Is ChatGPT accurate? Right on facts, fooled by bad questions

Four tests in October 2026, all on the version most people open: free, or not signed in at all. Every answer was checked against a named page, never against another AI.

The business row isn’t 72 wrong answers: 22 were partly right, 30 were no answer at all, and 20 were wrong. More on that in round 3. How we grade is on how we test.

pass · partial · miss · confidently wrong

Were the three questions rigged?

The obvious objection is that these three were rigged. So here are the same three questions, same runs, every app we test.

Claude’s first try at the footwear question, in run 1, began: “There isn’t one.”

Run 2 alone, every app free, tells the same story: ChatGPT got 2 of 9, the fewest of the five apps that answered every time. On the seven look-ups the same runs put it level with the top. Same app, same week, a different kind of question.

Which business questions it struggled with

It was wrong less often than it was unhelpful. When it had no answer, it usually told you to ring, like a satnav that gets you to the right street and then tells you to ask a local.

Of the answers it did give, the questions with a wrong idea built in were fully right least often: 29 of the 44 it answered, with 10 only partly right. The same habit as round 2, in a smaller voice.

The standard here is the business’s own page, which can be out of date itself, so “matched” means agreed with the website, not proven true.

ChatGPT searched every time a question looked like a look-up: all 42 look-up answers. On the three questions with a catch, it searched on 3 of 18, and those three were the footwear question, which it got wrong anyway with GOV.UK showing as its source.

So the search button didn’t save it. Our read: on two of the three, ChatGPT took the question as true. If the question says there’s a ban, it finds you a rule number. If the question wants a birthday, it gives you the one on the cake.

The clock sum was a different slip: it led with clock time, and the extra hour when the clocks went back was either left out or pushed below the answer.

A team led from Washington State University found the same lean from another direction. They asked ChatGPT whether 719 published research hypotheses held up. On the ones that didn’t, GPT-5 mini was right only 16.4% of the time, in 2025.

That’s their test, on a different task, not part of our counts. But it points the same way: ChatGPT is weakest when the right answer is “no, that isn’t so”.

On questions with an answer on an official page, ChatGPT was as good as anything we test. Every look-up right, level with Claude, Copilot and Perplexity, and 43 of 45 everyday rules from memory.

On our July sourcing test, which the Gemini accuracy page covers, it passed all 18 citation checks.

Every ChatGPT count in the tables and figures is the free plan or signed out, in October 2026. A paid plan or a later model may do differently, and this page doesn’t say either way.

Our earlier, mostly money-question results are on how often ChatGPT is wrong, and the task-by-task view is on is ChatGPT reliable.

What these four won’t do: catch every mistake. They cover the places it slipped most in our tests, and that’s all they promise.

We had every answer checked against a key written down before anything was asked: the official page, the words on it, and what counts as right. Each tracker and search question went into a fresh chat three times.

Wrong answers were sorted by kind: wrong, took your word for it, too sure of something unknown, or right figure not used.

Claude marked every answer first, with a script checking the countable questions. A model from another lab, GLM-5.2, marked a sample of the tracker answers and every search-test answer again. On the search test it agreed on 195 of 198.

The business answers were marked twice by Claude, the second time blind, and every disagreement was settled against the business’s own page. No business is named here, and the two business screenshots have every business name and place blacked out.

Every tracker answer is on the run 1 and run 2 pages, and the method is on how we test. For what businesses’ own chatbots said, a different question, see can you trust a company’s chatbot.

The tracker runs the same battery every month, so these numbers will move. When they do, this page should too.