Building a Europe AI Can See: Croatia’s National AI Profile

Five homepages, one clean page and four copies of a different one. All five are invisible in the same two ways: no structured identity and no dates.

• National average: 40.4, Grade D. The lowest of the three countries profiled so far. The five sites span only 8 points (38 to 46), and all five sit in Grade D.
• Zero JSON-LD on all five, and almost no dates. Freshness averages 0.8: four sites score 0 and one scores 4. Trust and authorship is 0 on every site.
• Four of five carry competing H1 tags (3 to 4 each, 15 in total). Only the portal has a single H1. Austria and Belgium each had one such site.
• Four ministries share one pattern. Entity clarity is 45 on each, and each ministry’s organization is detected as a domain name (mup.gov.hr, mrosp.gov.hr, zdravlje.gov.hr, mpu.gov.hr) rather than an institution.

The profile in one view

Scores are the Inspector’s dimension scores (0 to 100) for each homepage. Read across a row to see a national pattern; read down a column to see an institution’s character.

Shading: 75 and above strong, 50 to 74 moderate, 20 to 49 weak, below 20 absent. Trust and authorship is the Inspector’s E-E-A-T density score. The portal is gov.hr, the others are mup.gov.hr, mrosp.gov.hr, zdravlje.gov.hr and mpudt.gov.hr.

Croatia against Austria and Belgium

Averages of the five homepages in each country.

Croatia is weakest on structure, schema, freshness and authorship. Austria and Belgium both had near-perfect architecture, so Croatia is the first country where the structure itself is part of the problem.

Where Croatia needs to improve

Ranked by how many institutions are affected and how much one fix would unlock, not by the size of the gap alone.

1. Structured identity (all five)

No Organization, Person or Article schema anywhere, no sameAs links, and entity connectivity and knowledge-graph anchoring are 0/100 on all five. On the four ministries the organization is detected only as a domain name, so a machine learns the URL but not the institution. The highest-leverage fix is one Organization template carrying the full institutional name, an English alternateName and a Wikidata sameAs. If the ministries share a platform, one template change could reach four sites.

2. Dates and accountability (all five)

Freshness runs 0, 4, 0, 0, 0. The portal’s candidate queries include six items announced as “nova usluga” (new service), among them eRehabilitator, the Central Population Register, a cultural card and a state-aid register. “New” is a relative word with no date attached, and it ages by the day. The Interior ministry’s news about 522 new police vehicles has the same problem. Undated announcements read as stale to retrieval engines that weigh recency.

3. One H1 per page (four of five)

The four ministries carry 4, 4, 4 and 3 H1 tags, which holds their structure score at 70 against the portal’s 100. Claude’s score on these four sits 11 to 15 points below the portal’s (57 to 61 against 72), consistent with the engine’s stated sensitivity to heading disorder. This is the fastest fix in the set, a template-level change with no new content needed.

4. Interface text and name fragments in the entity layer (all five)

The four ministry homepages produce 36 candidate queries between them, and apart from the police-vehicle item and a justice-links block they are navigation and accessibility labels. “e-Građani” appears on all four, “Prilagodba pristupačnosti” (accessibility adjustment), “Važne poveznice” (important links) and “Istaknute teme” (featured topics) on three. Pensions, family policy, health policy and justice, the actual subjects, barely register. On the Health site the minister is detected three times (“Ministrica Hrsti”, “Irena Hrsti”, “Irene Hrsti”) with the surname cut at the “ć”.

5. Metadata, alt text and depth (all five)

Meta descriptions are missing on all five, and canonical URLs on three (Interior, Labour and Pensions, Health). A total of 116 images lack alt text across the five homepages, with coverage of 42%, 24%, 6%, 0% and 6%. Gemini is the lowest-scoring engine on all five sites (average 27.6, estimated citation about 9% on every one), and Perplexity is next (33.0). Claude is the highest (61.2, tied with ChatGPT on Health), then ChatGPT (58.2) and Copilot (47.2). Where word counts are reported (Interior 482, Labour and Pensions 520, Justice 546) they fall short of the Inspector’s 600-word target.

Croatia’s homepages tell a machine what to read. They do not tell it who is speaking, or when.

