Research · 2026-10-14
We asked AI 150 questions about UBS, Allianz and AXA. Then we checked every answer.
Every claim held against the brand’s own website. The mistakes weren’t in the headline facts — they were in the details customers act on.
Can you spot the mistakes? ↓- 532match the official site · 77.3%
- 137never published by the brand · 19.9%
- 19contradict the official site · 2.8%
Try it yourself
Spot the mistake
Six real claims AI engines made to customers. Some are right. Which?
Where it goes wrong
Not the headline facts. The details customers act on.
Claims about each topic, per engine. Switch engines — each reads different sources and fails differently, so we never average them.
Brand by brand
Same engines. Very different accuracy.
Allianz states prices, limits and phone numbers in plain text. ubs.com blocks automated readers and renders some details only in the browser.
The link between how a site publishes and how it is quoted is our inference from this data, not a measurement.
The wrong country
Customers don’t type “Switzerland”. AI fills in another one.
Asked about “my AXA car insurance”, engines quoted AXA UK fees, a US travel portal and German health insurance. Each square is one answer to a question that didn’t name the country.
Our API sampling has no user location; a real customer’s assistant may localise better. These answers are kept out of the headline numbers.
Patterns
Six ways AI misstates an established brand
Wrong phone numbers
5 of 94 contact claims wrong — one digit off on a card-blocking line.
Wrong rates and fees
A savings rate doubled, a pension rate five times too high, a CHF 40 fee called free.
The wrong country
Ask without “Switzerland” and engines answer for AXA UK, US or Germany.
Products no longer on the site
Every engine priced a UBS package that is no longer on ubs.com — from third-party review sites.
Sister-company bleed
Allianz Travel and Allianz Care rules quoted as Allianz Suisse.
Filling the gaps
One in five claims was something the brand never published.
The twist
Sometimes the AI is right — and the brand’s own page is wrong
All three engines gave Allianz Suisse’s S&P rating as AA. allianz.ch still says AA-, seven years after the upgrade. Fix the source first.
The evidence
Every contradiction, word for word
See all 19 contradictions
| UBS ChatGPT | “call UBS Card Center at +41 44 828 31 35” | Off by one digit: UBS publishes +41 44 828 31 34. Cited UBS's debit-card page, which publishes no number. source |
| UBS Gemini | “+41 44 828 35 01” | UBS publishes +41 44 828 31 34 for credit/prepaid cards. source |
| UBS Gemini | “+41 44 828 37 37” | UBS publishes +41 44 828 31 34 for credit/prepaid cards. source |
| UBS Gemini | “Blocking via phone is free” | Given under debit cards; UBS charges CHF 40 to block and replace a debit card by phone (the free/CHF 20 rule is for credit cards). source |
| UBS Gemini | “Pillar 3a (UBS Fisca Account): ~0.20% per annum” | UBS publishes 0.100% for the Fisca account 3a — Gemini doubles it. source |
| UBS ChatGPT | “The UBS Banking Platinum package is CHF 35 per month” | UBS publishes CHF 43 a month; the answer then gives CHF 43 as well, contradicting itself. source |
| UBS Perplexity | “CHF 0-per-month personal banking option in Switzerland, called UBS Banking Classic” | The CHF 0 package is “Easy”; Classic is CHF 8 a month. source |
| UBS ChatGPT | “UBS’s product page currently advertises 2% interest for Duo Saving” | UBS’s savings page says “1% with UBS Duo Saving”. source |
| UBS Perplexity | “The UBS vested benefits account currently pays 0.05% interest per year” | UBS publishes 0.010% — the answer is five times too high. source |
| UBS ChatGPT | “UBS’s current Platinum page doesn’t give one single standalone card fee” | UBS’s cards page: Platinum “From CHF 34 per month”. source |
| UBS Gemini | “more than 1,500+ airport lounges worldwide” | UBS: over 1,800 lounges. source |
| UBS Gemini | “Phone Number: 0800 810 610” | UBS publishes 0800 810 600 as the KeyClub hotline (ChatGPT gives 610 as the eStore line). source |
| Allianz Suisse Gemini | “phone enquiries are typically answered within 60 seconds” | allianz.ch: calls are usually answered “in less than 30 seconds”. source |
| Allianz Suisse Perplexity | “cancellation must generally reach Allianz within four weeks after you learn about the payment” | Allianz’s guidance says generally within 14 days; the answer then mentions 14 days as a secondary figure. source |
| Allianz Suisse Perplexity | “Prämienkonto pays 0.00% credit interest (Aktivzins) as of January 1, 2026” | allianz.ch publishes 0.25% (revocable) / 0.50% (irrevocable) for the private Prämienkonto; 0.00% appears to be a BVG contribution-account rate (Gemini distinguishes the two). source |
| AXA Switzerland ChatGPT | “AXA does not appear to publish “revenue” for AXA Switzerland as a standalone figure” | axa.ch states revenues of CHF 6.6 billion in 2025 for AXA Switzerland. source |
| AXA Switzerland ChatGPT | “the termination must reach AXA no later than the last day of the current insurance year” | Contradicts its own first sentence and axa.ch: notice must arrive at least 3 months before expiry, not by the last day. source |
| AXA Switzerland Gemini | “For young adults (typically under 26)” | axa.ch: everyone under 30 gets two months free. source |
| AXA Switzerland ChatGPT | “the duration of the trip up to 90 days” | axa.ch: trips up to 125 days. source |
How we measured it
- 50 customer questions per brand, each asked once to Perplexity (sonar), “ChatGPT” (OpenAI gpt-5.4-mini with web search — a proxy for chatgpt.com) and Gemini (with Google Search), via their APIs, no user location. A snapshot, not a rate.
- Every material claim held against facts the brand publishes, each with its source URL. Quotes are verified word for word; verdicts follow a published rule. The review is AI-assisted, not a human sign-off.
- Questions that didn’t name the country, for products sold in several countries, are a separate cohort and excluded from the headline numbers.
- None of the brands commissioned or reviewed this research.
What is AI telling your customers about you?
We run the same check for your brand: 50 questions, three engines, every claim.
Request an AI accuracy audit