GEO and AI Visibility Agencies in Hungary and Central Europe: Who Do AI Answer Engines Actually Name? (Measured, Q3 2026)

The unit is a (prompt, engine) detection — in how many answers a name occurred on the 31-prompt panel, on this one day. This is not market share; the panel asks about the market, but it is not the market.

The table contains only agencies whose existence we verified against their own websites. Further company names occurred in the answers that we could not verify — including at least one with no operating company behind it (see "Engines sometimes name companies that do not exist").

To the question "which agency does AI visibility (GEO) in Hungary?", the four engines returned four different lists: twenty different company names in total, and not one player named by all four. Only two names appeared on more than two engines: 'Vantgard Digital (three engines) and HD marketing (two). The verified Hungarian mentions: Marketing21, 'Vantgard Digital, HD marketing, CRS AI Marketing, BP Digital, Contented, Tartalom Design.

The practical lesson: whoever checks their visibility on a single engine is not seeing "the market" — they are seeing one engine's momentary reading of it. The same spread shows up in whether engines are willing to name anyone at all: on category questions, Gemini named at least one verified player in 10 of 14 answers, AI Overviews in 9 of 13, Claude in 8 of 14 — ChatGPT in 2 of 14.

The 13 measured Google AI Overviews answers produced 92 citations across 68 different domains — 55 domains appear exactly once. And within the Hungarian-language category answers a clear pattern emerges: a large share of the citations are the named companies' own websites (marketing21.hu, hdmarketing.hu, vantgarddigital.hu, tartalomdesign.hu, seoklikk.hu) — and this pattern holds across all four engines on the Hungarian question.

Where English-language answers are backed by ranking listicles and aggregators (on the French market, a single ranking page made it into the sources of two different questions), the Hungarian answers have no such layer behind them: on the Hungarian market today, a company's own machine-readable website is the citation source. The single most-cited source in the entire measurement, incidentally, was not an industry site but youtube.com (9 citations).

The new category's most common abbreviation — GEO — collides with older, established meanings in several languages: a magazine brand and geomarketing in Germany and France, land surveying in Poland, and in English business language also the Global Employment Organization. The measurement quantified this: on the bare-acronym "GEO" questions in German, French, and Polish, 3 of 12 answers were about a different industry entirely (geoinformatics, geomarketing, surveying) — all three from the same engine. And on the Hungarian–Central-European question, a different engine's answer largely listed satellite and mapping companies instead of agencies.

The competition for the category is therefore partly a competition over what the category's name means — and any measurement built on a single engine inherits this distortion too.

While assembling the measurement panel, a publicly available ranked answer carried a non-existent agency in 2nd place. We excluded it; the exclusion is documented. In the present measurement's answers, further company names occurred for which our verification found no operating company. This is why we publish only source-verified names: AI answer engines sometimes hallucinate companies, and only verification filters this out. If you choose a provider from an AI answer, verify that the company exists — unfortunately, that sentence is not a joke.

The measurement also included 13 term-definition questions on the new category's expressions — such as machine equity (how well machines know and recommend a brand), Model Mindshare (who an engine names unprompted), KEO (knowledge engine optimization: presence in the models' trained knowledge), and AI visibility engineering. Of the three language models' 39 answers, 28 gave a substantive definition — and zero attributed the term to anyone. The picture is consistent with the confusion around the category's name: the engines already use this new field's words, but the category is not yet anchored inside them — no stable name, no named source. This will be one of this report series' time series: with documented first-use terms, it is measurable when source attribution first appears, and on which engine.

We measured a single day — no number here is a trend. Due to search-quota exhaustion, only the category questions are available from AI Overviews. Detection counts are not market share. The English prompts ran with a fixed Germany location; the numbers are valid for that surface. The measurement table knows 34 verified players — whoever it does not know, the roll-up does not count; such names are preserved raw. And we claim no causality: we publish dated observations.

31 fixed prompts (category, term, and control questions) in five languages with fixed locations · four engines with pinned model versions (gpt-4o, gemini-2.5-flash, claude-sonnet-4-5, Google AI Overviews via the SerpApi surface) · 106 of 124 calls successful · every raw answer stored unchanged; every quote in the processing must be locatable verbatim in the raw answer, otherwise the system rejects it (11 such rejections in this measurement) · all 34 players in the table verified against their own websites · the measurement's documented total cost: USD 2.33. The next edition runs on the same fixed panel, in comparable form.