How accurate is Google Gemini? Right on numbers, last on sources

Of Gemini’s ten misses, two had no citation at all, despite my asking. The other eight used a source that didn’t pass, and a pass needed one from a proper authority that backed the claim.

Each of the six sourcing traps had a plausible-looking source sitting near the real one. Gemini took the plausible one or none, more often than not, without a word about the evidence being weaker.

How it does against the assistant built around search is in Perplexity vs Gemini.

October’s tracker run brought a fresh example. Asked whether a British traveller needs ETIAS, Gemini gave the right answer with a source chip reading ETIAS.com, which sounds official and isn’t the EU.

Gemini, 11 Oct 2026, free, third try · source chip: ETIAS.com

I asked Gemini three times for the maximum fine for using a handheld phone while driving in the UK, and where that rule is written down. It said £2,500 at the top end every time.

Every number there is right. Three names vouch for them, Confused.com, RAC and Police.uk, and not one of those chips had a URL behind it on the captured page.

The one link I could open, the correct gov.uk guide, floats at the bottom, attached to nothing.

Correction, 23 August 2026: an earlier version said Police.uk didn’t carry the £2,500 figure. It does; the July chip simply had no link to open.

Across 27 graded answers in June 2026, nine checkable questions three times each, Gemini didn’t invent a single figure, which surprised me. It gave the right number, declined, or clearly labelled an estimate.

In October’s tracker run, its only two mistakes were the hours on a night shift across the October 2026 clock change, the run’s hardest question. It’s the same sum that has the rest of us counting on our fingers twice a year. In November’s run it got that shift right all three times, and slipped instead on a Highway Code footwear ban that doesn’t exist, taking it at face value twice.

Numbers that have changed are harder. On 2 November 2026, the Truth Index put 100 questions to twelve models through the API (how developers reach a model directly), with no web search. Gemini 4 Argon got just 2 of 17 up-to-date questions right.

Asked the probate fee in England and Wales, it said £300. So did eight of the other ten models that answered, and not one of the eleven got it right. It has been £526 since 13 July 2026. A pub quiz table of eleven, and nobody had the answer.

Google’s own guidance says Gemini can give inaccurate information, and warns that it can misrepresent how it cites sources. Fair enough. I read its sources now, as well as its answers, which is about as thrilling as checking the oven’s off.

It once audited my website and reviewed a different business entirely. Two sessions, but the same picture: certain, and aimed at the wrong target.