Your Amazon data is ready for AI analysis when every settlement in the period has been split into its components and tied to a bank deposit, every SKU carries a landed cost under a method you can name, fees and ad spend and refunds sit against the SKU that caused them, inventory counts by location match the books, and the period you are asking about is defined by order date rather than deposit date. Miss one of those and the AI still answers. The answer is wrong in a way you cannot see.
The ten checks below are grouped into eight phases. Each one states what done looks like and the failure it prevents. Run them before you ask an AI a single question about margin.
Done looks like: a list of settlement IDs for the period, each with a posting in the ledger, none missing and none duplicated. Amazon's Seller Central page on sell.amazon.com describes the Payments dashboard as showing your current and previous settlement periods and recent payouts. Pull the list from there and match it to the ledger line by line.
Failure prevented: a month with three settlements booked in one system and two in another. An AI comparing that month to the prior one reports a 33 percent revenue swing that never happened.
Done looks like: the settlement total equals the deposit (after any reserve), and the total is broken into product sales, referral fees, fulfillment fees, refunds, reimbursements, storage, advertising invoices where deducted, and other charges. The sum of components equals the deposit to the cent.
Failure prevented: a lump "Amazon deposit" posted to revenue. The AI reads net cash as sales, understates revenue, and has no fee lines to analyze at all.
Done looks like: a unit cost for each SKU that includes purchase price, inbound freight, duty, and prep, computed under FIFO or another method you can name and defend to your CPA. The method is applied the same way in every period.
Failure prevented: margin computed on invoice price alone. A SKU that costs $5.10 to buy and $6.40 to land shows a margin 13 points better than reality on a $10 sale, and the AI recommends pushing ad spend into a product that cannot support it.
Done looks like: no PO with a zero freight line unless freight was billed elsewhere and allocated from there. Allocation basis (by weight, by value, by unit) is recorded.
Failure prevented: a newer, costlier layer of inventory that never enters COGS. The AI reports margin holding steady while the real number has been sliding for two POs.
Done looks like: referral fees, FBA fulfillment fees, and monthly storage tied to the SKU and order that generated them. A per-SKU fee total for the month that you can trace back to settlement lines.
Failure prevented: fee allocation by revenue share, which misstates every SKU whose weight or price differs from the catalog average. Illustrative numbers show the size of the error. Take three SKUs in a month:
The 15 percent referral rate is Amazon's published rate for Home and Kitchen on sell.amazon.com; the fulfillment figures are illustrative. Allocating the $21,850 by revenue share gives SKU-A 50 percent ($10,925), SKU-B 25 percent ($5,462.50), and SKU-C 25 percent ($5,462.50). SKU-C is understated by $987.50. With landed COGS of $6,000 on SKU-C, actual contribution is $15,000 minus $6,450 minus $6,000, or $2,550, a 17.0 percent margin. The allocated version says $3,537.50, or 23.6 percent. Six and a half points of margin invented by an allocation rule.
Done looks like: each campaign's spend for the period tied to the SKUs it advertised, with attributed sales alongside. Spend that cannot be attributed (brand campaigns, catch-all auto campaigns) is held in a named bucket, not spread by revenue.
Failure prevented: a SKU that looks profitable because its $1,800 of monthly ad spend is sitting in a company-wide advertising line. Ask the AI whether pausing ads on that SKU would hurt and it has nothing to work with.
Done looks like: refund amounts and refunded units by SKU for the period, with the fee credits Amazon returns on a refund netted against the fee lines they reverse. Reimbursements for lost or damaged inventory tied to the SKU and the case.
Failure prevented: a return spike hidden inside a total refund figure. A catalog-wide refund rate of 4 percent looks healthy while one SKU has gone to 11 percent and is losing money on every third unit.
Done looks like: units on hand at each FBA location, in transit, and in your own or 3PL warehouses, matching the inventory value on the balance sheet at landed cost. Differences are explained (reimbursed, disposed, in a removal order), not tolerated. The reconciliation procedure is in how to reconcile FBA inventory to your books.
Failure prevented: a "how much should I reorder" or "should I liquidate this" question answered against a count that is 300 units off. Amazon's FBA page on sell.amazon.com states that aged inventory charges apply monthly to items stored in a fulfillment center for more than 181 days; an AI cannot weigh that charge if it does not know how many units are sitting there.
Done looks like: one canonical SKU per product, mapped to every ASIN and FNSKU it has carried, including retired listings and merged variations. Parent and child relationships recorded. The same product on Shopify or Walmart maps to the same canonical SKU.
Failure prevented: one product appearing as three, each with a third of the sales and all of the fixed costs. The AI reports three underperformers instead of one healthy SKU.
Done looks like: a monthly view built from order-date transactions, with the last settlement of the month accrued so the month is complete before you compare it. A note on which report basis you used, so a week-over-week question and a month-over-month question are answered from the same definition.
Failure prevented: a settlement that closes on the 3rd of next month carrying the last four days of this month's sales into the wrong period. Every comparison that touches a month boundary shifts by a few percent, in different directions each month, and the AI chases noise.
ConnectBooks does the first seven checks as a matter of course. It reconciles each Amazon settlement into QuickBooks or Xero at the transaction level, posts each fee and refund to its own account against the originating SKU, carries landed cost per unit under FIFO, and tracks inventory by warehouse. The accounting overview covers the mechanics. That reconciled layer is what Crunch, the analytics AI inside ConnectBooks, reads when a seller asks which products were profitable last quarter but are losing money this quarter, or what a fee band change did to a SKU's margin.
Checks 3, 4, 8, and 9 still need a human. Crunch cannot know your freight allocation basis if nobody entered it, cannot count units in a warehouse it has no feed from, and cannot decide that two listings are the same product when the catalog says otherwise. It will compute on whatever cost is recorded, and if that cost is missing freight, the margin it reports is too high. The broader failure mode is covered in the data quality problem behind AI accounting tools. Run the ten checks, then ask.
Running an e-commerce business comes with plenty of challenges, but ConnectBooks is here to make your life easier. With real-time insights, seamless integrations, and detailed tracking of your profitability and inventory, you can stay ahead of the game. Whether you’re selling on Amazon, Shopify, Walmart, TikTok or eBay, ConnectBooks helps you manage your finances with 100% accuracy and confidence, so you can focus on growing your business.
Ready to level up? Start making smarter, data-driven decisions every step of the way. Try ConnectBooks Free Today or Schedule a Demo
