ChatGPT vs. Gemini vs. Claude: Compare Them Using Your Everyday Tasks

1. Draft an email with real constraints

Create a fictional customer email about a delayed delivery. Give all three assistants identical facts: the package is two days late, tracking has not updated, and a replacement requires manager approval. Request a reply under 120 words that acknowledges the inconvenience without promising a refund, replacement or delivery date.

Compare the drafts without looking at their brand names. A friendly message that invents a promise fails, regardless of style. Then ask, “Make it warmer without adding commitments.” This second turn tests whether the assistant preserves constraints while revising, a more useful distinction than which initial draft sounds most impressive.

Watch for invented operational steps, too. “I have contacted the carrier” asserts an action nobody supplied. Mark that as a factual error even when the draft avoids prohibited promises and otherwise sounds ready to send.

• Check every statement against the supplied facts.
• Count words and remove empty reassurance.
• Record edits needed before you would send it.

2. Find answers inside a document

Use a public handbook or a sanitized PDF of roughly 15–30 pages. Pick three questions whose answers you have already located: an eligibility rule, an exception and a deadline. Ask each assistant to answer only from the document, quote supporting language and identify the page or section. Add one question the document cannot answer to test whether it admits the gap.

Open every cited location yourself. Viewer page numbers may differ from printed page numbers, so accept a clearly identified section when appropriate. A correct answer with an invented citation is still a failure. If your file contains scanned pages, include those in the test; readable text and reliable navigation are separate capabilities. Record upload restrictions rather than assuming every tier handles the same files.

Choose an exception that modifies a rule elsewhere in the file. This exposes answers built from one matching passage rather than the complete policy. Your answer key should identify both passages and explain their relationship.

3. Ask questions about an image

Photograph a fictional event flyer beside a few household objects. Include a small-print deadline and one deliberately blurry detail. Upload the same original image to each service and ask it to extract the event information, describe the objects and distinguish visible facts from uncertain interpretations.

Reward an assistant that says it cannot read the blurry detail. Confident guessing is not better image understanding. Check whether it confuses an object's appearance with its identity or misses a qualifying note. Avoid medical images, identity documents and photographs exposing addresses or children's information.

Specify an output format: event name, date, location, deadline and unreadable fields. Leaving a field blank with an explanation is preferable to completing it from context. This also makes omissions easier to compare.

• Compare extracted details with the original photograph.
• Ask which answers are uncertain and why.
• Reject unsupported conclusions about people or sensitive traits.

4. Build a plan that survives changes

Ask for a Saturday outing in your city for two adults and a child, with a $100 total budget, no car and a required return home by 5 p.m. Provide a starting neighborhood, accessibility needs and meal preferences. Give every assistant the same permission to search the web, if available.

Then introduce rain and a one-hour delay. Check whether the revised plan still meets every constraint instead of merely swapping one attraction. Verify hours and admission on official venue sites, and routes with the local transit agency. Treat unsupported prices and travel times as estimates, not confirmed facts.

Require a timed itinerary with a running cost total. Recalculate the arithmetic yourself, including whether quoted admission is per person or per group. A plausible schedule can conceal a budget overrun or an impossible return journey.

• Include fares, food, admission and a budget cushion.
• Check transfer times and realistic walking distances.
• Require a fallback if a venue closes.

5. Test connections without surrendering control

Connections can decide the contest when answers alone are similar. Use a test calendar, harmless document or dummy email account. Check each provider's official help center for current app support, US availability, subscription requirements and administrator restrictions. Product names and permissions change; an advertised integration may not perform your intended action.

Plant two similarly named test documents with different dates and ask for the newer one. Require a source link, not just a summary. This checks whether the connection retrieves the intended record rather than a convincing substitute.

Test retrieval first, then a reversible action such as creating an event on the test calendar. Confirm it actually appeared. Review permissions and disconnect afterward. Add your five scores, but let a serious privacy concern or failed must-have task override the total. Choose demonstrated usefulness, not an ecosystem promise.