Institution profiles

gov.hr, the national e-government portal – Grade D – 46/100

The best of five, named only “Naslovna”

The gateway to e-Građani is parsed as “Naslovna”, a generic label meaning front page. It is the only site with a single H1 and a structure score of 100, and it has the highest score on four of five engines (38, 66, 72, 29, 58).

Strengths: The richest service catalogue in the set: e-Usluge, e-Informacije, My Profile, life situations, foreigners in Croatia, Croats abroad. Its 23 candidate queries describe what citizens actually do.
What AI sees: No named institution. It detects “Artificial Intelligence” and a truncated phrase, “Pristupite Sredi”, as entities, plus “Conversion Rate” as a topic, none of which describes the portal.
Why it matters: Every citizen service starts here, and the portal serves foreigners and diaspora, yet a “This page is available only in Croatian” notice surfaces among its candidate queries.

mup.gov.hr, the Ministry of the Interior – Grade D – 38/100

A named ministry on a fragmented page

The page parses under its full institutional name, but it carries 4 H1 tags, no canonical URL and 13 images without alt text (24% coverage). It has the lowest entity graph stability in the set (38/100, “unstable”).

Strengths: Full name detected, and the only freshness score above zero (4). Live news shows active publishing.
What AI sees: The organization as “mup.gov.hr”, the domain, with no topic clusters. Candidate queries are led by the 522-police-vehicles handover, then refunds and navigation labels. It has the lowest ChatGPT, Claude and Copilot scores in the set.
Why it matters: Its newest, most citable content is exactly what lacks a machine-readable date.

mrosp.gov.hr, the Ministry of Labour, Pension System, Family and Social Policy – Grade D – 38/100

The pensions ministry AI cannot see

Tied for the lowest score and the weakest alt text in the set: 33 images missing, 6% coverage. It has 4 H1 tags and no canonical URL.

Strengths: Parsed under its full name, with a Programmes and Projects section and a directory (Adresar) among its detected elements.
What AI sees: Only seven candidate queries, all institutional or navigational. None mentions pensions, family or social policy. Co-occurrence is 0/15 and query alignment 0/10.
Why it matters: Pension and family-benefit questions are among the highest-volume citizen queries, and the engine can name only the URL.

zdravlje.gov.hr, the Ministry of Health – Grade D – 40/100

The best entity graph, built from the wrong entities

It has the best extractability (65), the highest semantic richness (52) and the strongest entity graph in the set (65/100, moderate, with co-occurrence 12/15). But it also has 4 H1 tags and 0% alt-text coverage, with all seven images lacking alt.

Strengths: The richest content and the best entity graph of the five, with Claude (61) tied with ChatGPT.
What AI sees: The minister’s name as three entities, “Crne Gore” and “Vlade Pa” as fragments, and the organization as “zdravlje.gov.hr”. Candidate queries are announcements, featured topics and navigation labels.
Why it matters: The most content-rich ministry produces the noisiest entities, and the fix is markup, not more writing.

mpudt.gov.hr, the Ministry of Justice and Administration – Grade D – 40/100

Which ministry is this?

The page parses as “Ministarstvo pravosuđa i uprave”, the detected organization is “mpu.gov.hr”, and the audited domain is mpudt.gov.hr. Three names for one site.

Strengths: The only site above 10 on schema and metadata (20), and the best Copilot (49) and Gemini (30) of the four ministries.
What AI sees: A person-like entity (Damir Habijan, no role attached) and candidate queries such as justice-system links and navigation labels.
Why it matters: A justice ministry whose own page cannot agree on its name is hard for an engine to anchor or cite.

One template fix, five local ones

The same gaps repeat so closely that most of the uplift is a template-level job.

Scope and method

Each institution’s homepage was audited once on 4 October 2026 with the AI Visibility Inspector (v1.9.3) and the Ivica Srncevic Framework. This is a homepage snapshot, not a full-site audit, and the Croatian-language pages were audited, not any English versions. The sample covers five sites, not the whole state, and regional and local government sit outside it. The shared-template reading is an inference from identical scores and repeated labels and should be confirmed in the page source. Engine citation percentages are the tool’s estimates and are best read relatively. The Inspector’s “content gaps” list and commercial labels (such as “Conversion Rate”) use vocabulary that does not suit government sites and were left out of the analysis.

Profile network. For the comparison countries, see National AI profiles of Austria and Belgium.
Also see how Croatian TOP 10 enterprises perform in AI.