Choosing a GEO agency in Switzerland: six questions

You have three GEO offers on the table. They all promise roughly the same thing: make your company more visible in ChatGPT, Google or Perplexity. On paper, they are hard to tell apart.

In practice they do not necessarily measure the same engines, the same questions or the same markets. Two proposals at the same price can therefore cover very different ground.

That is not surprising. The discipline is under two years old, its vocabulary is still moving, and everyone presents their method differently. For an SME, the useful question is not who uses the most impressive terms. It is what will actually be checked.

There is a quieter risk too. An agency can know GEO well and still target the wrong websites. In French-speaking Switzerland, the sources that matter are not the French ones, nor always the Zurich ones.

We do not rank agencies and we recommend none. We measure daily what the assistants answer across ten sectors of the Swiss-French market. Since April 2026, that observatory, Swiss Atlas, has collected 65,106 citations.

Those figures produce six simple questions to ask before signing. They will not automatically identify the best provider. They will make vague offers show themselves quickly.

Accuracy comes before visibility

Many companies want to appear more often. Few have checked what is already being said about them.

In our corpus, 97.1 % of the citations in an AI Overview come from a third-party site. In other words, what the AI says about your company was almost always written elsewhere: a newspaper, a comparison site, a directory, a forum, sometimes a competitor. Your own site accounts for 2.9 %.

That changes how you fix an error. An outdated opening time, a price that has moved, a certification credited to someone else: updating your own site will not always be enough. You have to go back to the page carrying the wrong information.

An agency cannot guarantee what ChatGPT will say tomorrow. It does not write the answer. What it can guarantee is a control, and that is a very different thing.

In practice that means three things. A written list of what is true about you: opening hours, prices, what you do and do not do, who runs the company, which certifications. A regular check of what each assistant says on those points. And an alert when a gap appears, with the correction pursued at source.

This is joint work. Nobody can build that list in your place.

Ask precisely what will be measured

Tracking "AI" means nothing. Engines have to be named, and they do not behave alike.

Swiss Atlas, April to August 2026.

Nine sources instead of six sounds secondary. For a small brand it is not: there are simply more places to take. Conversely, ChatGPT answers with no source one time in five, which makes counting links less telling there.

The choice depends first on your customers. An industrial buyer, a patient and a private individual do not use the same tools. AI Overview deserves particular attention: it appears directly in Google, without anyone asking to speak to an AI. It therefore reaches people who use no assistant at all.

Each additional engine increases the measurement volume. It therefore increases the price. A very broad offer on a small budget usually hides a trade-off: fewer questions, or less frequent checks. That is not necessarily a problem, but it is worth knowing.

The questions matter as much as the engines

Everything depends on the questions asked. If they do not resemble your customers' questions, the score says little. A list written from the company's internal vocabulary always produces flattering results.

These questions are built with your sales team, who hear the objections, and your support team, who receive the requests.

One point is often forgotten: price and purchase questions.

Swiss Atlas, April to August 2026.

This table reads both ways. On price questions the AI rarely answers, and your classic search optimisation keeps the upper hand. But one purchase search in four still produces an AI answer. That is the one arriving just before the decision.

A list without price questions therefore measures your awareness, not your commerce.

Once the list exists, three indicators are enough: the share of answers where your brand appears, your share of mentions against the competitors being tracked, and where you are named. Ask to see them question by question. A single average hides exactly the one where you are losing ground.

GEO does not replace your website

"Classic search optimisation is over, GEO is next." The argument comes up often in offers. The figures say rather the opposite.

In our corpus, 61.5 % of the pages cited by AI Overview are also cited by at least one other engine. The work is therefore not redone engine by engine. A page that becomes citable is citable more or less everywhere. Google itself describes AI Overview as a layer over its existing index, in its documentation on AI features in Search.

Then there are the 23.4 % seen above. On purchase searches, Google mostly defers to the classic links. The part of the journey closest to the sale remains the one you already work on.

An agency that sets GEO against search optimisation will have to explain how it gets there.

Which pages, then? We looked at which ones keep their citation over several days.

Swiss Atlas, pages cited on at least 5 distinct days.

The product page comes clearly first. GEO is often said to play out on the blog: in our corpus, a page describing a service precisely, with its conditions, is cited longer than an article. The homepage accounts for 6.0 %.

Two technical details are worth asking about. Among the Swiss pages we analysed one by one, 63 % use structured data and 41 % show a named author. And of those 143 pages, 50 come from sites with no established authority. An SME is not ruled out in advance.

