📋 Overview
Most sellers either skip their weekly review or spend two hours rebuilding the same spreadsheet and still end the session without a decision. An AI assistant can compress the reading and comparing work, but only if you feed it the same reports on the same schedule and hold it to a strict output format. This article covers which exports to standardize on, a reusable review prompt, what to check before acting, and where weekly AI reviews commonly go wrong.
🗂️ Freeze a Fixed Report Pack
A weekly review is only useful if this week’s numbers are comparable to last week’s. That means the same reports, the same date windows, and the same metric definitions every time. Pick your pack once and stop changing it.
A workable pack for most private-label sellers, all available in Seller Central:
- Business Reports — sessions, Detail Page Views, units, and Unit Session Percentage at the child-ASIN level. This is your traffic and conversion baseline.
- Sponsored ads reports (under Measurement & Reporting in the advertising console) — a campaign-level and a search-term-level export for the same window.
- Account Health — your current Order Defect Rate, Late Shipment Rate, Pre-fulfillment Cancellation Rate, and any active policy notifications.
- Manage Inventory or the Inventory Ledger Report (Reports > Fulfillment > Inventory Ledger) — available units, inbound, and unexplained movement.
- Voice of the Customer — returns and complaint themes by ASIN.
Add one file that Amazon does not give you: a small cost table listing each SKU’s landed unit cost, current FBA fulfillment fee from Fee Preview, and your referral-fee percentage for that category. Referral fees run from 5% to 45% depending on category, so look up your own rather than assuming a single rate. Without this table, any AI comment about profitability is arithmetic on missing inputs.
Use the same window length each week and align it to the same weekdays. A seven-day window that starts on Monday every week is comparable; a “last 7 days” pull taken on Wednesday one week and Saturday the next is not.
🔁 Build the Review Loop in Five Steps
1. Decide the questions the review has to answer
Write down four to six questions before you involve any tool. Typical ones: Did unit volume change, and was it traffic or conversion? Which SKUs moved outside their normal ad-efficiency range? What will stock out or go long-term-storage-aged before the next review? Did any account health metric move toward its threshold? Which return or complaint theme is growing?
These questions decide your report pack, not the other way round. If a report does not answer one of them, drop it from the pack.
2. Write a definitions block you reuse every week
AI assistants will happily invent their own version of “conversion rate” or “ad efficiency” if you leave the definition open. Keep a short block of text that states exactly how each metric is calculated in your business, and paste it into every session:
- Conversion = Unit Session Percentage from Business Reports, child-ASIN level, not parent.
- Contribution margin = net revenue minus landed cost, FBA fulfillment fee, referral fee, and ad spend for that SKU.
- Ad efficiency = ad spend divided by total SKU revenue, not just ad-attributed revenue.
Definitions that drift break your week-over-week comparison more quietly than bad data does.
3. Use an exception-based prompt, not a “summarize this” prompt
Asking for a summary gets you a restatement of numbers you already have. Ask instead for variances against a threshold you set. A reusable pattern:
You are reviewing weekly performance data for my Amazon business. Attached: Business Report (this week and prior week), campaign report, search term report, inventory snapshot, and my SKU cost table. Use the metric definitions below exactly as written; do not substitute your own.
Report only exceptions. An exception is any SKU where units, sessions, or Unit Session Percentage moved more than 15% against the prior week, any campaign where spend moved more than 20%, or any SKU with under three weeks of cover. Ignore everything inside those bands.
For each exception, output: SKU or campaign, the metric that moved, both weeks’ values, the arithmetic you used, the two or three most plausible explanations, and which report or Seller Central page I should open to confirm which explanation is right.
List separately any figure you could not calculate because the data was missing. Do not estimate or fill gaps.
The 15% and 20% figures are starting points you set yourself, not benchmarks — tighten them if the output is noisy, loosen them if nothing ever surfaces. The important parts are the instruction to ignore stable SKUs and the instruction to declare gaps instead of filling them.
Also ask for explanations rather than conclusions. An assistant looking at a conversion drop cannot see that a competitor launched a coupon or that your main image changed; it can only tell you the drop is on the conversion side rather than the traffic side, which is the part worth knowing.
