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

Sweden’s five audited pages average 68.8 on the AI Retrieval Index, the highest of the six countries compared in this series so far. Norway, averaged 58.0, Hungary 55.8, Ireland 51.8, Germany 51.8 and Greece 49.4.

This profile differs from the earlier ones in one way. The five audited pages are all on one site, government.se, the English-language site of the Government Offices: the front page and four ministries (Finance, Justice, Climate and Enterprise, and Health and Social Affairs). When someone asks an AI assistant what the Swedish government has decided, which ministry is responsible for a policy area or where a national strategy stands, this is the site the answer should come from. Because the pages share one publishing template, the result describes how that template reads to AI more than it describes five separate institutions.

The audits measure three things about each page: whether an AI system can read it, whether it can tell who published it, and whether it can judge that the content is current. All five audits were generated on 13 October 2026.

The main finding: Swedish government pages are the best built in the series and still declare almost nothing about the people and bodies behind them. Structure scores 100 on all five pages, extractability averages 83.0 and freshness 65.0. Four of the five pages carry no JSON-LD. Entity connectivity, KG anchoring and E-E-A-T density are 0 on every page, including the front page that has a valid structured data block.

This research is part of the Building A Europe AI Can See series, which also includes National AI profiles of Austria, Belgium, Croatia, Cyprus, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Norway, Sweden.

Read also Sweden Flagship companies research.

The Profile in One View

The spread between the best and worst page is 6 points, against 27 for Norway and 19 for Ireland. All five pages grade C, with the same sub-grades of A, B and D. Three ministries share the same index of 68.

Sweden leads on four of five dimensions, and the biggest gap is freshness. Against Norway, Sweden has 33.2 points more freshness, 11.6 more extractability, 8.0 more structure and 7.0 more schema. It is 3.0 points behind on Entity Clarity, where Norway (52.0) and Hungary, Greece and Germany (all 55.0) lead it. Every Swedish page scores 65 on freshness.

Other indicators across the five pages:

• Entity graph stability averages 72.2, and every page reads “moderate”. Scores run from 69 (Justice, Climate and Enterprise) to 76 (front page). Norway averaged 44.0.
• Entity connectivity is 0 of 100 on all five pages. That includes the front page, which has a valid structured data block.
• KG anchoring and E-E-A-T density are 0 of 100 on all five pages.
• Four of five pages carry no JSON-LD at all. The one block, a GovernmentOrganization on the front page, is valid and field-complete.
• Dates are detected through page text or meta tags only, on all five pages. None is in JSON-LD.

How the Five Engines Read the Swedish Pages

Each audit scores the page for five engines and estimates how likely each is to cite it. Averaged across the five Swedish pages:

Per page, as score (estimated citation):

Every engine scores Sweden higher than Norway, and the order is the same except for Perplexity and Copilot, which swap places. Claude (80.8) and ChatGPT (78.4) read the pages best, because both reward the clean, hierarchical structure that all five pages share. Perplexity rises to 66.6 and passes Copilot (61.2). In Norway it was 53.0 against 57.2.

The front page is the only page where Perplexity finds core citation signals, and it shows. The audit notes core citation signals detected and estimates ~63% citation. The four ministries carry the note that there is no JSON-LD, so Perplexity cannot anchor citations, and sit at ~44–45%. The front page is also the only page with a valid structured data block.

Copilot follows the same split. The front page scores 74 with ~47% estimated citation. All four ministries score 58 with ~34%, and each carries the note that no valid JSON-LD blocks were detected.

Gemini is the weakest reader again at 55.2, with an estimate near 23%. All five pages miss the Person and Organization pairing and lack Article or Person JSON-LD. Alt-text coverage is partial on four pages: 58% on the front page, 70% on Climate and Enterprise, 85% on Finance and 90% on Justice. The front page has the highest Gemini score (64) and the lowest estimate (~20%). The Inspector does not explain why the two move in opposite directions.

ChatGPT’s findings are identical on all five pages. Each carries the note that entity signals are weak and that key topics should be named and described explicitly, and each is credited for depth and structure optimized for semantic retrieval. Claude’s findings differ only in degree. All five lack author or Person schema as an expertise signal, and the four ministries add that contextual disambiguation is low.

What the Five Audits Reveal

The template decides the score. Finance, Justice, and Health and Social Affairs score identically on all five dimensions (100, 85, 45, 35, 65) and on every engine score. Climate and Enterprise is a point lower on the index because its extractability is 80, not 85. Only the front page stands apart.

Structure is perfect (100 on every page). Each page has a single H1, H2 sub-sections and logical nesting. No page triggers a structural warning.

Extractability is strong (83.0, range 80–85). Finance, Justice, and Health and Social Affairs score 85. The front page and Climate and Enterprise score 80. The difference is image alt text: five images are missing it on the front page (58% coverage), three on Climate and Enterprise (70%), two on Finance (85%) and one each on Justice (90%) and Health and Social Affairs.

