92%, Gemini 9.9%, Perplexity 5.88%, Claude 3.19%, Copilot 3.07%. Claude ranks third here against fourth globally with more than double the share, and Copilot also runs materially higher. A plan built on worldwide figures underweights two of the five engines that matter in this market.

Why do different sources give wildly different platform shares?

Because two separate metrics are in circulation and they run in opposite directions. Referral share counts which assistants send visitors onward to websites. Web-visit share counts how much traffic the assistants themselves receive from users. Gemini has a large user base and comparatively few outward referrals, which is what an assistant embedded in Google's own surfaces would produce, and that single fact explains most of the apparent contradiction. The web-visit family also excludes native app usage, which reportedly adds 15% to 60% on top depending on platform. Both are valid. Neither substitutes for the other, and they should never be averaged.

What share of Swiss adults use AI assistants?

Between roughly 47% and 76% depending on the threshold measured, and the range is the answer. Comparis and Innofact, surveying 1,035 adults in March 2026 and publishing April 2026, found 76.1% use AI chatbots at least occasionally, up from 62.4% in 2025 and 49.7% in 2024, with 41.6% using AI for search rather than a search engine. The University of Zürich WIP-CH 2025 study, fieldwork 2 June to 27 August 2025 with 1,078 respondents, found 73% had ever used generative AI but only 47% did so at least monthly, with 21% weekly and 10% daily. Occasional, ever-used, monthly and daily are four different questions.

Is adoption the same across the language regions?

No, and the gap is measurable in one of the surveys. Comparis and Innofact recorded 81.6% chatbot adoption in Romandie against 74.1% in German-speaking Switzerland in their March 2026 survey of 1,035 adults. Ticino was not published separately in that survey, so we do not interpolate a figure for it and neither should anyone else. Note that other Swiss AI surveys did not publish regional breakdowns at all, which is a difference in source coverage rather than a contradiction, and it is worth checking which survey a regional claim actually came from.

What share of Swiss searches trigger an AI Overview?

Nobody has published a Swiss figure, and an agency quoting one should be asked for its query set. What is documented is availability rather than prevalence: Google launched AI Overviews in Switzerland on 25 March 2025 for signed-in users aged 18 and over, in German, French, Italian and English, while Germany and Austria received German and English only. Switzerland was the only country in that wave with four languages, and coverage has since expanded well beyond it. One circulating figure covers 374 French-language retail-banking keywords in Romandie, which is one vertical, one language region and an interested publisher, so it cannot support a national claim.

Is there peer-reviewed evidence that GEO techniques work?

Yes, with qualifications that matter as much as the finding. Aggarwal and colleagues, at ACM SIGKDD 2024, built a benchmark of 10,000 queries and reported visibility increases of over 40% across various queries, with the strongest methods reaching 41% on position-adjusted word count and 28% on subjective impression. Adding statistics, citations and quotations helped; keyword stuffing did not. Later academic work warns that rewriting purely for citation can degrade retrieval, and that GEO is multistage across retrieval, context allocation, mention, citation, prominence and fidelity, so improving one stage does not prove improvement in another.

Why does one run per month per engine not count as measurement?

Because identical prompts return different source sets on different runs. A review of roughly 45 GEO studies published between 2023 and 2026 found low source overlap between runs, substantial run-to-run variability and persistent fidelity gaps, with topical relevance and position in the retrieved context the most reproducible levers and generic heuristics transferring poorly. Visibility also depends on engine, date, location, phrasing and whether search was activated. A single run per engine is a sample presented as a result, and the figure will move next cycle whether or not the work succeeded, leaving nobody able to tell which.

Can any Swiss agency show audited proof of citation improvement?

Not that four independent research passes could locate, and that includes us. Three of the four passes reported explicitly that no named Swiss agency publishes an independently audited citation uplift. That cuts both ways: it means nobody in this market can offer you third-party verified proof today, and it means an agency presenting a Swiss citation case study should be asked who audited it, over what period, on which engines and with how many runs per prompt. Where audited market evidence does not exist, verifiable method is the only differentiator a buyer can check.