Test an AI data tool on linked sheets: free test set with hand-counted answers
Before you connect your team’s sheets to any AI tool, test it on data where you already know the answers. This page gives you three small linked sheets (orders, customers, regions) and the answers we counted by hand. Load them into datatomcp, ask your agent the questions below, and compare.
The set has one trap on purpose: order 31 belongs to customer C-011, who is missing from the customers sheet. A tool you can rely on should tell you about that row, not quietly put its revenue somewhere.
The three sheets
| Sheet | Rows | Columns | Link |
|---|---|---|---|
| orders.csv | 48 | order_id, order_date, customer_id, amount, status | customer_id → customers |
| customers.csv | 10 | customer_id, name, segment, region_id | region_id → regions |
| regions.csv | 3 | region_id, region, office |
Made-up data, small enough to check in any spreadsheet.
1. Load them
Sign in at app.datatomcp.com, open Add data, choose Paste a link and add the three links above, regions and customers first. You can also download the files and upload them.
datatomcp checks the columns the sheets share:
customers.region_id→regions.region_id: every value is found, so the link is confirmed.orders.customer_id→customers.customer_id: 10 of the 11 customer IDs are found, so the link is only suggested. Open the relationship map, look at it, and confirm it. Nothing crosses sheets until you do.
2. Ask your agent
Connect Claude, ChatGPT, Cursor or any MCP client (how), then ask these questions in plain words.
Paid revenue by region
| Region | Paid revenue | Paid orders |
|---|---|---|
| Southern Europe | 11,690 | 16 |
| North America | 9,930 | 16 |
| Western Europe | 2,490 | 7 |
| No region (order 31, customer C-011 not found) | 250 | 1 |
| Total | 24,360 | 40 |
The answer should come with a note that 1 order had no matching customer. In datatomcp that’s meta.unmatched in every result that crosses a link, and the agent is told to mention it.
Orders by customer segment (all 48 orders): SMB 21, Enterprise 14, Mid-market 12, no customer found 1.
Paid revenue in September 2026: 6,670 from 13 orders.
Refunded orders: 8, worth 6,040.
3. What to look for
- Same answer every time. Ask the same question twice, or from two different agents. A counted answer doesn’t change; one estimated from pasted rows can.
- The missing customer. A plain join drops order 31 or hides it in a blank group. The answer should say so.
- A way to check. Every datatomcp result has a link that opens the rows behind the number (
meta.view_url). - What went through the chat. Only the question and the answer: a few hundred tokens instead of 48 + 10 + 3 rows. On a real sheet with 50,000 rows that’s the difference between an answer and “file too large” (what pasting costs).
When the test set checks out, add your own sheets the same way. Deleting a dataset deletes our copy of it at once.