Worksheet · Measurement
AI answer recording worksheet
Use a repeatable worksheet for prompts, AI answers, source links and failed captures while keeping the original evidence.
Download worksheetOne row per recording
Give every attempt a unique identifier. Record the exact prompt, question set version, assistant, surface and date.
Save the full answer in your own records and put its file reference in the worksheet.
The surface is the place you asked, such as a consumer app, a search page or an API. Record it even when the provider is the same, because the same company can answer differently in different places.
Start each original prompt in a fresh conversation. Earlier messages can shape the reply. Put follow up answers in separate rows to preserve that context.
Keep an unavailable attempt with its reason. Never write no mention when no usable answer was saved, because that turns missing evidence into a result.
If you have to reuse a thread, say so in the notes. A reader can then set that row apart from the others.
Record observations separately
Keep a business name, a linked source and an accuracy concern distinct. One yes or no column cannot describe all three.
Don't paste private customer details, sign in tokens or confidential material into a shared file.
Use a short, fixed set of values for the mention column, such as yes, no and unknown. Free text like sort of or partly makes the column impossible to count.
Fill the mention column only for completed captures. A failed attempt has no answer to read, so its mention field stays empty.
A link can support one detail without supporting the whole answer. Use your notes to say which statement you checked.
Keep accuracy apart from whether you like the wording. An unflattering but accurate description is different from a wrong business fact.
Record how the answer described you in a short note, in your own words. Then mark whether each business fact in it matched your records, didn't match, or couldn't be checked.
Keep a comparison traceable
Give your prompt list a version. Add a capture status and a reason for each failed recording.
Add a separate page change record with the owner's publication date. Keep it apart from the capture date so a task cannot become an observation.
Name each prompt list version with its date, and never edit a version once recordings use it. A reworded prompt belongs in a new version with its own identifier.
When you correct a logging error, record the correction and keep the original observation where it helps explain an earlier report.
Store saved answers in a folder named by prompt list version and date, and don't edit them once saved. A file changed after the fact can't support the row that points to it.
If you take a screenshot, keep the date and the address bar in view where the surface allows it. Crop out personal details instead of the context.
Use the download
Start with the blank header file and keep an untouched copy. The worksheet provides a recording structure rather than collected results.
The template works on its own. It doesn't connect to assistants, analytics or your website.
Add columns if your review needs them, but keep the originals in the same order. Anyone comparing two worksheets can then match them column by column.
Fill in each column the same way
The download has one column for each fact a later reader needs. Agree what goes in each one before the first recording, and write that down beside the file.
The question set version and question identifier tie each row to the exact wording. The question column repeats that wording, so a reader doesn't need a second file to follow it.
The assistant and surface columns say where you asked. The country and language columns say the context you set or could see. Leave either blank and add a note when you couldn't tell.
The capture status column holds a short, fixed value, such as completed or unavailable. Put the reason for a failure in the source notes, not in the status itself.
The citation addresses column holds the full links the answer gave. The answer file column points to the saved copy of the whole response.
The last three columns belong to follow up work. They hold the page change date, the person who owns the next step and the date of the next review.
Write every date and time one way
Mixed date formats cause quiet errors. The same row can read as two different days to two different people.
Write dates with the year first, then the month, then the day. That order sorts correctly in any spreadsheet and leaves no doubt about which number is the month.
The internet standard for timestamps recommends stating how a time relates to Coordinated Universal Time. It notes that local daylight saving rules are complex and can change.
That matters in the UK, where clocks move twice a year. A capture time with its offset stays clear after the clocks change.
Keep the capture date and the page change date in separate columns, both in the same format. You can then sort the rows and see which answers came before the change.
Note the request settings for API captures
An API capture comes from a request someone wrote, and that request has settings. Two requests with the same prompt but different settings aren't the same observation.
Record whether web search was switched on. OpenAI and Anthropic both offer web search as a tool the developer adds to a request.
Anthropic says the model decides when to search based on the prompt. It may answer directly, without searching, when the question draws on stable knowledge.
Record any location sent with the request. OpenAI's guide says its web search tool accepts an approximate user location made of country, city, region and time zone.
Record any list of allowed or blocked websites as well. Both providers document domain filters that limit which sites a search may use, and a filtered answer isn't a general one.
