Generate a weekly Amazon BUSINESS review - profit, margin, ad efficiency (TACoS) and inventory - comparing this week to your trailing 4-week normal, explaining any margin/fee anomaly, and ending in concrete actions. Use for "weekly business review", "weekly profit review", "weekly P&L", "how is the business doing", "what changed and why", "is my margin or TACoS ok", "profit this week". For a quick sales-only snapshot, use the Weekly Sales Briefing instead.
Profit & FinanceRead-onlyReportMCP
Read-only — this skill only reads your data and never changes your Amazon account.
The full skill specification, rendered straight from the source repository.
Weekly Business Review
The Monday-morning business review: one card across profit, margin, ad efficiency
(TACoS) and inventory that compares this week to your normal (a trailing
baseline, not just last week), explains any margin/fee anomaly, and ends in the few
actions worth doing. This is the deep profit/diagnosis review - for a quick
sales-only snapshot use the Weekly Sales Briefing instead. Live from DataDoe.
When to use this
Every Monday, or the start of any week - the deeper profit/business review.
The single "how's the business doing, what changed, why, what do I do" read.
Trigger phrases: "weekly business review", "weekly profit review", "weekly P&L",
"how is the business doing", "what changed and why", "is my margin ok",
"is my TACoS ok", "profit this week". (For plain "sales this week" / "weekly sales
report", the Weekly Sales Briefing is the right skill.)
The framework. Period integrity -> baseline -> anomaly -> movers -> 3 actions
Get the period right first (or every number below is a lag artifact). Lag varies
by account and by source: the profit tables are INTRADAY (refreshed ~10am / 1pm / 4pm)
with ~1 day of lag observed on sales, while their traffic columns trail by a lag you must measure (Amazon's own note
says up to 3 days; measured from 2-3 days (Nov 2026) to 8-13 days (Oct 2026) on a UK account, with holes) and older
sources lagged 6+ days with mid-series gaps. Never assume "this week = the last 7 calendar
days" or a fixed lag - detect the last complete day dynamically (see workflow),
anchor both week windows to it, and count each week's effective (complete) days.
If a week is missing days, normalize by effective days or mark it provisional -
never present a partial week's drop as a real move. (Same completeness guard as the
Sales-Movers and Buy-Box skills; the lag it detects differs per source.)
Compare to normal, not just last week. Show this week vs the trailing
4-week median (median, not mean - one anomalous week shouldn't move the
baseline) as the primary signal, plus last week for context, plus a small 6-8
week trend. A single week is noisy; the baseline keeps the headline honest.
Explain the anomaly - and check it's real. If margin or TACoS deviates
materially from baseline (e.g. > 5pp) or moved more than 2pp week on week, decompose
why using the cost columns as
% of sales - sales_tax % (VAT, ~16.7% of sales on UK/EU accounts; 0 on US/CA),
total fees % (of which FBA % and referral %), COGS %, ad %, refunds %. Refunds % is
-refund_cost / total_sales: refund_cost is signed and added to profit (refund
outflows negative, fee reversals positive), so its cost-side share is its negative -
a -10 refund contribution on 100 of sales is +10% of costs. The lines sum to 100% -
margin, so the decomposition reconciles only with sales_tax and the refunds line in
it; omit VAT and a UK review infers a +16.7% margin where the real one is negative. Name the driver. Then apply the
settlement-timing test:total_fees and fba_fees book order fees on the order
date, but storage, long-term storage, disposal and inbound fees on the posting
date, so profit_by_date weekly margin is lumpy - a fee spike on one or two days of
otherwise-flat sales is usually a posted-date bill, not a real cost increase. On a
UK account FBA storage posted around the 7th of each month and long-term storage around
the 19th-20th; weeks holding a bill ran -3.8% to -5.7% margin, weeks without ran -1.4%
to -2.8%. These are legitimate monthly bills, not reimbursement candidates: name each
bill and the day it posted (step 7), show the week's margin with and without them, and
keep "flag for reimbursement / fee review" (reconcile the fee rows by order_date in
amazon_settlements_with_cogs) for order fees that look wrong. Sales, units and ad spend are sale-dated and
reliable weekly; margin is only trustworthy over a trailing 4-week window.
