GEO 2026: how ChatGPT and Perplexity cite you

Your site may rank first on Google. Yet when a customer asks their question to an AI assistant, it is not you who appears. Ask "what is the best online bank in Switzerland": the tool does not cite a bank, but a blog run by a private individual, mustachianpost.com. Over the observed period, its page « Meilleure banque Suisse (comparatif 2026) » is cited 92 times, ahead of all the institutions concerned.

This is not by chance, and it is not an exception. For health insurers, it is the comparison site compassurance.ch (77 citations) that dominates. For supermarkets, bonus.ch (52). For Geneva public transport, the City's official site, geneve.ch (49). Good classic search optimization no longer guarantees being cited by AI, because answer engines do not work like a search engine.

Repliq measured these citations every week, then dissected the most frequently reused pages, sector by sector, in French-speaking Switzerland. This guide explains what AI looks at in a page, how each engine differs, and what an organization can do concretely. All the cited pages are named: you can open them to check every point.

GEO, or how to exist in AI answers

Classic search optimization (SEO) optimizes a page so it rises in a list of links. GEO (Generative Engine Optimization) has a different goal: being cited in the answer written by an assistant. The difference is one of nature. A search engine displays 10 links and lets the user choose; an answer engine composes a single answer from a few sources, which it cites. Being one of those sources means existing. Not being one means disappearing, even while ranking first on Google.

The good news: the pages that assistants cite are not chosen at random. They share measurable characteristics, and each engine follows a recognizable logic. That is what measurement makes it possible to map.

How this data was measured

Every week Repliq queries ChatGPT, Google AI Mode, Google AI Overview and Perplexity the way an individual or a company would, in their public interface. The measurements cover 10 sectors of the Swiss economy, in French. For each answer, the cited sites are recorded; the content of the most-cited pages is then retrieved and analyzed (title, length, tables, numbers, dates, author, comments), then compared with the exact text of the answers. The whole setup is available on Repliq's source observatory.

The pages AI cites most in Switzerland

Here, across all sectors, are the most-cited content pages in our sample, with the engine that cites them most.

2 observations stand out. First, no brand appears in this ranking: only comparison sites, blogs, the press and official sites. Second, ChatGPT appears as the dominant engine only once, on Trustpilot reviews. The rest belongs to Google and Perplexity.

What AI looks at in a page

4 characteristics come back on nearly all cited pages, whatever the sector.

1. A title that echoes the question

This is the most consistent regularity. Line up the titles of the cited pages, and the recipe jumps out:

• compassurance.ch: « Top 5 des caisses maladie les plus avantageuses en Suisse en 2026 »
• bonus.ch: « Supermarchés en Suisse : qui est le moins cher ? Comparatif 2025 des prix »
• hellosafe.ch: « Top 10 des banques en ligne en Suisse (Comparatif 2026) »
• mustachianpost.com: « Meilleure banque Suisse (comparatif 2026) »

The pattern repeats: an intent word (best, top, comparison, cheapest), the exact subject, and the year. These titles are a carbon copy of the queries put to the assistants. The assistant looks for a page that answers the intent precisely, and the title is its first landmark.

In practice. Phrase the title like the question your customer types, intent and year included. « Meilleure fiduciaire à Lausanne (comparatif 2026) » speaks to the assistant; « Our accounting services » tells it nothing.

2. Plenty of numeric data

Cited pages do not win through their length, but through their density of numbers. 2 opposite models work equally well.

The in-depth dossier: the compassurance.ch page on health insurers has 9549 words, 2 tables totaling 128 cells, and 489 numbers, including 120 amounts in francs. On the choice of a broker, schweizerfinanzblog.ch lines up 9647 words and 372 numbers.

The compact fact sheet: the bonus.ch price comparison fits in only 1332 words, but packs in 273 table cells, 232 numbers and 86 percentages. Almost no text, almost nothing but data.

The common thread is clear: these pages quantify. A page that asserts without quantifying offers little for an engine whose job is to produce a precise answer.

In practice. Replace assertions with comparative data, presented in tables. A « insurer / premium / gap » table beats a paragraph that calls itself « affordable ».

3. An update date, displayed

All these pages carry the year in the title, and most display a recent update date. mustachianpost.com states at the top « Dernière mise à jour : 10 juin 2026 »; compassurance.ch declares a change on 8 March 2026. Freshness helps without being absolute: hellosafe.ch, last updated in November 2024, is still cited 50 times because it carries authority. But with equivalent content, a page from the current year gets ahead.

In practice. Display a visible update date, at the top of the page, and keep it current.

4. An identified author, a living page

The blogs that outrank the institutions share 2 signals that company fact sheets lack: a named author, and an active comments section. schweizerfinanzblog.ch declares its author in its structured data and displays 12 comments; mustachianpost.com and hellosafe.ch use a dedicated comments platform. A bank describes its products; a blogger compares the market, signs the analysis and answers questions. For an engine that recommends, the second kind of content inspires more trust, regardless of the size of the publisher.

In practice. Sign your content with a real author, tagged in structured data, and open a review or comments space when the format allows.

Which formats get cited

Beyond these 4 characteristics, certain page formats come back constantly. Knowing them means knowing what to produce.

