How I test AI answers
The rules I follow when I put ChatGPT, Gemini and Claude up against each other.
Why I compare AI answers
More and more people ask AI about exercise, food, sleep and health. A smooth answer can still skip safety context, lean on weak evidence or sound far more certain than the research is. My comparisons make those differences visible. I'm not trying to crown a permanent winner.
What every test must show
Each test records the product and model name, the mode or tier when it matters, the test date, the exact prompt, and whether web search or other tools were on. Every model gets the same core question. Any follow-up question is shown or clearly summarised.
I separate the model name the provider shows from guesses about what runs underneath. If the interface doesn't say which exact version it is, I say that.
How I judge an answer
I check whether it:
- answers the actual question,
- separates general information from personal medical advice,
- flags the important risks and exceptions,
- is honest about uncertainty,
- gives sources that exist and support the claim,
- avoids invented numbers or citations,
- gives practical, proportionate next steps,
- corrects itself when shown contrary evidence.
Scores, when I use them, follow criteria written in the article. They describe that prompt on that date, not the permanent quality of a model.
Fact-checking and limits
Important claims are checked against official public-health guidance, recognised professional organisations or peer-reviewed research. I list my own verification sources separately from the ones the AI suggested.
AI systems change all the time and can answer the same prompt differently. A test is a documented snapshot, not a lab trial and not medical validation. I never invent outputs, quietly improve an answer, or claim to have tested a model I couldn't access.
Corrections and money
If I get something wrong, I correct it on the post and say what changed. Free access, subscriptions, sponsorships or any other relationship with an AI company must be disclosed in the comparison. Money does not buy a better result.