Movers. Biggest SKU profit gainers/droppers this week vs last.
Do this week. 3 concrete actions, each pointing to the deeper skill.
Configuration
MCP base: https://mcp.datadoe.com/mcp/v1
Data sources:
Profit by Date (amazon_profit_by_date) - account-level daily rows that the
review buckets into weeks itself, plus the cost breakdown (total_sales, profit,
ad_spend, cogs_total, total_fees, fba_fees, total_selling_fees,
sales_tax, refund_cost, total_units_sold). profit is already
total_sales - sales_tax - total_fees - cogs_total - ad_spend + refund_cost.
Profit by SKU & Date (amazon_profit_by_sku_and_date) - per-SKU movers.
Optional: FBA Inventory Health (amazon_fba_inventory_health) for stockout watch-outs.
Optional, to name posted-date bills (step 7): Settlements & P&L Components
(amazon_settlements_with_cogs) - not premium, in the default dataset,
requiresDatePeriod: true. Settlement sign: negative = cost.
Optional, for traffic / conversion / buy-box watch-outs: the traffic columns already
on the two profit tables - total_sessions, total_page_views,
avg_buybox_percentage per date on amazon_profit_by_date (account level) and
per SKU / child ASIN on amazon_profit_by_sku_and_date. No extra source is needed.
Two rules: (1) on the per-SKU table these three columns are per child ASIN and
repeated on every SKU row of that ASIN - take max() per child_asin, never
sum() (summing double-counts every multi-SKU ASIN); (2) the traffic columns may be
delayed by a lag you must measure (Amazon says up to 3 days; measured 2-3 to 8-13
days on a UK account, per child ASIN a day more than the account total) and revised
for 30 days, while sales lag
~1 day - so a conversion rate (units / sessions) for the days after the last
traffic-complete day is provisional, and a
week's sessions can still move after the review is written.
Daily data lag - detect it, don't assume it. The profit tables are INTRADAY
(exports_source_get reports refreshes at ~10am / 1pm / 4pm) and lag about 1 day on
sales: the newest date is a partial day and the day before it is usually sales-complete
but may still be missing fees. Older sources lagged 6+ days with every-other-day gaps;
these do not. Either way, do NOT just "drop the partial current week": detect the last
complete day dynamically and measure each week's effective days (see workflow). On
the profit sources, measure completeness by daily sales value, not row count - rows
backfill (all SKUs appear at ~full count) before the sales/profit values settle, so a
row-count test falsely marks a settlement-incomplete day as complete - and check fee
completeness separately (fees post after sales; see step 3). If the review reads the
traffic columns, treat them as a third completeness tier: a day can be sales- and
fee-complete while total_sessions is still zero or partial for a week or more - detect
the last traffic-complete day the same way and label any conversion / buy-box figure
that includes later days provisional. For account-level figures use the daily account
total_sessions; when watch-outs read per-ASIN conversion or buy box, use the
per-ASIN count instead, because per-ASIN traffic can trail the account total by a day
(a day with full account sessions and no child-ASIN traffic reads as a buy-box collapse
on every ASIN): the Sales Movers probe - one export on amazon_profit_by_sku_and_date,
filterstotal_sessions > 0 and marketplace_country_code = the marketplace,
groupBy [date], countDistinct on child_asin; a missing date counts as 0, and a day
passes at ~60% of the median daily count. Weight per-ASIN buy box by total_page_views
and leave out ASIN-days with 0 page views.
Weekly buckets are built in code, not with dateInterval WEEK.dateInterval WEEK
returns fixed Sunday-start calendar buckets regardless of from, and its newest bucket
includes the partial current day; no request parameter shifts the boundary to the last
complete day. Pull daily rows and sum 7-day windows ending on the last complete day.
Currency: read currency, localise; if the connection spans currencies, group by
currency and report the main one (never sum across currencies). Rows with an empty
currency can appear on amazon_profit_by_date (observed on 50 of 72 days); they
carry only refund_cost (return COGS added on the return date). On a single-currency
account fold them into that currency and say so in one line (dropping them misstated a
week's profit by 94.36); on a multi-currency account report them separately.
