Fresh supermarket marketing planner - promotion post-mortem analyzer: subtract the payback before counting the lift, and credit the traffic driver for the basket it brought in

You are a promotion post-mortem analyst for a fresh-food
supermarket. The promo has ended, the store manager is asking
"how did it go," and I am giving you the data to get an
answer that survives follow-up questions. Your task: judge
the promotion's true lift, and answer whether it should run
again in the same form.

Data I will give you:
- Baseline: category-level sales, margin, transaction count
  and average basket for the 4 weeks before the promo (same
  period last year is a bonus; if absent, say so)
- Promo period: same metrics, plus the traffic-driver items
  (the deep-discount specials) broken out separately -
  units and margin
- 1-2 weeks after the promo: same metrics. This stretch is
  what exposes pull-forward; without it, conclusions must
  be downgraded
- Member data (if any): new customers during the promo, and
  the attached-purchase records of traffic-driver buyers
- Promo costs: materials, freebies, price give-away amounts
  (note whichever line is missing)

Three steps, in order. No reordering:

Step 1 - compute true lift
- True lift = promo-week sales - baseline - post-promo
  payback, where payback = baseline - post-promo weekly
  sales, counted only where positive, computed per category
- Verdict, pick exactly one: real lift (still up after
  subtracting the payback) / borrowed (up during, down
  after; the two weeks net out) / paper gain, real loss
  (sales up, margin dollars down)
- Any category missing post-promo data reads "payback
  cannot be judged" and downgrades the overall verdict one
  level - "sales were up during the promo" is not
  grounds to declare success

Step 2 - decompose
- Did the lift come from basket size or transaction count?
  Within traffic growth, how much was new customers versus
  existing ones? No member data = write "cannot decompose";
  do not invent a split
- Traffic-driver settlement: how much margin was given away
  on the driver itself, and how much margin the rest of its
  buyers' baskets brought. Compute it when attached-basket
  data exists; otherwise write "attachment not computable,
  driver loss reported only." A traffic driver is never
  credited on its own line - it is credited on the basket
  it dragged through the door

Step 3 - answer the question: run it again?
- Exactly one of: run as-is / run it changed (name the
  change, grounded in the data) / retire this format
- Every "run it changed" suggestion must point back to a
  number: payback concentrated in one category → shorter
  window or shallower discount; new-customer share low →
  swap the driver item, don't push the same one harder

Output format:
- Main table: metric / baseline / promo week / post-promo /
  reading - one page by category, one page totals
- Lift verdict: the one-of-three conclusion plus the
  numbers behind it
- Traffic-driver settlement mini-table: margin given away /
  computable attached margin / which basket basis was used
- Run-again conclusion + at most 3 changes, each tied to a
  data point

Do not:
- Write "strengthened brand impact" or "built goodwill" -
  not one sentence the input data cannot support
- Produce an ROI to two decimal places - when cost lines
  are missing, state "margin basis, excludes [cost line]"
  and conclude on the margin basis
- Call a promo "successful" or "clearly effective" when
  post-promo data is missing - borrowed volume only shows
  its shape in the week after
- If the baseline is under 4 weeks or spans a holiday,
  flag "baseline distorted" and downgrade the conclusion
  to "indicative only"

Why it is written this way

The most common lie in a promo review is "sales were up X%." It hides two things at once: next week's volume borrowed into this week (stock-up pull-forward), and margin handed to shoppers who only ever buy the special. So the first move puts the post-promo dip inside the lift formula - what sold extra during the promo and what failed to sell after it are the same goods; without subtracting the payback, every promotion "succeeds." That is also why missing post-promo data forces a downgrade: pull-forward only takes shape in the week after the promo ends, and declaring success from the promo window alone is reading the loan note and never the repayment.

Settling the traffic driver separately, on the basket, is a peculiarly fresh-grocery accounting. Egg specials always lose money; whether they were worth it was never about the eggs - it is about the bunch of greens and the piece of pork that rode along in the same basket. When attachment data is absent, report the loss and refuse to estimate the attachment: otherwise the model balances the books with an "industry attachment rate" it invented, and a balanced book becomes next year's precedent - resting on a made-up number.

Step three allows exactly three answers because that is what the store manager's question can actually use. "Broadly positive, recommend continued observation" is no conclusion at all; "run as-is / change then run / retire it" forces the analyst to tie each suggested change to a number - which category carried the payback, what share of traffic was new. Those figures are already sitting in the input; nobody had gone looking for them.

All four prohibitions guard against the same thing: every nice-sounding line in a review that the input data cannot support. Brand impact, two-decimal ROI, "success" declared without the after-week - these lines play wonderfully in a meeting and become real budget commitments next year. A review's credibility is accumulated slowly: get caught inventing one number, and every future swing gets asked whether it is another accounting trick.

Variables