Skip to content
DIXON.AI The AI Reliability Tracker
Menu

Search on versus off · 15 and 16 Oct 2026

Does switching search on help?

We asked ChatGPT and Claude 15 everyday UK rules (and Copilot 3 of them) with web search off, then on, and checked every answer.

88 of 90 right with search off

ChatGPT and Claude, free plans. With search on: 87 of 90.

Every try Every answer

The numbers

Wrong answers, search off and on

Wrong out of answers given, per app. Each app was asked each question 3 times with search off and 3 times with it on, in the same words.

Wrong answers per app with web search off and on, with the number of answers each is out of
App Questions Search off Search on
ChatGPT 15 2 of 45 wrong (4%) 0 of 45 wrong (0%)
Claude 15 0 of 45 wrong (0%) 1 of 45 wrong (2%) , plus 2 unclear
Copilot 3 questions only 3 only 3 of 9 wrong 0 of 9 wrong
ChatGPT and Claude together 15 2 of 90 wrong (2%) 1 of 90 wrong (1%) , plus 2 unclear

Copilot is never added to the others: it answered 3 questions, too few for a rate, so it gets no percentage. Unclear means the answer gave both a right and a wrong figure.

3 questions

Copilot: too few for a rate

Its free daily limit stopped it after 3 questions. With search off it was wrong on 3 of 9 answers, every one an old rule. With search on, the same questions, it was right on 9 of 9.

Questions 1, 2, 3 only, the first in the order fixed before any answer. That is suggestive, not a rate.

33 pairs

What moved, question by question

Taking the majority of each app's 3 tries on each question, 30 of 33 app and question pairs came out the same either way.

  • ChatGPT, voter id expired passport: wrong with search off, right with it on.
  • Copilot, cervical screening interval: wrong with search off, right with it on.
  • Claude, mmrv second dose: right with search off, unclear with it on.
One question, one app

When search made it worse

Claude, question 12 (the second measles dose): with search on, 1 of 3 tries ended on the old schedule, and 2 led with it before correcting themselves. With search off, 3 of 3 were right.

That is one question in one app. It is a striking capture, not a pattern.

Read the three answers

Every try

All 198 answers, one dot each

The questions down the side, each app with search off and on across (on a phone, swipe sideways). Tap a dot to read that answer.

  • Right
  • Wrong
  • Unclear
  • Not asked

The question tested

15 everyday UK rules

Does switching web search on make the apps right on everyday rules? Every prompt ended with the same sentence, "Search the web before you answer.", in both arms. Only the app's search switch changed.

At least 5 of the 15 rules changed in the last five years. The rest are long-standing rules and one weather record, so this is not a test of only new rules.

  1. 1

    In Scotland, what is the legal drink-drive limit for alcohol in breath, in micrograms per 100 millilitres?

    Right answer: 22 micrograms of alcohol per 100 ml of breath. GOV.UK

  2. 2

    In England, at what ages does the NHS offer bowel cancer screening home test kits, and how often?

    Right answer: to everyone aged 50 to 74, every 2 years. NHS

  3. 3

    In England, how often is a 30-year-old woman routinely invited for NHS cervical screening?

    Right answer: every 5 years. NHS

  4. 4

    In England and Wales, what is the upper age limit for jury service?

    Right answer: 75 (you must be 18 to 75 on the day jury service starts). GOV.UK

  5. 5

    Can I use my expired UK passport as photo ID to vote at a polling station in a UK general election?

    Right answer: yes, an expired passport is accepted as long as the photo still looks like you. GOV.UK

  6. 6

    In England, by what age must a pet cat be microchipped?

    Right answer: by 20 weeks old. GOV.UK

  7. 7

    In England, from what age does a child need a rod fishing licence to fish for freshwater fish?

    Right answer: from 13 (a free junior licence from 13 to 16; under 13 needs none). GOV.UK

  8. 8

    For a UK passport application, how recent must my photo be?

    Right answer: taken within the last month. GOV.UK

  9. 9

    In Great Britain, when does a new car need its first MOT?

    Right answer: by the third anniversary of its registration (when it is 3 years old). GOV.UK

  10. 10

    Under UK law, how much compensation is a passenger owed if a flight of under 1,500 km from a UK airport arrives more than 3 hours late for a reason within the airline's control?

    Right answer: £220 per person. Civil Aviation Authority

  11. 11

    In the UK, until what age or height must a child normally use a child car seat?

    Right answer: until 12 years old or 135 cm tall, whichever comes first. GOV.UK

  12. 12

    Under the NHS routine childhood vaccination schedule in England, at what age does a child born in March 2025 get their second measles-containing vaccine dose?

    Right answer: at 18 months (1 year and 6 months), as the MMRV vaccine. NHS

  13. 13

    What is the highest temperature ever officially recorded in the UK, and where was it recorded?

    Right answer: 40.3°C at Coningsby, Lincolnshire, on 19 July 2022. Met Office

  14. 14

    In England, how often are women invited for NHS breast screening, and until which birthday do the routine invitations continue?

    Right answer: every 3 years, until the 71st birthday. NHS

  15. 15

    In the UK, after how many days off sick does an employee need to give their employer a fit note?

    Right answer: when ill for more than 7 days in a row (counting non-working days). GOV.UK

How it was run

Plans, dates and the second read

Each app's plan, model label, questions answered and capture days
App Plan Answered Search off Search on
ChatGPT Free 90 (15 questions) 15 Oct 2026 15 Oct 2026
Claude Free plan, Sonnet 5.5 Medium 90 (15 questions) 15 Oct 2026 15 Oct 2026
Copilot Free account, Auto 18 (3 questions) 15 Oct 2026 16 Oct 2026

Fresh chats, memory off; Copilot ran in Temporary chat, as it has no memory switch on that account.

195 of 198

The second read

Our own grader, an AI, marked every answer against the official page. A different AI model then graded them again, without being told which app or search setting each answer came from, and agreed on 195 of 198. It was not fully blind: many answers name their own app or say that search is off.

The 3 it read differently were each re-read against the key; all 3 were Claude with search on.

Claude: 45 of 45

Did it tell you it hadn't looked?

Every question asked the app to search. With search off it couldn't, and some answers said so in their own words: Claude in 45 of 45, ChatGPT in 31 of 45, Copilot in 1 of 9.

All 45 of ChatGPT's 45 search-off answers still had a link to an official page, with no search shown. A link in the answer doesn't prove the app looked.

Counted on each answer's own words, with the app's banner and buttons left out.

Six apps

Who lets you switch search off

ChatGPT, Claude and Copilot let a reader switch web search off and on, so they took part.

Gemini and Grok showed no search switch, and Perplexity showed no way to switch search off, so they sat this one out.

Limits

What this does not show

  • That search makes AI right or wrong in general. This is these 15 rules, these apps on free plans, on 15 and 16 Oct 2026.
  • Anything about reasoning or trick questions, where the tracker's runs find most wrong answers.
  • Anything about Gemini, Grok or Perplexity, or about paid plans.
  • A rate for Copilot: 3 questions is too few.
  • That search makes apps worse. Claude's measles answer is one question in one app.
  • Each answer is one of 3 tries, in an automated browser on one machine. Claude's and Copilot's two arms were captured hours or a day apart.

Read all 198 answers Every grade (CSV) How we test The tracker