• Open ChatGPT, Claude, Microsoft Copilot, Gemini, Grok, and Perplexity in separate tabs.
• Use a personal consumer account and select the normal free plan or free model shown in the product. Do not start a trial or upgrade.
• Record the exact plan or tier label, model or mode shown, and date tested. If a label is not visible, write Unknown or Not shown instead of guessing.
• Start one scorecard row for each service. You will add search, upload, result, and limit observations as you complete the guide.
• Use only the fictional practice information in this guide. Do not enter payment details, connect outside apps, or upload private files.

Use the exact labels shown in the product. Some services call no-cost access Free, Standard, or something else. Features and limits can change, so the date and model or mode make your scorecard easier to understand later.

• You cannot find a plan label: Record Unknown and capture the account state or usage message you can see. Do not convert an unknown tier into a free result.
• The model or mode is hidden: Record Not shown. Do not infer it from an old article or another person's screenshot.
• The product offers an upgrade or trial: Decline it and return to the normal free experience before running the test.
• The service is unavailable in your region: Mark availability No and the scored tasks N/A. Do not use a VPN or a different account to hide the limitation.

• Open a fresh chat in each service you recorded in Step 1.
• Use the ordinary chat entry point and paste the same prompt below into every service.
• Do not manually turn on web search, research, file analysis, or another special mode for this test. If a service searches or invokes a tool automatically, let it finish and record Automatic rather than hiding the behavior.
• Save the answer with the service, displayed free tier, model or mode, and test date.
• Check the public hours, staffing, budget, capacity, missing details, and flyer conflict before judging the writing style.

Copy this comparison prompt into each assistant:

Help me prepare a short public update and an internal run sheet for a fictional community event. Use only the facts below.

EVENT
Name: Northside Repair Café
Date: Saturday, November 21, 2026
Public hours: 10:00 a.m. to 1:00 p.m.
Venue access: 9:00 a.m. to 2:00 p.m.
Capacity: 40 people
Registrations: 34 people
Volunteers: 6
Stations: Small electronics, clothing and textiles, household items
Staffing: Two volunteers at each station

APPROVED BUDGET
Refreshments: $120
Replacement parts: $90
Signage: $35
Budget limit: $300

SOURCE CONFLICT
The approved run sheet says doors open at 10:00 a.m. A draft flyer says doors open at 9:30 a.m. The approved run sheet controls. Flag the conflict so the draft flyer can be corrected.

MISSING INFORMATION
No registration URL, street address, parking details, or contact email has been supplied. Do not invent them.

RETURN
1. A public event blurb of no more than 100 words.
2. An internal run sheet with setup, public hours, station staffing, and cleanup.
3. The budget total and remaining amount, with the arithmetic shown.
4. The number of spaces left and the percentage of capacity already registered.
5. A short list of missing details and the flyer correction needed.

Use clear, plain language. Keep public copy separate from internal notes. Do not add claims, dates, links, prices, or services that are not in the brief.

The prompt contains the complete fictional practice brief, so every reader can run this test without downloading a file.

Check the answer against these facts:

• Public hours are 10:00 a.m. to 1:00 p.m.
• Venue access runs from 9:00 a.m. to 2:00 p.m.
• Two volunteers staff each of the three stations.
• Approved spending totals $245, leaving $55.
• Six spaces remain. Thirty-four of 40 places is 85%.
• The draft flyer must change from 9:30 a.m. to 10:00 a.m.
• The address, registration URL, parking details, and contact email must remain missing.

A polished answer still needs correction if it invents a link, changes the opening time, or mixes internal setup details into public copy.

In this sample Sonnet 5.5 response, the event time is right, but "registration is open" and "Details on how to register will be shared soon" are not supplied facts. Remove those claims before using the blurb. Good writing and accurate writing are separate checks.

This ChatGPT result gets the capacity calculation right: 40 minus 34 leaves six spaces, and 34 divided by 40 is 85%. Check the rest of an answer with the same care. These are sample runs, so your wording and layout may differ.

• The service searched automatically: Record web search used as Automatic. Keep the result, but do not describe it as a no-search answer.
• The response stops because of a usage limit: Record the exact message and time. Do not upgrade or start a trial to finish the test.
• The service asks for missing details: Answer only with facts already in the prompt. If it requires invented details to continue, score that behavior accordingly.
• The response is cut off: Record it as incomplete. If you retry, give every service the same one-retry rule and keep the original result.

• Open a new chat in each service from Step 1.
• Use the product's normal web-search control if one is available. Record web search availability as Yes, No, or N/A. If search starts automatically, add that detail to Notes.
• Paste the same research prompt below.
• Open the cited page yourself. Confirm that it is the National Park Service page, not a travel blog, search summary, forum, or social post.
• Score the answer against the official page as it appears on your test date. Mark source quality N/A when web search is unavailable.

Copy this research prompt:

Use web search to answer this question from official sources only:

Can I bring a dog on hiking trails in Shenandoah National Park?

Find the National Park Service's current Shenandoah pet rules. Tell me:
- The leash length and type required.
- Whether pets are allowed on most trails.
- The names of trails where pets are prohibited.
- Any important exception for the Stony Man route.

Link the official NPS page. Quote or closely identify the wording that supports each rule. If you cannot find an official source or cannot browse in this account, say so. Do not fill gaps from memory, travel blogs, forums, or old social posts.

