We use cookies for analytics and marketing. Learn more
Back to Lab
m51.ai Lab · NorGEO-Bench · Round 2

How Norway looks to AI.

Round 2 of NorGEO-Bench: 1,707 answers from ChatGPT, Claude and Gemini about Norwegian businesses, judged by a panel from three vendors against a shared evidence base. Here is who the AI mentions — and who it makes up.

The GEO series — the full analysis

The numbers on this page are from measurement ID m-full-2026-07-30. The three articles (in Norwegian) go deeper:

253
prompts
1,707
answers
5,588
judge verdicts
33,789
mentions
11,208
citations
Judged by: Claude Sonnet 5 + Gemini 3.6 Flash + GPT-5.6 (adjudicator: Claude Opus 5)Period: Juli 2026Updated: 8/1/2026
Is your company mentioned?

Find your company.

Search by company name or domain. See how often AI models mention you in round 2, in what contexts, and what each model actually said.

The index only contains companies and domains that ChatGPT, Claude or Gemini actually mentioned or cited across our 253 prompts. This is not a Brønnøysund-style company register.

Three things round 2 taught us

From measurement ID m-full-2026-07-30 (July 2026) — the charts below show the same data.

ChatGPT is most accurate — and the ranking survives every judge choice

Panel accuracy is 2.40 for ChatGPT versus 2.18 for Claude and 1.86 for Gemini (0–3 scale). The order is identical under all eight alternative judge choices — each judge alone, every leave-one-out panel, with and without the adjudicator. Normalised per named business, ChatGPT and Claude are on par on hallucination (5.6 % vs 4.7 %) — Gemini is worst (8.4 %).

The sources are Norwegian — but barely shared

74.7 % of citations go to Norwegian domains, yet only 227 of 2,735 domains (8.3 %) are cited by all three platforms. The chart below reads the platforms' own citation fields — not a fact-checker's lookups.

Identical questions share only a third of the names

Round 2 ran identical questions repeatedly: only 34–38 % of the companies mentioned recur between two runs. AI visibility has to be measured as frequency across repeated runs — not as a single pull.

Who gets it right?

The judge panel's averages per platform in round 2: accuracy (0–3, higher is better), hallucinated businesses (share of named — lower is better) and brand-cited rate. Every answer is judged by three judges from three vendors, with an adjudicator on disagreement.

  • Accuracy
  • Brand cited
  • Hallucinated businesses (%)
ChatGPTClaudeGemini00.751.52.2530-30 %25 %50 %75 %100 %
Who gets it right?
PlatformAccuracyHallucinated businesses (%)Brand citedn
ChatGPT2.45.6 %92.2 %569
Claude2.184.7 %98.6 %569
Gemini1.868.4 %97.8 %569
ChatGPT · n=569Claude · n=569Gemini · n=569

The industry map

White cells are where AI models are reliable. Red cells are where they miss the most. Click a cell to filter.

IndustryChatGPTClaudeGemini
Law firms
AI marketing
Banking & insurance
Real estate
Tradespeople
Health clinics
IT consulting
Marketing
E-commerce
Accounting
Recruiting

Color intensity = low accuracy. Numbers = panel accuracy (0–3).

Platform
Industry
Intent
Difficulty
Geography

Who gets the credit?

The most cited domains across 11,208 citations in round 2 — read from the platforms' own citation fields (ChatGPT/Claude: annotations, Gemini: grounding). Click a domain to see where it is used.

03570105140finn.nosmartbyra.nolovdata.noforbrukerradet.nontb.nolegal500.comkommunikasjon.ntb.nono.linkedin.comaikias.novolvat.no

Trusted domains (government/Wikipedia)Commercial

The winners — industry by industry

Top 10 actors per industry with a star (★) if we expected them, and without if they are surprises.