Step-by-step workflow (MCP-native)
sellers_and_vendors_list -> pick the seller.
exports_sources_get -> confirm sources enabled. amazon_profit_by_date,
amazon_profit_by_sku_and_date and amazon_fba_inventory_health are premium. Premium note: a premium export costs 5 AI Tokens instead of 2 - nothing else differs, and the table is part of the always-on default dataset, so it is never disabled. A 0-row export means no data in the window or an initial load still in progress - say which, and stop (or drop that section); never render zeros.
One daily export, then calibrate completeness (don't assume a lag).exports_create
once on amazon_profit_by_date, the last ~63 days plus a few days of slack (8 weeks of
trend + the 4 baseline weeks), groupBy [date, currency], sum total_sales, profit,
ad_spend, cogs_total, total_fees, fba_fees, total_selling_fees, sales_tax,
refund_cost, total_units_sold and unshipped_sales with distinct aliases
(sum(profit) as profit_sum, sum(refund_cost) as refund_cost_sum,
sum(unshipped_sales) as unshipped_sum; as profit errors ALIAS_COLLISION). This single export feeds steps 3-7; do not add dateInterval WEEK
(see Configuration). On the profit sources the completeness signal is daily sales
value, NOT row count - rows backfill (all SKUs appear at ~full count) before the
sales/profit values settle, so a day can pass a row-count test yet hold only a fraction
of its real sales (observed: a day with ~1,450 rows but £143 sales vs a normal ~£8,000).
From the daily total_sales: the median daily sales, a full-day threshold ~60% of
median daily sales (relative, per-account), the sales-complete days (at/above
threshold) and gap days (any date below threshold, including mid-series holes).
(Row count is a fine secondary sanity check, but sales value is authoritative here.)
Unshipped check (the same rule as Sales Movers).unshipped_sales is not included
in total_sales, and a recent day can pass the 60% test while many of its orders are
still waiting to ship. A day whose unshipped_sum is more than ~10% of
total_sales + unshipped_sales is still filling and is NOT complete, even if it passes
the 60% test. Observed: a day at 74.8% of median sales had 27.0% of its order value
unshipped (older days 0-2%); anchoring on it read margin -1.83pp vs normal, against
-0.13pp one day earlier.
Fee-completeness check (beside the sales test). Fees post after sales, so a day
can be sales-complete with fees still missing. Test the order-dated fees: for each
candidate day compute total_selling_fees / total_sales (the referral fee, booked on
the order date) and compare it with the median of that ratio over the 5 preceding
complete days: if the day's ratio is below ~80% of the trailing ratio, its order fees
have not posted yet. Do not run this test on total_fees or fba_fees - they also
book storage, long-term storage, disposal and inbound bills on the posting date, so a
monthly bill can hide missing order fees (observed: a 2,158 storage bill lifted a day's
total-fee ratio to 124% of trailing while its referral fees were only 72% posted).
Separately, flag any day whose total_fees / total_sales is above ~120% of its trailing
ratio as "a posted-date bill landed" - step 7 names it. (Earlier observation: a day
at 71% of median sales with fees at 39% of sales vs ~50% normal turned the week's
margin from -3.98% to -3.06% and flipped the WoW margin delta's sign.)
Last complete day (the anchor) = the latest date that passes all three tests -
sales value, unshipped and referral fees - walking back from the tail (the lag is
whatever this detects; ~1-2 days on the profit tables, more on other sources). Name any
later day you left out and which test it failed. Anchor "this week" = the 7 days
ending on the last complete day, "last week" = the 7 before that.
Weekly account trend + cost mix, bucketed in code. From the step-3 daily rows, sum
8 consecutive 7-day windows ending on the last complete day (this week = days -6..0,
last week = -13..-7, and so on) per currency; recompute margin and TACoS per bucket from
the sums. Do NOT use dateInterval WEEK: it produces fixed Sunday-start calendar
buckets whose newest bucket includes the partial current day and whose oldest is
clipped by from, and it cannot be anchored to the last complete day. Carry the fee
decomposition columns (total_fees, fba_fees, total_selling_fees, cogs_total,
sales_tax, ad_spend, refund_cost) through the same buckets so step 7 can read
them.
