Fewer inquiries are coming in.
Yet website traffic is the same, your Google rankings haven’t dropped, and campaigns are running as before.
The explanation you won’t find in any monthly report may be this: the buyer put together their shortlist without you.
You never find out, because it happened months before anyone would have picked up the phone.
The decision about who is in the running happens before first contact
International research on B2B buying behaviour (US and UK data, summarised in a Search Engine Land analysis of the changing buyer journey) shows two shifts at once.
First: 51% of B2B buyers start vendor research in an AI tool, not a search engine. They ask ChatGPT, Perplexity or Gemini who knows their stuff in a given category and treat the answer as the first filter. (G2, The Answer Economy, 2026)
Second: the average shortlist has shrunk from 3.2 to 2.5 vendors. Fewer names make the cut because the model did the first round of screening for the buyer. (Apollo, 2026, cez Geisheker Group)
Both figures come from US and UK samples. A comparable measurement for the Slovak market doesn’t exist yet, and we won’t guess at one. We do see the same direction here, though. In the GEO projects we run, we hear a question more and more often that nobody asked two years ago: “A customer told us they found us through ChatGPT. What does that mean for us?” This is an observation from our own work, not a measured sample.
The maths of a shorter shortlist is uncomfortably simple. Three places on average have become two and a half. Whoever used to be third is often out of the running today. And because the buyer never contacted them, the third vendor never finds out. Lost demand doesn’t look like a lost tender. It looks like a phone call that never came.
What happens next matters. Buyers don’t treat the model’s answer as a final decision. They take two or three names from it and check them the traditional way: they open the website, look at case studies, click through LinkedIn profiles, ask in an industry group. A mention in an AI answer doesn’t replace the sales process, but it decides who enters it at all. It’s a ticket in, not a sale. That’s exactly why its absence is so hard to spot: everything else in the company works as usual, only fewer people reach the start of the sales process.
Why a standard SEO report won’t catch this shift
An SEO report answers the question of where you rank in Google for a specific keyword. The question that decides who ends up on the shortlist today is a different one: does the model mention you when someone asks it about vendors in your category?
These are two separate measurements because two different mechanisms sit behind them. Google ranks ten links and leaves the choice to the user. A language model generates a single answer with two or three names, and the rest of the market simply isn’t in it. AI search in B2B doesn’t copy the order of organic results. Ranking first in Google doesn’t guarantee a mention in the answer, and a company below the top spot can still appear in it.
We saw this on our own website with server-side tracking. In regular Google results we’re not in first place, yet the AI answer lists us among the recommended agencies. Headless CMS is similar: Google AI places us among agencies that understand the technology. In both cases, what made the difference was how our expertise is defined in the code, not the number of articles we’d written on the topic.
This difference has an uncomfortable side. Model answers stabilise over time and reinforce themselves. Once a model starts listing you in a category, the mention repeats across related questions, and more context about you builds up in both training and search data. If you’re absent from the category, nothing happens. Competitors’ lead doesn’t grow linearly, and the longer you put off measuring, the more it costs to catch up.
So if your monthly report covers rankings, traffic and conversions but has no line on how often AI tools mention you, you’re seeing only part of the picture. Visibility in ChatGPT is a metric of its own today and needs to be measured separately.
What a company can actually do about it
In GEO and AI visibility projects we follow one methodology and assign every activity to one of four areas. If an activity doesn’t fit any of them, we don’t do it. All four areas run alongside traditional SEO, never instead of it. Google rankings remain a source of inquiries and one of the signals models draw on.
- Measurement. We start by putting together a set of questions your buyer would actually ask and measure how often, and in what context, models mention you. The set stays fixed throughout the project, otherwise numbers from different months couldn’t be compared. Without this baseline there’s no way to tell whether things are getting better or worse.
- Technical foundation. A model has to be able to read your website and understand who you are. We check that robots.txt doesn’t block AI crawlers and that content is available in the initial HTML, so the crawler doesn’t have to assemble it from JavaScript. We also verify that JSON-LD schemas describe the company, services and authors as connected entities. llms.txt belongs here too, but it’s only a supplement and won’t secure citations on its own. Why AI crawlers only see the initial HTML, and how to check this on your own site, is covered in AI Only Reads What It Can See on Your Website..
- Content worth citing. For every important buyer question, your website should have one page that answers it best. The answer goes at the top, key sentences make sense when lifted out of context, and claims rest on a number or an example. Tables and comparisons get cited more easily than long paragraphs.
- Consistency and mentions beyond your website. Models don’t rely only on what a company says about itself on its own site. They look for the same information in industry media, directories, reviews and on LinkedIn. Your name, services, location and case studies need to match everywhere. When measuring the construction segment, we saw that three of the most cited sources were industry portals, not manufacturers.
The order is deliberate. Measurement comes first because you can’t evaluate the other three areas without it. The technical foundation comes before content, because even the best text won’t get cited if the crawler can’t reach it.
Timing is part of setting expectations. Based on what we observe in our own projects, the first changes in answers appear within weeks of deploying the technical layer, but the full effect takes months. The range is wide and depends on the category and on how many external mentions of the company exist. Anyone who promises you a specific date is guessing. What is predictable: without defining your expertise in code, nothing moves, even if you publish an article every week. If you want to check where you stand before tackling the technical side, start with an AI visibility test.
How to check it within a week
You can get a basic picture yourself, with no tool and no budget. It takes about five evenings of work.
- Write down 15 buyer questions. The questions are about your category, not your brand. For example, “who builds custom B2B portals in Central Europe” or “which agency handles server-side tracking”.
- Ask them in three tools. ChatGPT, Perplexity and Google AI Overviews. Each question once, with no follow-up prompts.
- Record who the models name. A simple table is enough: question, tool, companies mentioned, order, whether you appeared. This shows you not only how AI finds you, but also who it recommends instead of you.
- Compare it with your SEO report. Mark the questions where you’re strong in Google but missing from the AI answer. Those are your cheapest opportunities.
- Check the technical foundation. Does your website have JSON-LD schemas for the organisation and its services? Does the crawler see content in the initial HTML? Does the company description on your website match the one on LinkedIn? A developer can answer this in a few minutes. You’ll find a simple initial-HTML test in AI Only Reads What It Can See on Your Website..
After a week you’ll have one table and one number: in how many of the fifteen answers you appeared. Take that number to your next management meeting. It answers “how are we doing in AI search” more precisely than any estimate.
What surprised us
When we started doing GEO on our own website, we expected content to make the difference. It turned out the biggest shift came from the technical layer that no reader ever sees. We’d had articles published for years, but models only started noticing us once we defined our expertise precisely in code and aligned it with what we say about ourselves elsewhere.
The second surprise was how fast the topic moved from “that’s interesting” to “this is missing from our report right now”. A year ago it was a discussion point. Today it’s a slot on the shortlist that either you take or a competitor does.
If you want to know exactly where you stand, we’ll run an AI visibility test for you. We’ll measure which questions in your category the models mention you in and who they list instead. Then we’ll name what your website is missing to change that. You’ll get a table and a list of actions, not a slide deck. The full process is described on our AI visibility and GEO page.