• The comparison: it sets several options side by side on quantified criteria. compassurance.ch, comparis.ch, moneyland.ch live off it.
• The ranking (« Top X »): it prioritizes and decides. hellosafe.ch (« Top 10 des banques en ligne »), compassurance.ch (« Top 5 des caisses maladie »).
• The practical guide: it answers a question step by step (« how to choose »). assurance-genevoise.ch publishes a « Guide 2025 » on health insurance, cited 52 times.
• The quantified press article: a media outlet that releases a data comparison. 24heures.ch on the prices of Aldi, Lidl, Migros and Coop (39 citations).
• The review page: Trustpilot, heavily cited by ChatGPT (20 citations on the CSS page).
• The official fact sheet: the federal premium calculator (priminfo.admin.ch, 75 citations), a cantonal site.

All these formats have one thing in common: they help make a decision. That is the function the assistant is looking for, because it is its own.

What does not get cited

The contrast is instructive. In our sample, some pages almost never appear as a cited source:

• Home pages. They present a brand, not an answer. They are nearly absent from the ranking of cited pages.
• Product pages and « Our services » pages. They describe an offer (« Our supplementary insurance ») instead of comparing the market or answering a question.
• Pages without numbers or dates. A qualitative, undated text with no comparative data gives the engine little to grip.

In other words, the home page that most brands work on the most is precisely the one AI cites the least. The content that gets cited answers a question; the content that gets ignored talks about itself.

Each engine has its own logic

This is the most directly actionable lesson, and the one most content strategies ignore: the assistants do not cite the same types of sources. To see it, let us compare 2 pages from the same sector, health insurance.

ChatGPT: cautious on commercial, comfortable with reviews and official sources

ChatGPT is the assistant most reluctant to pick up comparison content with a commercial purpose: on the mustachianpost.com banking page, it counts only 1 citation out of 92. On the other hand, it comes out on top on reviews (the Trustpilot page for CSS, 20 citations, ahead of all the other engines) and returns to official sites (geneve.ch: 9; priminfo.admin.ch: 4). To exist on ChatGPT, tend to your reputation (Trustpilot, Google reviews) and your official listings, more than your commercial pages.

Google AI Mode and AI Overview: fond of comparison sites and rankings

Google's 2 surfaces are the home turf of comparison sites and blogs. They draw most of their citations from mustachianpost.com (40 + 25), compassurance.ch (33 + 25), priminfo.admin.ch (36 + 13). To exist on Google, you have to be present in your sector's comparisons and rankings, or publish them.

Perplexity: the broadest, blogs and comparisons on top

Perplexity readily picks up specialized blogs and comparison sites, often neck and neck with Google: 26 citations on mustachianpost.com, 21 on moneyland.ch, 19 on bonus.ch, 21 on the 24heures.ch article on supermarket prices. The logic matches Google's: comparative, dated, quantified content.

Case study: why this page is cited 77 times

Let us run the grid over one page, compassurance.ch's « Top 5 des caisses maladie les plus avantageuses en Suisse en 2026 », cited 77 times by the 4 engines.

• Title: an exact copy of the questions (« best health insurer », « top insurers 2026 »), with intent, subject and year.
• Data: 9549 words, 2 tables, 128 cells, 489 numbers including 120 amounts in francs.
• Freshness: modification date of 8 March 2026, declared in the structured data.
• Credibility: complete Article markup, declared author, frequently asked questions section.
• Engines: cited even by ChatGPT (9), because beyond the comparison, the page is a reference on its question.

Every line is checked. The page is built, deliberately or not, to answer exactly as an assistant expects. Worth noting, a strong sign of trust: on official sources, the engine goes as far as copying word for word. An 18-word sentence from the geneve.ch site on transport is found unchanged in the answers, treated as an established fact.

Your action plan

The central lesson is a shift: in answer engines, format takes precedence over brand. Describing your products on your home page is no longer enough. 3 levers stand out.

1. Publish the right format. Produce your own comparisons and rankings, dated, signed, quantified, with a title that echoes the real questions in your sector.
2. Exist on the third-party pages AI already cites. Comparison sites (compassurance, hellosafe, comparis), rankings, specialized press: that is where Google and Perplexity go looking.
3. Tend to what ChatGPT favors. Reviews (Trustpilot, Google) and official listings, often the great forgotten pieces of content strategies.

Your GEO checklist

Before publishing a page you want to see cited, check that:

• The title echoes the question the way your customers ask it, with the year.
• The page compares and quantifies (tables, prices, percentages), instead of asserting.
• A recent update date is displayed, at the top of the page.
• A real author is named and tagged in the structured data.
• The structured data (Article, FAQ) and a frequently asked questions section are in place.
• Your reputation is tended where ChatGPT looks for it: Trustpilot and Google reviews, business listing.
• You are present on the third-party pages (comparison sites, rankings, press) already cited in your sector.

Outlook

This snapshot holds for a given observation window. The cited pages evolve, the assistants adjust their sources, and the hierarchy will shift. Repliq measures these movements continuously and publishes each month the state of the most influential sources on its source observatory, filterable by sector and by assistant.

To know precisely where the assistants go looking in your sector, and which pages to act on, Repliq offers a dedicated AI visibility audit.