AgentXray How we decide Blog

What the measurement caught

Findings from our own Amazon account, each with the date it was measured. Two of them are our own mistakes, which is the point: a tool that has never been wrong in public has either never checked or is not telling you.

Not included anywhere on this page: what we pay our supplier, or anything you could derive it from. The fees below are Amazon's numbers about our account. What we paid for the goods is not.

A referral rate that more than doubled, mid-month

What is usually assumed: Most tools take the referral fee as a fixed category percentage and multiply. It is the one input everybody treats as a constant, because Amazon publishes it as one.

What we measured: Splitting the same settlement table by posting date shows two different regimes on the same product, on the same account, weeks apart.

PeriodOrdersUnitsReferralPer unit
Before 15 Jul 2026751035.63%$1.93
From 15 Jul 20269511912.61%$4.41

Why it costs money: Break-even ACoS moved from 39.1% to 35.2%. Every bid, coupon and price decision made against the old figure was made against a number that had already stopped being true — and nothing in an advertising dashboard reports it, because the deduction does not happen there.

Measured 2026-08-18 · 222 units across 170 settled orders, amazon.com.au

The fee estimate is not what the settlement charges

What is usually assumed: Amazon publishes a fee estimate per ASIN, and a profit tool that reads it looks like it is reading Amazon's own number.

What we measured: On our own product the FBA fee the settlement deducted was $10.85 against a $10.16 estimate — and on amazon.com.au every fee carries 10% GST that the estimate excludes entirely.

Why it costs money: A break-even built on estimates read $18.94 when the settled number was $20.17. That gap is wider than the margin most sellers are optimising toward, which means the optimisation was pointed at the wrong target the whole time.

Measured 2026-08-05 · 30-day settlement window, 88 units, amazon.com.au

We got our own number wrong. Twice.

What is usually assumed: That a tool built by people who measure things carefully gets the measurement right the first time.

What we measured: Our published break-even went $18.94 → $20.17 → $20.88 across three corrections. The first came from the settlement backfill, the second from consumption tax on fees, the third from the referral shift above.

Why it costs money: The reason to say this out loud rather than quietly restate the latest figure: a tool that has never been wrong in public has either never checked or is not telling you. Ours reports the correction because the correction is the product working.

Measured 2026-08-18 · internal measurement history, amazon.com.au

The bug that manufactures a plausible wrong answer

What is usually assumed: That reading settlement line items is simply better, and getting it right is a matter of having the data.

What we measured: Filtering settlement rows on the description alone matches both the item-price row and the negative promotion row. That double-counts units, shrinks the revenue denominator, and produces a confident blended rate of about 9.6% that is not any rate the account was ever charged.

Why it costs money: It also manufactured a phantom defect — dozens of orders appearing to disagree about quantity. The fix is to filter on the amount type and the description together; with that done, item-price units equal fee units on every settled order. We publish this because anyone building the same thing will hit it, and because a number being specific is not evidence that it is right.

Measured 2026-08-18 · settlement line-item reconciliation, amazon.com.au

The limitation of this page

Nothing here updates itself. The engine detects an unstable referral rate on a rolling window and warns a connected seller in real time, but no table records the event — so a shift that has scrolled out of the window cannot be recovered later. Persisting those detections is the next piece of work. Until then these are case studies, not a feed, and saying so is cheaper than being caught pretending otherwise.