GEO / AEO Agency Zurich — Get Cited by ChatGPT & Perplexity in Swiss Finance

#GEO/AEO in Zurich: Why AI Search Visibility Matters for Swiss Financial Institutions

Zurich's enterprise decision-makers have changed how they research vendors and solutions. Assistants such as ChatGPT, Perplexity and Google AI Overviews now sit in front of the results page for a growing share of first-stage vendor research in Swiss financial services. When a CDO at UBS searches for the best FINMA-compliant AI vendor for claims automation, the answer is delivered directly by an AI system — no click required, just a citation. If your institution is not cited in those responses, you are invisible in this rapidly growing research channel.

The Zurich market adds two specific layers of complexity to GEO/AEO strategy. First, language: English is the primary B2B language for most enterprise interactions and international communications, but German is essential for domestic queries, Swiss media citations, and FINMA regulatory content. A single-language GEO/AEO strategy misses half the addressable query volume. Second, compliance: advertising rules for financial services restrict how capability claims can be made publicly. GEO/AEO content for regulated institutions must be factual, sourced, and compliant — which happens to align precisely with what generative systems reward when they choose a source to cite.

#What actually determines whether you get cited

It helps to be exact about the mechanism, because most of the advice circulating on this subject is search-engine folklore with new vocabulary. When a user asks a grounded assistant a question, the system rarely searches once. It expands the question into several sub-queries, retrieves candidate passages for each, and then synthesises an answer from the passages it kept. The citation is attached at synthesis time to the passage that supplied the claim. This has three consequences that change how you write.

Selection happens at passage level, not page level. A high-authority domain whose relevant answer is buried in the eleventh paragraph of a 3,000-word thought-leadership essay loses to a mid-authority page that answers the sub-query in a self-contained 80-word block with a date and a source. Domain authority still influences which candidates get retrieved; it does not decide which sentence gets quoted.

Specificity beats fluency. The passages that survive synthesis are the ones carrying a checkable object: a named regulation and its article, a threshold, a defined term, a date of entry into force. Marketing prose describing capability in the abstract contributes nothing an assistant can attribute, so it is discarded even when the page ranks well in classic search.

And the answer is assembled from whatever the retriever found in that moment, which makes the output non-deterministic. Two identical prompts an hour apart can cite different sources. Any measurement approach that treats a single observation as a rank is measuring noise.

#Why Swiss financial institutions are systematically under-cited

Zurich's finance sector has a publishing culture that generative retrieval handles badly, and this is the opportunity. Substantive material — regulatory position papers, economic research, risk methodology notes — is published as PDF, frequently behind a registration form. Retrieval pipelines parse HTML reliably and PDFs unevenly; anything behind a form is simply absent. Meanwhile the HTML that is public tends to be brand narrative, which contains nothing quotable.

Second, much of the genuinely authoritative Swiss content exists only in German while a large share of the high-value queries in this sector are typed in English, often by people sitting in London, Singapore or New York evaluating a Swiss counterparty. Third — and we find this in roughly half of first audits — the robots directives on the corporate domain block or partially block the crawlers that feed these systems. Teams frequently disallow GPTBot, ClaudeBot or PerplexityBot as a reflex against training-data scraping, without distinguishing the training crawler from the retrieval crawler that fetches a page at answer time. Blocking the latter guarantees exclusion. There is a legitimate debate to have about the former, and it should be a decision, not an inherited default from a template robots file.

#The Work: How We Build AI Visibility for a Zurich Institution

#1. Prompt panel and citation baseline

We do not audit keywords; we audit questions. We build a panel of the prompts your buyers actually type — vendor selection ("which providers do FINMA-compliant claims automation in Switzerland"), technical evaluation ("can a Swiss bank run an LLM without data leaving the country"), regulatory research ("what does the FINMA circular on outsourcing require of a cloud AI provider"), and their German equivalents. Each prompt is run repeatedly across ChatGPT, Perplexity, Google AI Overviews and Copilot to average out non-determinism, and we record who gets cited, which URL, and what the cited passage says. The output is a citation share table against the competitors who are actually appearing — usually not the ones the marketing team names.

#2. Retrieval readiness

An unglamorous engineering pass that decides whether anything downstream works. We reconcile the robots policy crawler by crawler, distinguishing training access from answer-time fetching and documenting the choice for whoever will be asked about it later. We check that the pages carrying your substance render server-side rather than requiring JavaScript execution that assistant fetchers may not perform. We convert the PDF corpus that matters into indexable HTML with stable URLs, keeping the PDF as the download. We put dates on everything — published and updated — because recency is a live signal in source selection. Where it fits, we publish an llms.txt inventory pointing to canonical sources, while being straightforward with clients that this is an emerging convention with uneven support rather than a standard anyone must honour.

#3. Answer-shaped content architecture

We restructure the site around questions rather than services. Each target question gets a self-contained answer block at the top of a page: a direct response in two or three sentences, then the qualification, then the evidence. Definitional pages for the terms your market argues about. Comparison pages that state trade-offs honestly, because assistants asked to compare will cite whoever wrote the comparison, and a page that only says you win is filtered as promotional. Genuinely useful regulatory explainers — what the revised Data Protection Act changed for automated decisions, what the DLT Act did and did not do for token custody — that a compliance officer would forward to a colleague. This is where regulated-sector constraints turn into an advantage: content you can defend to your legal department is content built out of verifiable statements, which is exactly the profile that gets quoted.