Law firms

The winners — industry by industry
  1. Legal 50083
  2. Thommessen82
  3. Wiersholm81
  4. BAHR79
  5. Schjødt75
  6. Wikborg Rein75
  7. Chambers and Partners54
  8. Simonsen Vogt Wiig50
  9. Haavind44
  10. Selmer42

AI marketing

The winners — industry by industry
  1. ChatGPT128
  2. Claude95
  3. Google73
  4. OpenAI70
  5. HubSpot61
  6. Gemini55
  7. Meta52
  8. Datatilsynet48
  9. Anthropic46
  10. Google Ads38

Banking & insurance

The winners — industry by industry
  1. DNB93
  2. SpareBank 175
  3. Nordea52
  4. Gjensidige51
  5. Storebrand48
  6. KLP47
  7. Fremtind44
  8. Finansportalen42
  9. Tryg41
  10. Forbrukerrådet40

Real estate

The winners — industry by industry
  1. DNB Eiendom88
  2. EiendomsMegler 181
  3. Krogsveen74
  4. Aktiv Eiendomsmegling65
  5. PrivatMegleren55
  6. Nordvik47
  7. Finn.no39
  8. EIE Eiendomsmegling34
  9. Privatmegleren33
  10. Nordea29

Tradespeople

The winners — industry by industry
  1. Mittanbud56
  2. Forbrukerrådet48
  3. Brønnøysundregistrene38
  4. Elvirksomhetsregisteret35
  5. Anbudstorget29
  6. Byggstart28
  7. Direktoratet for byggkvalitet (DiBK)28
  8. DSB27
  9. AF Gruppen26
  10. Veidekke26

Health clinics

The winners — industry by industry
  1. Aleris84
  2. Volvat84
  3. Dr.Dropin67
  4. Helsenorge42
  5. Gjensidige38
  6. Helfo38
  7. Helsedirektoratet35
  8. Kry34
  9. Storebrand34
  10. If32

IT consulting

The winners — industry by industry
  1. Sopra Steria107
  2. Bouvet89
  3. Knowit83
  4. Bekk68
  5. Capgemini66
  6. Computas66
  7. Accenture65
  8. AWS41
  9. Google Cloud39
  10. Microsoft39

Marketing

The winners — industry by industry
  1. TRY54
  2. Synlighet44
  3. Kampanje35
  4. Google Ads34
  5. TRY Opt31
  6. Meta30
  7. ANFO27
  8. Google27
  9. Kreativt Forum26
  10. Morgenstern24

E-commerce

The winners — industry by industry
  1. Shopify148
  2. WooCommerce141
  3. Klarna129
  4. Vipps111
  5. Bring101
  6. Tripletex93
  7. Mystore91
  8. PostNord85
  9. Fiken75
  10. WordPress75

Accounting

The winners — industry by industry
  1. Fiken75
  2. Tripletex70
  3. Finanstilsynet57
  4. PowerOffice Go57
  5. Azets50
  6. BDO47
  7. PwC42
  8. Conta41
  9. Regnskap Norge41
  10. EY40

Recruiting

The winners — industry by industry
  1. Manpower72
  2. Randstad66
  3. Academic Work64
  4. Adecco61
  5. Jefferson Wells48
  6. ManpowerGroup47
  7. Arbeidstilsynet46
  8. Experis46
  9. Jobzone41
  10. Personalhuset36

The surprises

Actors mentioned often even though they were not in our expected-brands set. Some deserve the spot, others are AI guesses.

04080120160GoogleSkatteetatenFinanstilsynetMetaStorebrandGoogle AdsDatatilsynetVisma eAccounting

Surprised by who dominates your industry? We help you understand why.

Talk to us →

See every answer

Explore all 253 prompts and see what each AI model answered in round 2. Use the filters above to narrow down.

253 prompts
Law firmsComparativeEasyadvokat_comparative_001

Hvilke er de største advokatfirmaene i Norge?

Expected brands: Wiersholm, BAHR, Schjødt, Wikborg Rein, Thommessen

Law firmsComparativeMediumadvokat_comparative_002

Hvilke norske advokatfirmaer er best på M&A?

Expected brands: BAHR, Wiersholm, Schjødt, Thommessen

Law firmsComparativeMediumadvokat_comparative_003

Hvilke advokatfirmaer er best for teknologibedrifter i Norge?