Entity clarity is middling (49.0, range 45–65). The front page scores 65 and the four ministries 45. All five pages score full marks for primary entity and clarity. What differs is the schema layer, and only the front page has one.

Schema and metadata is the weakest of the build dimensions (38.0, range 35–50). The front page scores 50, with one JSON-LD block, a GovernmentOrganization, that is valid and field-complete. The four ministries score 35 and have no JSON-LD. The audit flags no missing canonical URL on any page, which separates Sweden from Norway, where four of five pages lacked one.

Freshness is moderate and uniform (65.0 on every page). The audit detects dates through page text or meta tags and recommends moving them into datePublished and dateModified fields in JSON-LD. No page triggers a staleness warning.

The trust layer is empty. Entity connectivity is 0 on all five pages, including the front page. The Inspector lists Person, Organization and Article schema as missing there, so it appears not to count the GovernmentOrganization block as an Organization, and the audit does not say why. KG anchoring is 0 everywhere because no page has a sameAs link to Wikipedia or Wikidata. E-E-A-T density is 0 everywhere, with zero expertise, experience, authoritativeness and trustworthiness signals on each page. Semantic richness is 54 on every page.

The entity graph is the strongest part of the Swedish profile. Totals run from 69 to 76, all labelled “moderate”. Every page scores 20 of 20 on primary entities, 20 of 20 on secondary entities, 14 of 15 on co-occurrence and 15 of 15 on clarity. The gaps are schema (7 of 20 on the front page, 0 elsewhere) and query alignment (0 to 6 of 10). The detector also finds the same people on every page: Ulf Kristersson, Jessica Rosencrantz, Ebba Busch and Romina Pourmokhtari, with the titles Prime Minister and Deputy Prime Minister. All are labelled “emerging concept” with no schema, so none is tied to a Person.

Query eligibility sits at about 50%, with the front page above it. The audits generated 119 queries across the five pages. Average eligibility is 57.4% on the front page and 50.2–51.0% on the ministries. The highest-scoring query in the set is “startpage” at 61%.

Automated detection picks up interface text and mislabels headlines. Several ministry pages produce queries from interface labels such as “Shortcuts”, “Areas”, “Responsible ministers” and “Filter your results”. A headline about a prime minister leading a caretaker government is tagged CONVERSION on all five pages. Headlines about Gripen fighter jets, nuclear power state aid and the economy are tagged DIGITAL ANALYTICS or AI. “Conversion Rate” appears as a topic on the Health and Social Affairs page. These are artefacts of how the pages are read, and the Boundaries section says how far to trust them.

Page Profiles

Government Offices front page (government.se)

The audit reads the front page as being about “Startpage”. It scores 73, grade C – Fair, the highest in the set.

Strongest findings. It is the only page with JSON-LD, and its GovernmentOrganization block is valid and field-complete. It has the best entity graph in the set (76, “moderate”), the only non-zero schema sub-score (7 of 20) and the only entity the audit marks as schema-backed: “Regeringskansliet”, the Swedish name of the Government Offices, at 95% clarity against 68% on the ministry pages. It is the only page with Perplexity core citation signals (~63%) and has the best Copilot score (74, ~47%). Its 25 queries average 57.4% eligibility, led by “startpage” at 61%, the Budget Bill question and a joint declaration on an anti-ballistic missile coalition, both at 60%.

Weaknesses. Five images lack alt text, and coverage is 58%, the lowest in the set. There is no Article or Person schema, dates are detected only through text or meta tags, and entity connectivity is 0 despite the valid block. Gemini’s estimate is ~20%, the lowest of the five pages even though the Gemini score is the highest. The audit’s own label for the page is “Startpage”, which is what an engine reads as its subject.

What it means for AI systems. The front page shows that the template can carry structured data, and that it pays off: it leads on four of the five engines’ citation estimates and is the only page where the audit reports a schema-backed entity. But the block stops at naming the organization. It has no sameAs links, no people and no dates in JSON-LD, so an engine still cannot connect the Government Offices to a Wikipedia or Wikidata entry or to the people named on the page.

The four ministries

All four score 100 on structure, 45 on entity clarity, 35 on schema and 65 on freshness. None has JSON-LD, a Person or Organization block or a sameAs link.

Ministry of Finance. The detector finds “Artificial Intelligence” as a primary concept on this page, with 73% clarity and query alignment, the only page in the set where an AI topic appears among the detected entities. Query alignment scores 3 of 10. Two headlines carry the lowest eligibility on the page, “Strong Swedish economic recovery despite global uncertainty” and “State aid for nuclear power projects”, both at 45% and both tagged DIGITAL ANALYTICS.

Ministry of Health and Social Affairs. It has the best query alignment in the set (6 of 10) and the second-best entity graph (75). Its headlines include “Government adopts new national strategy against antibiotic resistance” (49%), “Government presents updated national dementia strategy” (45%) and new legislation on the rights and safety of children taken into care (54%). The only topic the detector finds is “Conversion Rate”, which looks like an artefact. Its Gemini estimate is ~28%, five points above the other ministries, and the Inspector does not say why.

