How We Optimized Our Own Site for AI Search: A GEO Case Study

When we asked ChatGPT to recommend an agency for migrating to Shopify, it named URich first – in English and in Ukrainian. Over the previous 90 days, Google had sent us exactly 2 clicks from non-brand searches. Those two facts sit side by side in our own data, and together they explain why we stopped treating AI search as “SEO with a new name”.

This is Part 1 of our own GEO case study: the baseline, what we audited, what we changed and the first signals. We will add the results to this article once there is enough data to compare.

We sell AI Search & Agent Readiness (GEO) as a service: auditing and fixing how eCommerce sites read to ChatGPT, Perplexity and AI Overviews. So when our analytics showed real, unprompted traffic already arriving from ChatGPT, we had two options – write another generic “what is GEO” explainer, or run our own audit on ourselves and publish exactly what we found, numbers included. We did the second one.

It helps to know what you are optimizing for. An AI assistant can know about your brand in two ways:

Live retrieval is the part a website can influence. To be cited, a page has to be reachable by the crawler, easy to parse, and contain a sentence the model can lift as a fact – a definition, a number, a clear “who it’s for”. Brand facts also need to agree across the web: if your site, your Clutch profile and your LinkedIn page describe you differently, the model has less reason to trust any of them.

One honest note on llms.txt: it is a 2024 proposal for a plain-text map of a site written for language models. No major AI provider has confirmed that it reads the file, and Google representatives have publicly downplayed it. We treat it as cheap insurance and a useful forcing function for writing clear summaries – not as a ranking factor.

GA4 already groups traffic into an “AI Assistant” channel – most teams never look at it because it’s small. Over the trailing 30 days it was 14 sessions, 100% of them referred from chatgpt.com: a real person asked ChatGPT a question, ChatGPT answered with a link to our site, and they clicked through.

The landing pages told us where ChatGPT already found us worth citing: mostly case studies (/cases/aurora-crystal/, /cases/janado/, /cases/) and a handful of service pages (headless commerce, AI solutions, Shopify development).

Two things surprised us. First, ChatGPT was already sending people to our Ukrainian pages – /uk/services/headless-e-commerce/ and /uk/services/shopify-development/ – within weeks of the Ukrainian version going live. A second language version is a second chance to match how people actually phrase their question. Second, the trend was going down: 14, then 13, then 10 sessions in the trailing 30-day window. Left alone, AI visibility doesn’t grow by itself.

Before changing anything, we pulled 90 days of Search Console data as a baseline. Brand queries (searches containing “urich”) had 1 892 impressions and 36 clicks – a normal branded-search pattern. Non-brand queries had 11 548 impressions and just 2 clicks – a click-through rate of 0.02%.

In plain terms: Google shows our pages for a lot of long-tail queries, but almost nobody clicks, which usually means low average positions on those terms. Yet ChatGPT was sending us real clicks in the same period. GEO and classic SEO are adjacent disciplines, not the same one – a page can be practically invisible in Google’s top results and still be cited by an AI assistant that read it directly.

Classic SEO isn’t standing still either: in the latest 28-day window, “woocommerce to shopify migration” reached 588 impressions at an average position of 18.9. We track both, separately.

Scrolling through long-tail queries, we found searches no human types. One read: “agent monetization, commercial, mofu, primary, ai/product teams, hero h1 directly references this concept” – a fragment of what looks like a content brief. Another: “accuracy of ai seo geo platforms tracking position in ai shopping guides”.

Our working hypothesis: AI tools – research agents, SEO platforms, assistants with web search – are running Google searches on someone’s behalf, and their prompts sometimes leak into the query. We can’t prove who sent them, so treat this as an observation, not a fact. But it is a useful reminder that part of your “search audience” is already software, and it reads your pages the same way a model does.

Before adding anything, we checked what was already in place:

That’s a solid foundation – most sites have none of it. But two concrete gaps stood out once we checked the data instead of assuming: llms.txt listed only 3 of our 28 published case studies, and most service pages opened with marketing copy (“unlock unlimited flexibility…”) rather than a definition an AI system could lift and cite as fact.

Two of the case studies already pulling ChatGPT referral traffic – Aurora Crystal and Janado – weren’t in llms.txt at all. We expanded the file to cover all 28 published cases, one line each, every line carrying a real, already-published metric rather than a generic description:

The logic: an AI system answering “who built a Shopify store for a luxury brand” or “eCommerce agency with a fintech case study” needs a concrete, citable fact to work with – not a paragraph of adjectives.

Marketing copy and citable copy read differently. Our headless commerce page used to open like this:

That’s fine as a pitch, but an AI system can’t lift “unlock unlimited flexibility” as a fact about what headless commerce actually is. We rewrote it as a definition first, positioning second:

We applied the same treatment to the intros of our main service pages.

A simple test we now use on every intro: could this sentence appear, unchanged, in an answer to “what is X and who does it?” If it only makes sense as an ad, it won’t be quoted.

We added a visible FAQ section (plus matching FAQPage schema) to Aurora Crystal, Janado and Oroy specifically – the three case studies GA4 showed were already receiving AI referral clicks – rather than all 28 at once. Each FAQ answers a question grounded in that project’s own published facts (e.g. “Why did checkout abandonment drop by 34%?”), not generic boilerplate. Prioritizing by evidence meant we could ship, verify and measure before committing to the remaining 25.

We set up a standing weekly report (GA4 + Search Console) with a dedicated “AI Assistant referrals” section – which landing pages, which AI sources, and the trend over time. It’s an imperfect proxy: a chatbot can read and answer using our content without anyone ever clicking through, and no analytics tool shows that. A referral click is the floor of AI visibility, not the ceiling – but it’s the part we can actually measure today, so that’s what we track.

After the changes went live, we asked ChatGPT the question a potential client would ask. We asked from our own account, so the answer may be personalized – treat it as a signal, not proof.

What we read in this answer:

We can’t attribute this answer to any single change – it may draw on older data too. That is exactly why Part 2 will compare numbers, not screenshots.

Our site sells services, so case studies and service pages were the assets that mattered. For an eCommerce store, the same principles land on different pages:

The ten checks we run first – you can do most of them in an afternoon:

In a few weeks we will update this article with the first results: AI-assistant sessions and landing pages, branded search as a proxy for zero-click recommendations, and a repeat of the ChatGPT test – plus Perplexity and Google AI Mode. If the case-study FAQs show a measurable effect, we will roll them out to the remaining 25 cases and report that too.

If you want the same audit run on your own store – what AI assistants can currently read, trust and cite about your brand – that’s exactly what our AI Search & Agent Readiness service does. Book a free consultation and we’ll show you your own numbers first, the same way we started here.