4. Spot-check before you act
Pick two or three numbers from the output each week and trace them back to the source export. Row-level arithmetic across large files is where these tools most often slip, especially when a report has parent and child rows mixed together or when a date column is read as text. If a number does not reconcile, say so in the session and re-run that section rather than accepting a corrected guess.
Two structural caveats worth knowing: ad-attributed sales continue to be assigned back to earlier click dates, so the most recent day or two of ad data will keep rising after you pull it — don’t build a bid decision on the newest rows. And any file export reflects the moment you pulled it, so pull everything in one sitting.
5. End with a dated decision log, not a narrative
Close every session by asking for no more than five actions, each with the SKU or campaign, the specific change, who does it, and the metric you will check next week. Keep this list in one running document with the date at the top. Next week, paste last week’s list in as the first input so the review starts by grading the decisions you already made.
This is what turns the review from reading into managing. It also gives you an honest record of how often the AI’s suggested explanation turned out to be the right one.
💡 Pro Tip: Keep an explicit “unchanged” request in your prompt — ask the assistant to name the SKUs that stayed inside every variance band. A stable list you can scan in ten seconds is what makes an exception report trustworthy; without it, you can’t tell whether a quiet week means nothing moved or the file failed to parse.
⚠️ Where Weekly AI Reviews Break Down
Asking for profit from data that contains no costs
Amazon’s traffic and advertising reports contain revenue, not margin. If you ask which SKUs are most profitable without supplying landed costs and current fees, you will get an answer built on revenue ranking with confident margin language attached. Supply the cost table or restrict the question to revenue and units.
Changing the report pack whenever something looks interesting
It is tempting to add a new export each week because last week’s question needed it. Within a month you have no consistent series and no way to tell a real trend from a reporting change. Add reports at a planned interval — quarterly, for instance — and keep the core pack fixed in between.
Reviewing account health metrics as ratios instead of against thresholds
A week-over-week percentage change on a small metric is misleading. What matters is distance from Amazon’s threshold: Order Defect Rate below 1%, Late Shipment Rate below 4%, and Pre-fulfillment Cancellation Rate below 2.5% for seller-fulfilled orders, plus an Inventory Performance Index of 400 or above for FBA storage capacity. Put those thresholds in your prompt and ask for distance-to-threshold, not percentage change. Confirm the current values in Account Health, since Amazon revises both thresholds and enforcement mechanics.
Letting the tool choice dictate the workflow
File size limits, spreadsheet handling, connector availability, and how much data an assistant can hold in one session change often, and different tools handle multi-file analysis very differently. Check the current limits in your vendor’s own documentation before you design a pack around them. The pack, the definitions block, and the exception prompt are portable; a workflow built around one product’s upload behavior is not.
❓ FAQs
How long should a weekly review take once it’s set up?
Pulling a fixed pack of exports is usually the longest part, and it is the part you can shorten with saved report configurations. Budget more time in the first few weeks, when you are still tuning variance thresholds and catching arithmetic errors, and expect the session to get shorter as your definitions block stabilizes.
Can AI pull the reports for me instead of manual exports?
Programmatic access to Amazon data goes through the Selling Partner API, and some tools expose it to an assistant through a connector. That removes the export step but not the review discipline — you still need the same windows, the same definitions, and the same spot-checks. Whatever you do, pull data through Amazon’s sanctioned API rather than any tool that scrapes Seller Central, which conflicts with Amazon’s terms.
Weekly or monthly — which cadence is right?
Weekly suits anything you can still act on: bids, stock cover, suppressed listings, a conversion drop worth investigating. Monthly suits questions where a single week is mostly noise, such as SKU-level margin trends, return-rate shifts, and catalog pruning decisions. Running both, with different questions in each, avoids treating normal weekly variation as a trend.
Should I compare against last week or the same week last year?
Use week-over-week to catch operational problems and same-period-last-year to judge whether a seasonal move is normal. Give the assistant both comparison columns and tell it which one applies to which question, otherwise it will pick one and present it as the whole picture.