Ministry of Justice. The cleanest case of the template: it scores the same as Finance and Health and Social Affairs on every dimension and every engine score. It has the fewest queries (21), among them “Contents on government.se” and “Filter your results”, which are interface labels. One image lacks alt text.

Ministry of Climate and Enterprise. It scores lowest at 67. Three images lack alt text (coverage 70%), extractability is 80 against 85 on the other ministries, and its Perplexity, ChatGPT and Claude scores are each one or two points lower.

What it means for AI systems. The ministries are easy to parse and say nothing about who publishes them. An engine reading a ministry page gets a clean heading hierarchy, current dates and well-formed headlines on antibiotic resistance, dementia care or nuclear power. It gets no structured statement that this is a Swedish government ministry, which body it belongs to or who leads it. Because the four share one template, fixing it once would fix all four.

What This Means

Swedish government pages are built to be read and not yet built to be identified. Structure is 100 on every page and freshness 65, and entity connectivity, KG anchoring and E-E-A-T density are 0 on every page. The pages tell an engine what they say and when. They do not tell it who is saying it.

The template is both the strength and the lever. Three ministries score identically and the fourth is a point lower, so one change to the ministry template reaches all four. The front page shows the first step: the one page with a JSON-LD block has the highest index (73), the only Perplexity core citation signals (~63% against ~45%) and the best Copilot result (74 against 58).

Dates are in the page but not in the data. All five pages show dates through text or meta tags. Moving them into JSON-LD with a dateModified is a small change, and the audit recommends it on every page.

People are named but not marked up. The detector finds four named people, plus the titles Prime Minister and Deputy Prime Minister, on every page, each as an “emerging concept” with no schema. Person markup is the missing link between the pages and the people who lead the government.

Scope and Method

What was audited. Five Swedish government pages, one page each, assessed with the AI Visibility Inspector and Srna Frameworks. All five audits were generated on 13 October 2026.

• government.se, the front page of the Government Offices
• government.se/government-of-sweden/ministry-of-finance/
• government.se/government-of-sweden/ministry-of-justice/
• government.se/government-of-sweden/ministry-of-climate-and-enterprise/
• government.se/government-of-sweden/ministry-of-health-and-social-affairs/

These five were chosen because they are the English-language public door of the Swedish government: the front page and four ministries that cover finance, justice, climate and enterprise, and health and social affairs.

What the research measures. Each report gives an AI Retrieval Index and grade, Structure, Depth and Schema sub-grades, and five scored dimensions: Structural Integrity, Data Extractability, Entity Clarity, Schema & Metadata, and Freshness & Decay. It adds scores for five engines (Perplexity, OpenAI / ChatGPT, Claude, Google Gemini, Microsoft Copilot) with estimated citation percentages, a Schema Intelligence section (validation, entity connectivity, KG anchoring, E-E-A-T density, semantic richness), query eligibility and an entity graph stability score.

Boundaries

• Five pages of one site, not five institutions. The ministry pages share one template, so the audits are not five independent observations. Finance, Justice, and Health and Social Affairs return identical dimension scores.
• One page per entry. Subpages, PDFs, the Swedish-language site and other domains are not covered.
• Findings describe each page on the audit date. The headlines quoted are those on the pages on 13 October 2026.
• Averages are simple means of the five page scores. Norway, Ireland, Hungary, Greece and Germany comparison figures are those published in the earlier profiles.
• Each report produces two freshness blocks per page. The scores matched within each page.
• The audit’s scoring is not fully reconciled. The front page has a valid GovernmentOrganization block, yet Organization is reported as not detected and entity connectivity is 0. Its Gemini score (64) is the highest in the set and its Gemini estimate (~20%) the lowest. Justice has one image missing alt text and a 90% coverage note, while Health and Social Affairs has one image missing and no coverage note. The Inspector does not explain any of these.
• The Inspector reads pages without running scripts. If any of these pages adds JSON-LD with JavaScript after load, the audits cannot show it.
• Entity and query detection is automated and produces artefacts. Interface text is read as queries and entities, “Product / Service” is read as a commercial entity, headlines are tagged DIGITAL ANALYTICS, AI or CONVERSION, and “Conversion Rate” is read as a topic. Queries and entity labels are quoted as the audit displays them.
• Every report lists a similar set of generic secondary content gaps, such as Large Language Model and Dark Funnel. The list is not interpreted here.
• Citation figures are the Inspector’s estimates, not observed citations in live engine answers.
• Headlines, names and figures in the page profiles are those detected on the audited pages. They are reported as page content and are not independently verified.

Calculate Your Potential Exposure

Reading poorly to AI has a cost. The AI Visibility Cost of Inaction Calculator lets you enter your own figures and see the exposure for your organization.