In Romandie, the good sources are often local

Over twelve weeks our observatory collected 65,106 citations across 313 sites. Of those, 216 are Swiss, seven in ten. But those 216 sites carry only 35.5 % of the citations.

The Swiss landscape is therefore wide and scattered. There are not three large sites to convince, there are dozens of mid-sized ones. Among the most cited: fr.comparis.ch, moneyland.ch, blick.ch, 24heures.ch, bonus.ch, ge.ch. The rest of the citations go to a handful of large platforms, Reddit and YouTube first.

Behind each of those Swiss sites there are journalists and teams. Not a form.

Getting a listing corrected at a Swiss comparison site, a mention in a Swiss newspaper, an update on a cantonal site: that is press relations work, here. It takes knowing the newsrooms and knowing who to write to. A method designed for another country produces target lists that carry no weight here.

Language adds a layer. We asked the same question about health insurance on the same day, on google.ch, in three languages.

A single question, as an illustration. Not a representative measurement.

No common source in first position. A brand well cited in German can be absent in Italian on the same question, and a single-language measurement would never show it.

The question is therefore not whether the agency is Swiss: "Name the ten sites the AI cites most in my sector and my language region. Which of them have you already worked?"

A one-day citation is not a result

Across all the pages cited by AI Overview in our corpus, 59.3 % are cited on a single day. Those holding for ten days or more account for 1.6 %.

One day is not enough.

A screenshot showing your brand in an AI answer is a signal. Not yet a result. The reverse holds too: your absence on a given day proves little either.

This applies on both sides of the table. On the agency side, showing an isolated citation and calling it a result. On the client side, typing your name once into ChatGPT and drawing a conclusion.

The only solid reading asks the same questions, of the same engines, over several weeks. That is also why history matters: it makes the figures comparable over time. Switching provider without it costs several months of hindsight.

The six questions

To ask as they are, in the meeting.

1. "Who checks that what the AI says about us is accurate, and how often?" A useful answer names a list of information and a frequency.
2. "Which engines do you measure, and why those?" The choice is justified by your customers. "All of them" is not a choice.
3. "Who writes the questions being tracked, and are there price questions?" Without price questions, the measurement covers awareness.
4. "Do you also measure in German, and are the results separated?" A Swiss average mixes two markets with different sources.
5. "Which sites will you contact? Name three." 97.1 % of what is said about a company comes from somewhere other than its own site.
6. "If I stop in a year, do I leave with the history?" History is what makes the figures comparable over time.

Ask them. The one answering with figures will not necessarily be the best, but at least you will know what you are buying.

Three sentences worth a follow-up question

Aiming at no one: the sector's vocabulary is still loose, and these three formulations come up often.

"We guarantee a place in the answers." There is neither a purchasable slot nor a registration procedure. The selection is also highly dispersed: our 10,759 citations spread across 2,002 distinct sites, and the top ten gather only 14.1 % of them. No mechanism supports a placement guarantee.

"Classic search optimisation is over." The 23.4 % above say otherwise. Google presents AI Overview in its rollout announcement as an entry point to websites, not as their replacement.

"Results in a few weeks." Possible, but unverifiable at that horizon, since 59.3 % of citations do not last beyond a day. What is measurable in the short term is the starting point.

What the engagement covers beyond Google

One last point to clarify: scope. The work does not play out only on Google and the assistants. Video and social platforms count too, and Reddit is the most cited site in our observatory across all sectors, ahead of YouTube.

These are spaces you do not control and where your company is described nonetheless. An offer covering only one of these surfaces should say so clearly.

Method and limits(4)

1. Corpus. Swiss Atlas, Repliq's observatory across ten sectors of the French-speaking Swiss market. Questions asked daily in French, from 20 April to 12 August 2026. The observatory of Swiss-French sources covers twelve weeks, August 2026 edition, 65,106 citations across 313 sites.
2. Stability. A page counts as cited on a given day if it appears in at least one answer that day. Duration groupings use distinct, not consecutive, days.
3. Formats. Markup and authorship figures come from 143 cited pages, analysed one by one. 80 further pages were not crawlable, which biases towards open sites.
4. What these figures do not say. They describe French-speaking Switzerland and do not transpose unchanged. The trilingual comparison covers a single question. The Swiss site split concerns the 313 tracked sites.

See the sources cited in French-speaking Switzerland→