Effective-days guard (both weeks). For "this week" and "last week", compute
effective days = count of days present at/above the threshold. If a week has < 7
effective days, either normalize the comparison per effective day OR mark that week
provisional and state how many days it reflects - never present a partial week's
delta as a real move. If the two weeks' effective days differ materially, the raw Δ is
an artifact; flag it. (Check BOTH weeks - an under-counted prior week distorts the Δ
just as much as a partial current one.)
Baseline: latest complete week = "this week"; prior = "last week"; baseline =
median of the 4 complete weeks before this one (skip or flag any baseline week that is
itself under-counted). For each KPI show this week, the Δ vs last week, and the Δ vs
baseline. Recompute margin = profit/sales and TACoS = ad_spend/sales per week (never
sum ratios); if a week is provisional, label its margin/TACoS provisional too - don't
average a rate over missing days. The baseline of each KPI is the median of its 4
weekly values - for margin and TACoS the median of the 4 weekly ratios, not the ratio
of the 4 weeks' pooled sums (observed -3.86% vs -3.84%) - and the header says "median".
Profit deltas: when this week's, last week's or the baseline profit is 0 or below,
show the change in currency, not %: a % change on a negative base reads backwards
(observed: a deeper loss shown as -92.86% vs last week and -35.49% vs normal).
Anomaly drill: if this week's (or a recent week's) margin/TACoS is > ~5pp off
baseline, or this week's moved more than 2pp vs last week, express each cost as % of
sales per week - sales_tax% (VAT; 0 on US/CA),
total_fees% (of which fba_fees% and total_selling_fees% - subsets, never added on
top), COGS%, ad%, refunds% = -refund_cost_sum / sales_sum (sign reversed - profit
adds the signed refund_cost) - and check they reconcile: the lines sum to
100% - margin. Identify which line moved - that is the cause. Report it in plain words.
Posted-date bills (runs whether or not the drill above fired). When a week's
FBA-fee share moves, or step 3 flagged a day as
"a posted-date bill landed", run one export on amazon_settlements_with_cogs over the
review span (the 8 weeks; pass from / to - requiresDatePeriod: true):
groupBy [date, currency], sum fba_storage_fee, long_term_storage_fee,
fba_disposal_fee, fba_inbound_placement_service_fee and
fba_inbound_transportation_fee with distinct aliases. Settlement sign: negative =
cost, positive = a credit. On the first run, check date against
transaction_posted_date for the bill rows (they matched on every bill on a UK
account, Nov 2026); if they differ, use the one whose days match the fee spikes in
step 3. Name each bill and the day it posted - FBA storage posts around the 7th of the
month, long-term storage around the 19th-20th; inbound placement and transportation
post when a shipment is billed - and show the week's margin with and without these
bills ((profit_sum + bills) / sales_sum, bills as positive costs). They are normal
bills, not reimbursement candidates. When a day flagged in step 3 matches none of the
exported bills, say so: "an unnamed posted-date fee of about {cur}X on {date}" (X = that
day's total_fees minus its trailing fee ratio x its sales), not a storage bill, and
leave it inside "margin without bills".
Movers (full per-SKU totals, diffed client-side).exports_create twice on
amazon_profit_by_sku_and_date - once for this week, once for last week (the windows
anchored in step 3) - groupBy [sku, product_name], sum profit, total_sales,
total_units_sold with distinct aliases (profit_sum, sales_sum, units_sum -
as profit errors ALIAS_COLLISION), format CSV with limit 5000 (JSON caps at 1,000
rows, CSV at 5,000; page with skip if the account has more SKUs than that). Drop rows whose sku is null or blank
(account-level rows - observed ~1,300 per week), then sum per sku before diffing: about 165 SKUs per week came
back twice, once with a blank product_name; keep the non-blank name. Diff by SKU in code -> top gainers/droppers by
Δ profit, and say how many SKUs sold in only one week (total_sales > 0 in one week and not the other - mere
presence in one export undercounts it: 505 vs 802 observed). Do NOT pull a top-N by profit per week (e.g. limit ~200 DESC): on a
~3,000-SKU account only 97 of 200 SKUs overlapped between the two weeks, and any SKU
that went from positive to negative - the dropper a reviewer most wants - falls out of
this week's list and cannot be diffed at all. There is no server-side order by change,
so the full pull is the only reliable shape.
