Prepare the trial before opening the chat
Choose material you know well and are permitted to share. Write down what a correct result must contain, how long the output should be and what the assistant must not assume. Save the original prompt so each candidate receives the same task.
Record the product, account tier, mode and enabled tools. If one assistant searches the web and another does not, the trial includes a product-access difference. That is useful information as long as you do not describe it as a pure model comparison.
Use these five tasks
Keep the trial small enough that you can inspect every result. Replace the examples with your own work while keeping the purpose of each task.
| Task | Example brief | Check |
|---|---|---|
| Summary | Summarize a supplied report in five points | Important exceptions remain visible |
| Current fact | Answer a dated question and cite sources | Sources support the exact claim |
| Revision | Shorten a draft without changing facts | Requested change happens without factual drift |
| Reasoning | Compare options under stated constraints | Conclusion follows the constraints |
| Uncertainty | Ask a question the source cannot answer | Assistant identifies the missing evidence |
Make the revision task difficult enough
Ask for a change that can expose unwanted edits: a new audience, a shorter length or removal of one unsupported claim. Compare the revised answer with both the first version and the source.
A useful assistant should preserve approved facts while changing the requested part. If each revision introduces a new issue, include that time in the result. Do not reward a polished tone while ignoring the effort required to keep the content correct.
Inspect sources and unknowns
Open cited pages rather than counting citations. Check dates, scope and whether the evidence applies to the exact question. A source about an announced feature does not prove that feature is available to your account.
For the uncertainty task, deliberately leave out a key detail. The assistant should not guess a deadline, price or owner merely to produce a complete-looking answer. Asking a useful clarifying question or labeling an unknown can be a strong result.
Keep a simple decision log
For each task, record pass or fail on the acceptance check, minutes spent correcting it and whether you would use the output. Add interruptions caused by limits or missing tools. Weight the results by how often that task occurs in your life.
Choose after several normal tasks rather than a single favorite answer. If the products are close, prefer the workflow with simpler accounts, exports and maintenance. Revisit the trial when your work changes instead of chasing every release announcement.
Frequently asked questions
How many prompts do I need to compare chatbots?
Five representative tasks can support a practical personal decision, but they are not enough for a general scientific ranking.
Should the prompts be identical?
For a controlled comparison, yes. You can also run a separate workflow trial using each product’s tools, but record the differences.
What is the most important metric?
Correct and useful output after review. Correction time, missing information and workflow friction often matter more than the style of the first answer.
Sources & update notes
Product facts were checked against the official pages below on 4 November 2026. Workflow advice is editorial guidance, not a claim of hands-on benchmarking. Availability, pricing and limits may change after this date.