Audit an authorized Amazon CSV with row-grain checks, synthetic subtotal math, missing-data handling and a source reconciliation.
Start with a deliberately small sum
week_start,marketplace,asin,units,ordered_sales_usd
2026-09-06,US,EXAMPLE-A,4,120.00
2026-09-06,US,EXAMPLE-B,2,80.00
2026-09-06,US,Total,6,200.00This file is invented, not an Amazon report schema. The two ASIN rows total 6 units and USD 200.00; the Total row repeats that amount as a check. Add all three sales cells and you get USD 400.00, a wrong answer that sounds precise. These selected ASINs are not the whole account, and sales are not profit.
Ask Claude for the rows it used and the formula, not just the final number. If it says USD 400.00, halt the analysis. The subtotal was counted as a detail row. Check the input before trusting any subsequent comparison. Reproduce USD 200.00 from the same filtered rows in the original report or an independent spreadsheet pivot. Only after the totals match should prose about change begin.
The attached file is an authorized, minimized snapshot, not a live Amazon feed. Its grain is one ASIN per completed US marketplace week; currency is USD and ordered_sales_usd is additive across these distinct ASIN rows. For the selected week, list included rows, count rows, sum units and ordered_sales_usd, and show each formula. Flag missing dates, duplicate ASIN-week keys, mixed marketplaces, blank numeric fields and total rows. If a needed column is absent, say so. Do not infer profit, ad return or the whole account from this subset.
Protect the raw file; inspect the converted copy
An authorized export has a report title, account, marketplace, reporting dates, filters and download time. Keep that download unchanged. When converting a spreadsheet to CSV, compare headers and row count before and after; Amazon notes that report fields and values can change. A column renamed ‘sales’ can silently erase the distinction between ordered and shipped measures.
For a weekly version of this check, connect Claude to 1stPage and request seller_analytics with asin_sales_ranking for the same seller account, marketplace and period. Compare its ASIN rows with the USD 200.00 example only if the filters and grain match. The MCP connection does not convert an uploaded file into fresh Amazon data; it asks 1stPage for the data that account actually holds.
Use a reduced copy that your organization permits you to share. Remove buyer identities, order identifiers and unrelated columns unless the approved task needs them; never include passwords or tokens. SKU and ASIN can be business-sensitive too. Anthropic offers chat uploads and persistent project files, so choose the file location under your organization's rules and Anthropic's current privacy terms. Deleting a local file is not a service-retention policy.
Write down separately how many rows were removed for privacy and how many were excluded by analytic filters. Otherwise a perfectly calculated subtotal of a reduced file can be mistaken for the original report's total.
Find rows that change the meaning of the number
Tell Claude whether a row is one order item, one ASIN for a week, or an inventory snapshot. Order-item files can repeat an order ID; ASIN-week files should not silently repeat the same ASIN-week key; end-of-day stock figures cannot be added across dates as units sold. Remove Total rows from detail aggregation, but retain the source total as a check.
Look for a blank date, mixed US and Canadian stores, currency columns without a common unit, text masquerading as numbers, and an incomplete last day. Have Claude list the offending keys and rows. Then inspect them yourself in the source. A blank numeric cell is unknown unless the report definition says blank means zero. This is not a contest to make the CSV pass; it is a test of what claim its rows permit.
Ask the question the columns can actually answer
A sales-only file cannot determine whether advertising caused a dip. It lacks spend, clicks and attribution rules. Ordered sales without fees, cost of goods and refunds cannot establish margin. The synthetic USD 200 subset cannot describe an entire seller account. Put the unsupported claim in the prompt and ask which missing fields would be needed to test it.
Now imagine EXAMPLE-B has an empty ordered_sales_usd in the next week's extract. Inserting zero would manufacture a decline. Check whether the original row exists, whether the period finished and whether the filters changed. Until the omission rules are known, say ‘not observed in this extract.’ That answer is less exciting than a stockout story and much more useful to a person deciding whether to replenish.
End with a reproducible snapshot, not a live-feed claim
Save the report title, source store, dates, copied file name, row count, exclusions and independently verified subtotal with the answer. Repeat the row and subtotal checks for each new export; an attachment does not refresh itself. Amazon's announced US-store seller plugin beta is a different, permissioned route, not an upgrade automatically applied to this CSV.
If the task recurs, name the person responsible for exporting and the permission review before looking for an approved connection. Test the connection's actual field coverage if one is chosen. Keep this particular answer tied to this particular file.
Sources and boundaries
Check the current permissions and report definitions before acting on an answer. Example prompts are questions, not live account results.