#4. Entity grounding

Assistants resolve organisations as entities before they trust claims about them. Swiss firms have unusually good raw material for this and rarely use it. We make your legal identity consistent and machine-readable across the commercial register entry and UID number, your Wikidata item, professional profiles, industry association listings and your own Organization markup with correct sameAs links. We give your named experts real author pages with credentials, publications and a stable identifier, and we attach Person markup to what they write. When a system is asked "who in Switzerland is credible on AI governance in insurance", it answers with entities it can resolve — an unnamed "our team of experts" resolves to nothing.

#5. The German track, done properly

German is not a translation exercise here. Swiss written German drops the ß, uses Swiss legal vocabulary (Geldwäschereigesetz, Sorgfaltspflichten, Vermögensverwalter, Datenschutzgesetz rather than the German or EU equivalents), and refers to Swiss instruments a de-DE writer will silently substitute with German ones. We produce the German track natively with Swiss writers and mark it up as de-CH with correct hreflang, because a page that mislabels its locale competes in the wrong retrieval pool. In our experience the sequencing that works is English first — the international query volume around Swiss financial topics is larger and the competitive field thinner — with German built out from month two, prioritised on regulatory and cantonal-market questions where the domestic corpus is strongest.

#6. Third-party corroboration

Assistants weight corroboration across independent sources, so a claim that exists only on your own domain is structurally weaker than one repeated by a publication the model already trusts. In this market that means the Swiss business and financial press — NZZ, Finanz und Wirtschaft, Handelszeitung — sector titles covering Swiss IT and fintech, industry association publications, and the Crypto Valley ecosystem for digital-asset topics. Contributed expert commentary, published research your people can be quoted on, and conference material that leaves an indexable trace all compound. This is slow work with a long half-life, which is the opposite of the paid-media rhythm most marketing teams are used to.

#Measuring a Channel With No Rankings

There is no rank tracker for this, and vendors who imply otherwise are selling a single sample dressed as a position. We measure four things instead. Citation share across the prompt panel, computed over repeated runs so the number reflects a distribution rather than one lucky answer. Passage coverage — what proportion of your priority questions have a page that answers them in a citable form at all, which is the leading indicator that moves first. Referral behaviour, acknowledging honestly that assistant referrals are systematically undercounted: many surfaces pass no referrer, and a substantial share of users read the answer, then type your name into a search engine. Which is why the fourth metric is branded query lift in Search Console, usually the cleanest evidence that the answer layer is doing work.

Timelines follow from the mechanism. Retrieval-based surfaces reflect a newly published, well-structured page within weeks, so the first citations on English-language questions typically appear early. Entity-level effects — being named as a category expert rather than quoted on one narrow question — accrue over quarters, because they depend on corroboration accumulating across sources you do not control.

#The Five Arguments We Have With Zurich Clients

"This is just SEO with a new name." It overlaps, and a site with broken indexation will fail at both. But classic SEO optimises a page to win a click, and this optimises a passage to be quoted in an answer the user may never leave. The unit of work is different, and so is the writing.

"Compliance will never approve it." Compliance objects to unsubstantiated capability claims, comparative superlatives and implied guarantees. None of those get cited anyway. Sourced, dated, specific content passes financial-services advertising review more easily than the marketing copy it replaces, and it is precisely what the retrieval layer selects.

"We should block the AI crawlers." Then you have chosen not to appear. That is a defensible position for some content and an accident for most. The decision should be taken per crawler and per content class, in writing, by someone who understands both sides.

"German is handled — we translate." Machine-translated German pages compete against natively written Swiss content on domestic regulatory questions and lose on terminology alone. Worse, an incorrect locale declaration puts the page in the wrong pool entirely.

"Let's do this after the redesign." Redesigns routinely break the things that matter here — URL stability, structured data, dates, hreflang. Doing the retrieval-readiness work as part of the redesign costs a fraction of doing it twice.

#What Clients Get Out of It

The honest framing is that this channel is small today in absolute traffic and disproportionately important in influence, because it reaches buyers at the moment the shortlist is formed. The compounding asset is a body of answer-shaped, sourced content and a resolved entity — both of which continue paying in classic search, in sales enablement and in analyst briefings regardless of how the assistant market consolidates. We fix the measurement baseline before publishing anything and report against the same prompt panel each quarter, so the trend is comparable rather than re-argued.

#Why Nehos for GEO/AEO in Zurich

• 47 specialists in AI, data and cloud, and 200+ projects in production since 2014 — we build the retrieval systems on the other side of this problem, which is why our advice about how citation selection works is engineering rather than inference.
• Regulated-sector fluency. We write content that survives financial-services advertising review because we understand what the review is actually protecting against.
• Native English and Swiss German production, not translation, with correct locale handling and hreflang.
• Technical execution included. Structured data, rendering, robots policy, PDF-to-HTML migration and Core Web Vitals work are part of the engagement, not a list of recommendations handed to your IT team.
• Measurement you can audit — a versioned prompt panel, repeated sampling, and reporting that distinguishes signal from a single lucky answer.
• A free 30-minute audit: bring five questions your buyers ask and we will show you who is being cited today.

Related areas we cover: AI visibility for Zurich and Switzerland, generative engine optimisation for corporate banking, answer engine optimisation for fintech and digital assets, ChatGPT Search visibility for Swiss B2B, Perplexity optimisation in Zurich finance, and generative engine optimisation for the wider Swiss market.

Continue with the Nehos Zurich hub overview, the full method on the GEO/AEO service — full methodology page, and the French-speaking equivalent at GEO/AEO Geneva — French-speaking Switzerland.

The questions below are the ones Swiss marketing and compliance teams raise most often before starting.