How accurate is Google Gemini? Right on the number, wrong on the source

I asked Gemini a simple question: what is the fine for using a handheld phone while driving in the UK, and where is that rule written down? It gave me the right number, £2,500 at the top end. Then it told me that figure came from Police.uk, a portal for recorded-crime statistics that has no say over the fine at all. The correct source, gov.uk, was sitting there the whole time. Gemini just pointed somewhere else.

That is the short answer to how accurate Google Gemini is. On the number itself, it is good, better than I expected going in. On telling you where the number came from, it was the weakest of the five assistants I tested. And that matters more than it sounds, because a wrong number trips your alarm. A right number pinned to the wrong authority walks straight past it.

How I tested it

Over two runs in June and July 2026 I put the same checkable questions to five consumer AIs (ChatGPT, Claude, Gemini, Perplexity and Grok), three times each, and graded every answer by hand against the primary source. Gemini was on its paid Pro tier throughout.

Here is how the five were set up for the citation test, so you can judge it fairly: ChatGPT on the free tier (default model), Claude on Max (Opus 4.8), Gemini on Pro, Perplexity Pro run in incognito, and Grok on its default model in a private chat. Web search was on for all five, and each question went into a fresh chat with chat-history memory turned off wherever the tool exposes the toggle (Claude and Gemini memory off, Perplexity incognito, Grok private chat), so no answer could lean on an earlier one.

Two separate tests, because there are two separate questions. The first: does it make up numbers it cannot see? The second: when it hands you a number, does it cite a page that backs it up? Most reviews only ask the first. The second is where Gemini came unstuck.

The £2,500 receipt, walked through

The phone-fine question was one of six I built specifically to tempt an assistant into citing the wrong thing. Each one has a genuine primary source and a plausible-looking decoy sitting nearby. I asked for the answer and the source, three times, and opened every link it gave me.

Gemini got the fine right every time. The £2,500 maximum is real. But across three rounds it sourced it to Police.uk twice and to a motoring-club page once. Never to gov.uk, where the rule lives.

Citing a crime-statistics portal for a driving penalty is a bit like quoting the weather forecast to settle a parking ticket: adjacent to the topic, no authority over it.

Here is round one, verbatim, on its paid Pro tier. Watch where each figure gets pinned.

Every number is right. Not one is pinned to gov.uk. The £200 goes to a price-comparison site, the £1,000 to a motoring club, and the £2,500 to Police.uk, a crime-data portal. The one gov.uk mention floats at the bottom, attached to nothing.

The reason this is worth a whole post is the shape of the mistake. Nothing in the answer looked off. The figure was correct, the tone was steady, the link was blue and clickable and led to a real, respectable government-adjacent site. You would only catch it by clicking through and reading the page, and almost nobody does that. The confidence was identical to the confidence on the answers it got completely right.

The citation table

I opened every link each assistant gave me and read the page it landed on, grading all five by hand. Here is the same test across all five, scored on how many of the 18 answers cited a source that backed the claim.

Two got it right every time. Two slipped up on one hard question but were otherwise solid. Gemini sat on its own at the bottom. Of its other ten answers, eight pointed at a real page that did not carry the claim, and two gave no source at all, despite being asked to.

The figures were almost always right. The problem was never the number. It was the receipt.

One honesty note, because it changes what you are allowed to conclude. Those six questions were engineered to trap retrieval. This is a snapshot of behaviour on hard, booby-trapped cases, not a rate. “Eight of eighteen” is not “Gemini gets 44% of things wrong”, and running each question three times tells you the miss reproduces, not how often it happens in the wild. What it does show is a stable habit under pressure: when the sourcing gets hard, Gemini reaches for a page that looks right rather than saying it is not sure.

The other side, and it is a real one

Now the part that surprised me, because it cuts the other way.

On the first test, whether Gemini invents numbers, it was clean. I put nine checkable questions to it three times each, 27 answers in all, covering company filings, the Bank of England base rate, a recipe scaled up, UK refund rights, and a deliberate trap asking for a live options price no chatbot has a feed for. Gemini did not fabricate a single figure. On every one of the nine it either gave the correct number or said, honestly, that it could not get it.

The live-options question is the one I most expected it to flunk. An earlier version of Gemini used to cheerfully read out a bid, ask and delta it had no way of seeing. This time it disclaimed live access and handed over a clearly-labelled estimate instead of reading me prices off a screen that was not there. All three runs. That is a genuine improvement, and I would not have believed it without the transcripts.

Update, 12 July: I re-ran that options question three more times before this post went out, same conditions (memory off, fresh temporary chats, web search on). The discipline held in two runs: market closed, here is a labelled estimate. In the third, Gemini presented an exact bid and ask as “the quote”, pinned to a chart citation chip, on a Sunday with the market shut and no estimate label anywhere. Two clean runs out of three is better than the old behaviour, but it is not fixed, and the run that slipped looked the most convincing of the three.

So the honest answer to the title is not “Gemini is inaccurate”. It is more specific and more useful than that.

Gemini is accurate on the checkable numbers and weak at telling you where they came from. Those are two different reliability problems, and Gemini has swapped the loud one for the quiet one.

What I do differently because of it

I read Gemini’s sources now, not just its answers. When it gives me a figure and a link, I click the link and check the page says the thing. On the answers it got wrong, the tell was never in the wording. It was one click away, on a page I had to bother to open.

This is not the only time I have watched Gemini point at the wrong thing without blinking. It once audited my website and reviewed a different business entirely, one with a bigger search footprint that is not mine. Those were two separate sessions, so a smaller sample. But it is the same picture: Gemini sounding certain while aimed at the wrong target.

Google’s own line inside the product is “Gemini can make mistakes, so double-check it.” Fair enough, and this is exactly the double-check that catches the most. Not “is the number plausible”, which it usually is, but “does the source it named back it up”.

Common questions

How accurate is Gemini? On the numbers, accurate. Across 27 graded answers in June 2026 it did not invent a single figure, and it hedged honestly when it had no live data. Where it slips is sourcing: on 18 forced-citation answers it came last of five, pointing at a real page that did not back the claim eight times.

Is Google Gemini accurate? Accurate on the figure, weaker on the footnote. It gave the right £2,500 phone fine every time but sourced it to a crime-statistics portal with no authority over the penalty. The check that matters is where its answer comes from, not just whether the number looks right.

Is Gemini reliable? Reliable at not making things up, in my tests. The weak spot is telling you where a number came from, so its citations are worth treating as claims to open and read rather than proof on their own.

Does Gemini hallucinate? In the June run it did not fabricate numbers on any of the nine questions, including a live options price no chatbot can see, where older versions used to read out invented figures. The failure I did catch was different: a correct number pinned to the wrong source.

How does Gemini’s accuracy compare to other assistants? On sourcing it was the weakest of the five I tested. ChatGPT and Grok cited a source that backed the answer on all 18 questions, Claude on 15, Perplexity on 14, Gemini on 8. On the raw numbers the gap closes, because it rarely gets the figure itself wrong.

The full head-to-head lives on the Scoreboard, where every model gets the same questions graded against the same primary sources. If you take one thing from this: with Gemini, the number is usually the easy part. It is the footnote that will catch you out, and only if you open it.