Use the current Shenandoah National Park pet page as the answer key. As checked on November 4, 2026, it requires a physical leash no longer than six feet and allows pets on most trails.

The listed pet-restricted trails are Fox Hollow Trail, Stony Man Trail, Limberlost Trail, Post Office Junction to Old Rag Shelter, Old Rag Ridge Trail, Old Rag Saddle Trail, Ridge Access Trail (Old Rag area), Dark Hollow Falls Trail, Story of the Forest Trail, Bearfence Mountain Trail, and Frazier Discovery Trail. At the beginning of the Stony Man route, dogs may use the Appalachian Trail section, but they may not turn onto the blue-blazed Stony Man Trail to the summit.

The live official page controls if this list or wording changes. Record the test date. Source quality and answer accuracy are separate: an answer can link the correct page but still omit a trail or misstate the exception.

This Gemini answer found the official NPS page and the leash rule. It also cited non-NPS sources elsewhere in the answer, despite the official-sources-only request. Open the citations before giving source quality a full score.

• Web search is unavailable: Mark search availability No and source quality N/A. Do not paste the NPS page into the chat and then claim the product found it.
• The answer cites a blog or search summary: Open the links and score the source quality based on what the assistant actually used.
• The citation opens the right domain but not the supporting page: Ask where the rule appears, but keep the original result for scoring.
• The NPS page has changed: Use the current page as the answer key and record the date. Do not score against a remembered trail list.

• Use only the fictional CSV below. Do not upload a real budget, invoice, customer list, or company file.
• In each service, look for an attachment or upload control and record Yes, No, or N/A for file-upload availability.
• Where upload is available, upload repair-cafe-budget.csv and use the analysis prompt.
• Where upload is unavailable, paste the CSV text and mark upload availability No. This tests calculation from pasted data, not file handling.
• Score CSV correctness separately from upload availability. Check the total, remaining budget, and every quantity multiplied by unit cost.

CSV content if you need to paste it:

expense,quantity,unit_cost,approved_total
Refreshments,40,3.00,120.00
Replacement parts,15,6.00,90.00
Signage,5,7.00,35.00

Copy this file-analysis prompt:

Analyze the attached repair-cafe-budget.csv. If no file is attached, use the CSV pasted above.

Calculate the sum of approved_total, compare it with the $300 budget limit, and report the remaining amount. Check whether each approved_total equals quantity multiplied by unit_cost.

Show the calculation. Do not add tax, fees, or unlisted expenses. If you cannot read the attachment, say so and ask me to paste the data. Do not guess.

The expected approved total is $245, and the remaining budget is $55. Each row matches quantity multiplied by unit cost:

• 40 times $3 equals $120.
• 15 times $6 equals $90.
• 5 times $7 equals $35.

A correct answer from pasted CSV text does not prove that the free tier accepts file uploads. A visible upload control does not prove that the service read the file correctly. Keep those observations separate.

In this sample, Sonnet 5.5 read the attached CSV, checked each multiplication, and returned the expected total and remaining budget. The workbook's Budget Answer Key lets you check those numbers yourself.

• The upload control is missing: Mark file upload No, paste the CSV, and continue with the calculation test.
• The service says it cannot read the attachment: Mark file upload or file reading No for that run. Do not score a guessed answer as file analysis.
• The service calculates a different total: Check each row multiplication and the sum. Keep the original answer so you can score the error.
• The service adds tax, fees, or another expense: Mark the unsupported addition as an accuracy problem.

• Open the scorecard workbook from Resources if you have access. Otherwise, create a sheet using the copyable columns below.
• Use one row for each tool tested on its normal free plan.
• Enter Yes, No, or N/A for web search and file upload availability. Score instruction following, accuracy, source quality, CSV calculation, and useful output with 0, 1, 2, or N/A.
• Add the account or plan, model or mode, test date, and any limit message you observed. Put extra context, such as region or automatic search behavior, in Notes.
• Use Chosen role to mark a tool Primary assistant, Research tool, or Backup, or leave it blank. Choose from your results rather than declaring one universal winner.

Copy these scorecard columns:

Tool
Account / plan
Model / mode
Date tested
Instruction following
Accuracy
Source quality
Web search available?
File upload available?
CSV calculation
Useful output
Observed limits
Notes
Chosen role

Use this scoring scale:

• 2: Passed without a meaningful correction.
• 1: Usable after an important correction.
• 0: Incorrect, unsupported, or invented.
• N/A: The feature was unavailable or the task was not tested.

You do not need one overall winner. A tool may be your best writing choice while another is better for research or files. Leave the role blank if it has not earned a place in your toolkit. Save the date so you can repeat the comparison after products change.

The downloadable scorecard starts blank. Fill it with your own results; the Budget Answer Key tab already contains the practice calculations.

Keep the tools that earned a useful role in your scorecard. You do not need to use all six.

• The model name is hidden: Write Not shown. Do not infer a model from old documentation or another person's screenshot.
• A feature was unavailable: Mark availability No and use N/A for the related score. Do not turn N/A into zero.
• You skipped a test: Use N/A and explain why in Notes.
• The scores are tied: Compare the actual errors, source quality, and limits. You can leave Chosen role blank.