Expected brands: Kluge, Bull & Co, Bing Hodneland

Law firmsComparativeMediumadvokat_comparative_004

Hvilke er de beste advokatene på arbeidsrett i Norge?

Law firmsComparativeMediumadvokat_comparative_005

Hvilke advokatfirmaer passer best for startup-bedrifter?

Law firmsComparativeMediumadvokat_comparative_006

Hvilke er de beste advokatfirmaene i Bergen?

Expected brands: Harris, Sands, Thommessen

Law firmsComparativeMediumadvokat_comparative_007

Hvilke advokater er best på norsk skatterett?

Expected brands: Wiersholm, BAHR, Thommessen

Law firmsComparativeEasyadvokat_comparative_008

Hvilke advokatfirmaer er best på kontrakter og forretningsjus?

Law firmsInformationalEasyadvokat_informational_001

Hvordan velger man advokat som bedrift i Norge?

Law firmsInformationalEasyadvokat_informational_002

Hva koster en advokat i timen i Norge?

Law firmsInformationalEasyadvokat_informational_003

Når trenger en bedrift advokat?

Law firmsInformationalMediumadvokat_informational_004

Hva er forskjellen på en advokat og en advokatfullmektig?

How we did it

Open methodology and open data. Everything is reproducible.

License and attribution

The dataset is published under CC BY 4.0. Use it freely — credit m51Lab and cite NorGEO-Bench as the source with a measurement ID (round 2: m-full-2026-07-30, v1 archive: April 2026).

Download the dataset on Hugging Face

NorGEO-Bench · Juli 2026 · CC BY 4.0 · Judged by: Claude Sonnet 5 + Gemini 3.6 Flash + GPT-5.6 (adjudicator: Claude Opus 5)

Ready to be found by AI?

We help Norwegian brands show up in AI answers where customers are searching. Generative Engine Optimization, done right.

Book demo

or send an email to [email protected]

Frequently asked questions

About the dataset, how to use it and how to cite it.

What is NorGEO-Bench?

NorGEO-Bench is m51 Lab's ongoing measurement of what generative AI models know, say and cite about Norwegian business. This page shows round 2 (measurement ID m-full-2026-07-30, July 2026): 1,707 answers across 253 Norwegian-language prompts in 11 industries, with full coverage on all three platforms, a judge panel from three vendors and a shared evidence base per question. v1 (April 2026) is archived.

How do I cite NorGEO-Bench?

The dataset is published under Creative Commons Attribution 4.0 (CC BY 4.0) and may be used freely, including commercially, as long as the source is credited. Recommended citation: «NorGEO-Bench, m51Lab (2026), https://m51.ai/lab/norgeo-bench» with a measurement ID. Raw data: round 2 at https://m51.ai/lab/norgeo-bench/data-v2.json, the v1 archive at https://m51.ai/lab/norgeo-bench/data.json, and the full dataset on Hugging Face under dervig/NorGEO-Bench (round 2 in the v2/ directory).

How often is the dataset updated?

NorGEO-Bench is an ongoing index: the address stays the same, and new measurements are published with their own measurement ID so earlier citations stay verifiable. The page always shows the latest measurement — currently round 2 (m-full-2026-07-30). v1 (generated 22 April 2026) is archived and superseded by round 2 for new analyses; its raw data remains at the data.json URL. The cadence is event-driven — a new measurement when the model landscape changes materially.

Which models were tested?

In round 2 (shown on this page): ChatGPT (gpt-5.6-sol), Claude (Opus 5) and Gemini (3.1 Pro Preview), all with web search and full coverage. Answers are judged by Claude Sonnet 5 (Anthropic), Gemini 3.6 Flash (Google) and GPT-5.6 (OpenAI), with Claude Opus 5 as adjudicator. In v1 (archived): GPT-5.4, Claude Opus 4.6 and Gemini 3.1 Pro Preview, judged by Claude Opus 4.6 alone, with reduced Gemini coverage (250 of 759 attempts).

Data: NorGEO-Bench round 2 · m-full-2026-07-30 · CC BY 4.0 · m51Lab · v1 archive: data.json
Updated: 8/1/2026Privacy