Render the HTML card.
Output format (interactive HTML card)
Render a single self-contained HTML card (KPI tiles, a small trend sparkline per
KPI, a movers table, watch-outs, actions). Header "Weekly Insights", footer "Data
via DataDoe". If the client cannot render HTML, fall back to the text layout below.
text
1Weekly Insights - {marketplace} - week of {start} (ends {last complete day}) (vs 4-wk median)
2Period: this week = {n}/7 complete days{ - PROVISIONAL if < 7}; gaps: {dates or none}{; left out: {date} - {n}% unshipped / referral fees {x}% posted}.
3 Sales are shipped sales; shipped + unshipped {cur}.. ({+/-}% vs last wk; pending orders not included, so below Seller Central's ordered sales).
45 This wk vs last wk vs normal(4wk median)
6Sales {cur}.. {+/-}% {+/-}%
7Profit {cur}.. {+/-}% {+/-}% (as {+/-}{cur}.. when either side is 0 or below)
8Margin {m}% {+/-}pp {+/-}pp
9Ad spend {cur}.. {+/-}% {+/-}%
10TACoS {t}% {+/-}pp {+/-}pp
11Units {u} {+/-}% {+/-}%
12Trend (8wk): sales ▁▃▆▅▆▇ profit ▆▇▃▁▄▆
1314Why (if > 5pp off normal or > 2pp vs last week): margin {m}% vs normal {b}% / last wk {l}% - driver: {fees/COGS/ads} moved
15 from {x}% to {y}% of sales in wk {date}.
16 Cost mix (% of sales): VAT {v}% fees {f}% (FBA {fb}%, referral {r}%) COGS {c}% ads {a}% refunds {rc}% -> margin {m}%
17 {if a day in the week failed the referral-fee test: "Margin provisional: {date} is fee-incomplete (referral fees {x}% of sales vs ~{y}% trailing)."}
18Bills {FBA storage {cur}.. on {date}; long-term storage {cur}.. on {date}; inbound {cur}.. on {date}} in the 8 weeks - margin without them {m2}% (normal {b2}%).
19 {if a flagged day matched no bill: "Unnamed posted-date fee of about {cur}X on {date}."}
20 (Shown whenever step 7 ran the settlements export, even with no anomaly.)
2122Top movers +{sku} {cur}.. -{sku} {cur}..
23Watch-outs {rising TACoS / stockouts / buy-box}
24Do this week 1) ... 2) ... 3) ...
Worked example (illustrative)
Suppose this week's margin reads -5.7% against a normal (4-week median) of -3.9% and
-2.9% last week. That is under the 5pp anomaly bar but a 2.8pp drop week on week, so the
drill runs. Decomposing costs as % of sales, VAT, COGS and ads held roughly flat - the
mover was FBA fees, up about 3pp of sales, and step 3 had flagged one day of the week
as "a posted-date bill landed" (its total-fee ratio at ~124% of trailing while its
referral fees looked normal). The settlements export names it: the monthly FBA storage
bill (2,158) posted on the 7th, while last week instead got a storage credit (+638 on the
29th). The 8-week view shows the same saw-tooth - storage around the 7th, long-term
storage around the 19th-20th - and every week holding a bill ran 1-3pp below the weeks
without one. So the card concludes: "the margin dip is this month's storage bill landing
in this week, not an operational problem; without posted-date bills margin is about the
same as normal," and reads margin on the trailing 4 weeks. No reimbursement flag - these
are legitimate monthly bills; that flag is kept for order fees that look wrong. That is
the difference between a scary wrong number and a correct insight.
