I’m trying to compare answers from OpenAI, Claude, Grok, Perplexity, and DeepSeek on the same research question, but the comparison feels uneven. Perplexity may search the web while another assistant answers from its model, and each product has different tools and default settings.

For a useful test, would you keep the prompt identical and turn browsing off wherever possible, or let each assistant use its strongest features? I’m also curious what you measure beyond whether an answer “sounds right”—for example, factual accuracy, useful sources, or how much checking the answer requires.

Has anyone settled on a practical model comparison rubric that accounts for these differences without turning the test into a week-long project?