We asked four AI assistants for five agencies. We got 19 names and not one in common.

Why we measure every month

"AI visibility" means nothing until you say which assistant, which question and what kind of recommendation. That is why every month we put the same set of Croatian questions to four assistants and compare what they say, whom they mention and whether they are right.

The September edition is the third, after July and August 2026. With three data points we can see a direction for the first time, not just a snapshot.

Six key findings

1. There is no single AI visibility

Asked "Recommend five digital agencies in Croatia for B2B lead generation and rank them" (in Croatian), ChatGPT named GoTraffic, Digital Matrix, Euro Growth Agency, Results Maker and Protagonist. Perplexity named Peakflow, OB2B, Logit and Arbona. In total we recorded at least 19 names, and the closest to a shared choice is Digital Matrix, which appeared with three assistants.

2. Some assistants are consistent, others change their answer

ChatGPT answered the agency question with the same five names, in the same order, in both runs. Gemini kept only two of its five names between two attempts.

3. Global software dominates, regional software stays in the background

In open questions about CRM, ERP and business software (56 answers), HubSpot appeared in 25 answers, Zoho in 23 and Microsoft Dynamics 365 in 19. PANTHEON and Minimax, solutions tailored to the Croatian market, appeared in 9 and 8.

4. Local knowledge is the biggest divider

How well answers are adapted to Croatia: Claude 63.4, Perplexity 60.3, ChatGPT 56.9 and Gemini 22.0 out of 100. On locality, Gemini scores two and a half to three times lower than the others.

5. Prices remain a weak spot

Of the 50 fuel price claims we could verify, 16 were wrong (32%). In August the share was 43%, and in July 26%.

6. AI reads the Croatian web, and Reddit

Croatian domains make up 54 to 62% of cited sources. The most cited are index.hr (93 citations), Reddit (88) and kompare.hr (81). Reddit is cited by Perplexity (66 times) and Gemini (22).

How the platforms differ

The platform ranking is unchanged for the third month in a row, and all four score 1 to 2 points lower than in July.

Topics: where AI knows Croatia best and worst

Tourism scores highest and the shopping basket lowest. Tourism is slipping as the season ends, while fuel is recovering after its August dip.

Tourism: Dubrovnik and Split still lead

In open tourism questions (56 answers), the most mentioned places are Dubrovnik (29 answers) and Split (27), followed by Zagreb (18), Rovinj (17) and Šibenik (15). Even when you ask for lesser-known places, the answer usually starts with Split and Dubrovnik as a point of comparison.

Shopping basket: big chains in, small ones rarely

In open questions about affordable grocery shopping (56 answers), Konzum and Lidl were mentioned in 19 answers each, Kaufland in 18, and Plodine and SPAR in 15 each. Studenac and Tommy appear in only 6.

Fuel: errors in price questions, not in trip estimates

Incorrect price claims are concentrated in questions about fuel and mobility, where 13 of 29 claims were wrong. In car trip plans AI was more accurate: 3 of 21 claims were wrong.

What this means for your brand

If one AI assistant recommends you, that does not mean the others see you. Five steps that follow directly from this measurement:

• Measure each platform separately. For B2B lead generation agencies, no single name appeared on all four platforms. An average "AI visibility" score hides exactly that difference.
• Measure more than once. Gemini kept only two of five names across two attempts at the same question. One screenshot is not a finding.
• Spell out your local advantage. PANTHEON (9 of 56 answers) and Minimax (8) trail HubSpot (25) despite being tailored to the Croatian market. That advantage needs to be stated clearly in the places AI reads.
• Know where AI gets its information. In Croatia that means index.hr, kompare.hr and tportal.hr first, but also Reddit. If you are not there, it is harder to make it into the answer.
• Keep prices and terms current and public. 32% of fuel price claims were wrong. If AI has no fresh, clear source for your prices, it will guess.

How we measured

Answers were collected on 9 September 2026. We asked 84 questions in Croatian across 10 topics: six consumer topics, three business topics and the labour market. Each question was asked twice on four platforms, via API with web search enabled: ChatGPT (GPT-4.1 mini), Claude (Sonnet 4.6), Gemini (2.5 Flash) and Perplexity (sonar). All 672 answers were collected successfully.

We scored each answer with the AI Reality Score from 0 to 100 across six dimensions: facts, locality, freshness, sources, adherence to the given constraints and stability. Fuel price claims were compared with the European Commission's Weekly Oil Bulletin, with a tolerance of EUR 0.06 per litre. September is compared with July (957 answers, 14 July) and August (672 answers, 26 and 27 August).

The figures should be read with a few caveats:

• One day, not a permanent ranking. Answers change with the model, the session and the wording of the question.
• A mention is not a recommendation. We count names that appear in an answer. Split, for example, also appears in answers about lesser-known destinations, but as a comparison.
• Visibility is not market share. It only shows how often a name appears in answers to the index questions.
• Questions that name a brand themselves are excluded from visibility counts. Example: "Compare Lidl, Kaufland, Konzum, SPAR and Plodine".
• API, not the app. Answers in consumer apps may differ, especially for ChatGPT and Gemini.
• Gemini sources were reconstructed from citation titles, because Gemini returns a redirect instead of the original link.