Put these settings in the source notes, or in extra columns if you use them often. Someone reading the row later can't see the request, only what you wrote down.
Record sources the way the assistant gave them
A source link in an app and a source field in an API response are different records. Note which kind you saved, because they aren't always equivalent.
OpenAI publishes a guide to its web search tool for developers. It says responses include inline citations by default, each with a structured record of the page address and title.
The same guide describes a separate sources field listing every address the model consulted. It says that list is often longer than the list of cited pages.
Anthropic's web search documentation says citations are always enabled for its web search tool. Each citation carries the address, the title and a short extract of the cited text.
Keep cited addresses and consulted addresses in separate columns or notes. A page the model read but didn't cite is a different observation from a page it linked.
In an app, copy the link addresses exactly as shown. Don't follow a redirect and save the final page instead, unless you note both.
Keep the file readable in any spreadsheet
Prompts and notes often contain commas, quotation marks and line breaks. The common CSV format says fields containing any of these should be wrapped in double quotes.
Let your spreadsheet app write the file rather than editing it in a plain text editor. One missing quotation mark can shift every later field in that row.
Keep long answers out of the worksheet itself. Save each one as its own file and put the file name in the answer file column.
After you reopen a saved copy, check a few rows by eye. Make sure the dates, identifiers and links still read the way you entered them.
Keep one master copy and work on dated copies of it. If two people edit the same file at once, one person's rows can be lost without warning.
Share the minimum useful record
Share only the context needed to assess the finding. Leave credentials and unrelated personal information out of the review copy.
The downloadable CSV holds column headings only. Make a working copy and enter your own recordings.
Give a reader the prompt set and the method alongside any summary. A spreadsheet total without its coverage can't explain what was checked.
Decide how long you need to keep the full answers, and delete what you no longer need. Keep the worksheet rows that a published summary relies on.
Check the log before you summarise it
A summary is only as good as the rows under it. Before you count anything, run a few simple checks on the whole file.
Look for repeated identifiers, completed rows with a blank mention field and failed rows with a mention filled in. Each one is a recording error to fix or explain.
Check that every row in one comparison uses the same prompt list version. A row from an older version belongs in its own group.
Then count mentions over completed captures only, and state the coverage beside the figure. Write how many attempts there were and how many produced a readable answer.
Whatever tool you use, the rule is the same. A figure without its coverage, method and prompt list can't tell a reader what was checked.
A worked example
A review has one answer and one failed attempt.
| Detail | Observation | Review or next step |
|---|---|---|
| Completed recording | The full answer and citation address are saved. | Record the prompt version and capture method. |
| Failed recording | No usable response came back. | Set capture status to unavailable and say why. |
Keep both rows. Only the completed recording counts towards answer-based totals.
Common questions
Can I combine two similar answers into one row?
Keep each recording separate. Similar wording doesn't make two attempts the same observation. A summary can refer to both rows without replacing them.
What should I do with a failed attempt?
Keep the attempt, its date and the reason it failed. Mark its capture status unavailable. It can't count towards mention totals because there's no completed answer to look at.
Can I combine two similar answers into one row?
Keep each recording separate. Similar wording doesn't make two attempts the same observation. A summary can refer to both rows without replacing them.
What should I do with a failed attempt?
Keep the attempt, its date and the reason it failed. Mark its capture status unavailable. It can't count towards mention totals because there's no completed answer to look at.
Check your understanding
How should you record a blocked capture?
Read the answer and explanation
- Missing, with the reason retained.
- A readable answer with no mention.
- A successful customer result.
Answer. Missing, with the reason retained.
No readable answer means no observed mention and no observed absence. Show capture coverage beside any rate.
Answer and explanation
Missing, with the reason retained.
No readable answer means no observed mention and no observed absence. Show capture coverage beside any rate.
Your next steps
Sources and method
A Carqo review method. Provider statements link to the relevant documentation.
- IETF, Date and Time on the Internet, Timestamps Link checked 29 October 2026
- OpenAI API documentation, Web search Link checked 4 November 2026
- Claude API documentation, Web search tool Link checked 4 November 2026
- IETF, Common Format and MIME Type for Comma-Separated Values Files Link checked 29 October 2026