GSO Seminar Case Study: Measuring AI Search Visibility

0.03 percent: That was the share of traffic from ChatGPT, Gemini and Co. on the website of an independent wealth manager when we measured it properly for the first time in February 2026. This case study describes a two-day one-on-one seminar "Generative Search Optimization (GSO)" that I ran for GFU Cyrus AG: from the SEO foundation and the mechanics of AI search engines to a prompt tracking that the participant now continues on his own. Company, competitors and URLs are anonymized.

Starting point: falling clicks, relaunch in summer

Google Search Console showed the picture I have seen on many finance and B2B websites since 2025: impressions stable, clicks 20 to 50 percent below the previous year depending on the topic. Part of it was an earlier relaunch, the larger part is AI Overviews. An Ahrefs study of 300,000 keywords puts the click loss for queries with an AI Overview at 44 percent, and exactly that pattern was visible in the participant's data.

On top of that, a relaunch was scheduled for the summer, implemented by an agency. The participant's question on day two summed it up: "From a technical perspective, is there anything else to consider for LLMs beyond the normal SEO items?" The answer was the real result of the seminar.

Day 1: Foundation and mechanics of AI search

SEO remains the foundation

The morning started with the pyramid I show in every GSO seminar: at the bottom technical SEO and on-page, above it citable content, above that brand visibility on trusted platforms, at the top sentiment, meaning how an AI talks about the brand. Anyone who does traditional SEO properly has already covered 60 to 80 percent of the optimization for language models. There is no sharp line between SEO and GSO, only a gradual transition.

Two numbers shaped the discussion: ChatGPT now ranks behind Instagram, Facebook, YouTube and Google in usage, but ahead of Amazon and Reddit. And 95 percent of people who use ChatGPT still use Google. AI search does not replace Google, it adds a step to the research in which the brand has to show up.

Query fan-out, grounding and entities

The afternoon went into the technology. An AI search engine breaks a question into several sub-queries, the query fan-out, sends them to its index and checks via embeddings which content fits best. If the similarity is below a threshold, it falls back to web search. For the participant this meant: an article has to serve the questions an AI splits the topic into, which is more than one keyword. With a fan-out tool we checked an existing guide: 67 percent overlap with the generated sub-questions, and the missing questions became the revision list.

Second building block: entities and grounding. An AI must recognize the brand as an entity, with consistent information on services, location, people and figures. That is what the grounding page concept is for, a facts page linked in the footer that gives the AI a reliable source about the company. For a wealth manager whose website contained few structured facts so far, this became the first concrete task for the relaunch.

Customer journey and motives

Day 1 closed with the customer journey in five phases, from awareness through information and evaluation to purchase and loyalty, combined with the motive compass to describe target groups. This matters for content planning because AI users ask different questions in every phase: very generic ones at the start, very specific ones about fees, strategy changes or minimum investments at the end. The participant mapped his existing articles to the phases and saw that almost everything sat in the information phase.

Day 2: Measure, track, prioritize

The AI traffic channel: 0.03 percent as the baseline

Day 2 started with measurement. In GA4 we created a dedicated LLM channel that bundles referrals from ChatGPT, Gemini, Perplexity and Copilot. The result: 0.03 percent of traffic. Very little, but visible for the first time, and therefore a baseline against which every measure can be assessed. In Bing Webmaster Tools the "AI Performance" view also showed 33 pages already cited by Copilot. The question "Which ones are they?" was the entry into prioritization.

A side aspect that is often underestimated: Search Console anonymizes 30 to 40 percent of queries, the BigQuery export only 10 to 20 percent, and it goes back further than 16 months. For a company that wants to trace click losses over years, the export is mandatory.

Brand prompts instead of keywords

The biggest difference to a classic SEO seminar was the tracking. Instead of keyword rankings we defined prompts: branded prompts that contain the brand name, and generic prompts along the journey, for example on investment strategies, custody fees or retirement planning. My recommendation: work out 50 to 100 prompts and check weekly whether the brand appears in the answers, with which sentiment and which sources are cited. The participant had the prompts pre-generated by ChatGPT, cleaned them up and stored them as a tracking set. In addition we used the Brand Radar to check how the AI describes the brand compared to competitors.

Content score, freshness and authorship

In the afternoon we worked on a specific article about gold as an investment. A content tool scored it above 90 points, but with too many repetitions of the main terms. My advice: aim for 80 to 90 points, reduce repetitions, write an answer-first opening that answers the question in the first paragraph. Content for LLMs is 60 to 80 percent content for readers; walls of text hurt both.

Two points surprised the participant. First, freshness: the team had deliberately removed article dates to appear timeless. For Google and for AI systems, a visible update date is a relevance signal. Second, authorship: in finance, who writes matters. Author profiles with qualifications, mentions in trade media and quotes with numbers matter more for citability than yet another guide.

Relaunch checklist for LLMs

Back to the question from the beginning: what does the relaunch need to consider for LLMs? The honest answer: there is no separate technical checklist for it, there is the SEO checklist plus four additions. A grounding page with structured facts, visible authors and dates, an answer-first structure in the guides and clean citability through sources and figures. The participant took the checklist for the technical part with him for the acceptance with the agency.

Learnings for your GSO seminar

• Measure first, then optimize. Without an LLM channel in GA4, every GSO discussion remains opinion.
• Prompts are the new keywords. 50 to 100 prompts along the journey, checked weekly, do not replace a ranking tool, but they show whether the brand appears in answers.
• Fan-out instead of keyword density. An article has to cover the sub-questions an AI splits the topic into.
• Grounding page before the relaunch. Structured facts about the company are the cheapest GSO measure.
• Make dates and authors visible. Looking timeless costs relevance, in finance twice over.

Conclusion

After two days the participant had three things that were missing before: a metric for AI traffic, a prompt set for tracking and a concrete list for the relaunch. The share of 0.03 percent will not rise overnight. But it is visible now, and every measure can be assessed against it. That is exactly what separates Generative Search Optimization from another buzzword.

More reports from seminars and workshops are in my AI case studies, the topic in detail on the page about GEO and SEO.