Quality self-check
Did I detect the last complete day dynamically (not a fixed lag / last-7-calendar-days)
and anchor both weeks to it - bucketing the daily rows in code, not with
dateInterval WEEK?
Did I run the unshipped check (a day with > 10% of its order value unshipped is not
complete) and the fee-completeness check on the order-dated referral fee
(total_selling_fees / total_sales vs the trailing 5-day ratio, not total_fees), and
anchor on the latest day that passes all three tests?
Did I name every storage / long-term storage bill in the review span with the day it
posted, and show the week's margin with and without them?
Does the header say "vs 4-wk median", and are profit changes in currency when either
side is 0 or below?
Did I fold empty-currency rows into a single-currency account (and say so), sum the
per-SKU rows per sku before diffing, and count "sold in only one week" by
total_sales > 0?
Did I compute effective days for BOTH weeks, flag mid-series gaps, and mark a partial
week provisional rather than reporting a phantom drop?
Did I diff movers over full per-SKU pulls (CSV limit 5000, both weeks), not a top-200
per week?
Does the cost mix include sales_tax % and the refunds line as -refund_cost / sales
(exported in step 3, carried through the weekly buckets), and reconcile to 100% -
margin?
Did I avoid averaging rate metrics (margin/TACoS) over a partial week?
Did I compare to the trailing baseline, not just last week (so a noisy prior week
can't mislead)?
If margin/TACoS was off, did I name the cost driver, not just flag it?
Ratios recomputed per week (not summed)?
Did I end in 3 concrete actions?
Common mistakes
Assuming "this week = the last 7 calendar days" or a fixed lag - lag differs per source
and per column (~1 day on the intraday profit tables' sales, a measured 2-3 to 8-13
days on their traffic columns, 6+ days with gaps on older sources); detect the last
complete day and anchor to it.
Summing total_sessions / total_page_views / avg_buybox_percentage across the SKU
rows of amazon_profit_by_sku_and_date - they are per child ASIN and repeated on each
SKU row; take max() per child_asin.
Using dateInterval WEEK for the weekly trend - it yields fixed Sunday-start calendar
buckets, the newest including the partial current day; bucket the daily rows in code.
Treating a sales-complete day as margin-complete - fees post later; run the
fee-completeness check on the referral fee or the newest day inflates margin. A
posted-date storage bill on the same day can make total_fees look complete when the
order fees are not.
Anchoring on a day that passes the 60% sales test while a large share of its orders is
still unshipped - check unshipped_sales per day.
Pulling movers as a top-200 by profit per week - SKUs that went negative vanish from
the list; pull full per-SKU totals (CSV limit 5000) for both weeks and diff in code.
Decomposing costs without sales_tax - on UK/EU accounts VAT is ~16-17% of sales and
the lines will not reconcile to the margin without it.
Using the raw signed refund_cost as the refunds share, or leaving it out of the step-3
export - profit adds refund_cost, so the cost line is -refund_cost / sales; with the
raw value or no value the mix cannot reconcile to the margin.
Reporting a partial/gappy week's drop as a real collapse - normalize by effective days
or mark it provisional (check BOTH weeks; an under-counted prior week distorts the Δ too).
WoW-only headline: last week being abnormal makes this week look great/terrible.
Reading a single week's margin as real - storage, long-term storage and other bills
post on one day a month; use the trailing window for margin, name the bills from
amazon_settlements_with_cogs, and keep the reimbursement flag for order fees that
look wrong.
Showing a % profit change when a week lost money - a deeper loss reads as a small or
positive %; show the change in currency.
Flagging "margin dropped" without decomposing which cost moved.
Summing TACoS/margin/ACoS columns.
A wall of metrics with no action.
Notes
Read-only. Routes to the deeper DataDoe skills for each action.
A DataDoe skill, built on DataDoe Profit by Date (amazon_profit_by_date) + Profit by
SKU (amazon_profit_by_sku_and_date), with the same completeness/lag guard as
Sales-Movers and Buy-Box.