Post-Promotion Review From Your Results Table: A Four-Prompt ChatGPT Chain
For Advertising & Promotions Managers ·
What This Builds
The review of a finished promotion usually gets written late, from memory, if at all. This chain takes a small summary table you already have and walks it through four prompts in one ChatGPT conversation: the differences and cost per redemption, what the numbers do and do not show, a repeat, change or drop list, and a half-page learnings entry you can file.
You supply every figure, and you recompute each one before going on. ChatGPT organizes and words. It never supplies a benchmark.
Prerequisites
- A ChatGPT account at chatgpt.com. The Plus plan is listed at $20/month, and that is the only ongoing cost of this chain.
- A summary table for one finished promotion: promotion code, planned and actual spend by line, and planned and actual redemptions or entries, in the units your finance team uses
- The entry format of your learnings library, if you keep one (see below)
- Your company's AI policy, checked for whether results and spend by promotion may go into this tool
The Concept
Think of a good analyst sitting beside you with a calculator and a notepad. You hand over a one-page table. The analyst shows the arithmetic on each line, tells you what the table cannot say, asks who would know, and then writes down what to do next time, always pointing at the line that supports it. You check the calculator, and you answer the questions.
The order matters. Arithmetic comes first so everything later rests on figures you have verified. Questions come second, because numbers alone rarely explain why a promotion over or under performed. Recommendations come third, tied to rows and to your notes. The entry comes last.
Before You Paste Anything
The table holds spend and results, which are commercially sensitive. Before step 1:
- Replace unreleased product names with a code name, and replace retailer and agency names with codes such as Retailer A and Agency B where terms are confidential.
- Leave out all consumer data. The table is aggregate counts only: no names, emails, addresses, phone numbers or individual entries.
- Check your AI policy and the data settings of your ChatGPT plan, including whether chats are used to train models and how long they are kept. A free consumer plan may train on chats unless the setting is off.
Build It Step by Step
Part 1: Prompt 1, differences and cost per redemption
Open a new conversation and paste your table into the placeholder:
Here is the summary table for a finished consumer promotion. For each spend line and for the total, show planned, actual, the difference (actual minus planned) and the difference as a percentage of planned. For redemptions, do the same for each retailer and the total. Then calculate cost per redemption, planned and actual, as total spend divided by total redemptions, and the difference between them. Show every calculation with the numbers written out so I can check it. Use only the numbers in the table.
[paste your table here]
Checkpoint. Recompute every figure yourself with a calculator or a spreadsheet. Fix any error by pointing to the line, and ask ChatGPT to redo the calculation. Do not go on until all of them agree with yours.
Part 2: Prompt 2, what the numbers show and what they do not
Using only the verified figures above, write two short lists. First, what the numbers show: statements each tied to a specific row. Second, what the numbers do not show: things I would need to know to explain the differences. Then write questions I could ask the sales team, the agency and the fulfillment vendor to fill those gaps. Do not guess at causes, and do not compare against any benchmark or industry figure.
Checkpoint. Ask the people the questions. Come back with short notes in your own words, for example "Sales: Retailer B's feature ran a week later than planned." Paste them into the conversation with a line saying they are your notes. Do not paste in anything you have not heard from a person.
Part 3: Prompt 3, repeat, change or drop
Based on the verified figures and my notes above, give me three lists: repeat, change and drop. Each item must name the table row or the note it rests on. If an item has no row or note behind it, leave it out. Do not use benchmarks, do not invent reasons, and do not recommend anything that the table and notes do not support.
Checkpoint. For each item, find the row or note it cites and confirm the item follows from it. Remove anything that reads as an opinion with no support. Add anything you know that the numbers cannot show.
Part 4: Prompt 4, the learnings entry
If you keep a learnings library in a Claude Project (the Level 3 guide on setting up a Claude Project as a promotions playbook and learnings library), paste in the entry headings from it. If not, use these: Promotion, What we planned, What happened, Why (from notes only), Repeat, Change, Drop, Open questions.
Write a learnings entry of about half a page using these headings: [list your headings here]. Use only the verified figures, my notes and the repeat, change and drop lists above. Write in plain sentences for a colleague who was not on the promotion. Do not add figures, causes or comparisons that are not in the conversation.
Checkpoint. Read the entry against your table one last time. Every number in it must appear in the verified figures. Then file it. If legal or finance should see the review, send it to them before it circulates.
Real Example: FALL-C
The promotion, the codes and every figure are invented. Amounts have no currency symbol.
| Spend line | Planned | Actual |
|---|---|---|
| Media | 40,000 | 44,000 |
| Print and point of purchase | 12,000 | 9,000 |
| Fulfillment vendor | 18,000 | 21,000 |
| Agency fees | 10,000 | 10,000 |
| Total | 80,000 | 84,000 |
| Redemptions | Planned | Actual |
|---|---|---|
| Retailer A | 5,000 | 5,100 |
| Retailer B | 3,000 | 2,100 |
| Total | 8,000 | 7,200 |
Prompt 1, the figures to recompute:
- Media: 44,000 minus 40,000 is 4,000 over, and 4,000 divided by 40,000 is 10 percent.
- Print and point of purchase: 9,000 minus 12,000 is 3,000 under, and 3,000 divided by 12,000 is 25 percent under.
- Fulfillment: 21,000 minus 18,000 is 3,000 over, and 3,000 divided by 18,000 is about 16.7 percent.
- Agency fees: no difference.
- Total spend: 40,000 plus 12,000 plus 18,000 plus 10,000 is 80,000 planned. 44,000 plus 9,000 plus 21,000 plus 10,000 is 84,000 actual. The difference is 4,000, which is 5 percent of 80,000.
- Redemptions: Retailer A is 100 over (2 percent of 5,000). Retailer B is 900 under (30 percent of 3,000). The total is 7,200 against 8,000, which is 800 under, 10 percent.
- Cost per redemption: planned 80,000 divided by 8,000 is 10.00. Actual 84,000 divided by 7,200 is about 11.67. The difference is about 1.67, roughly 16.7 percent more than planned.
Spend rose 5 percent while redemptions fell 10 percent, which is why cost per redemption rose by more than either.
Prompt 2 and your notes: ChatGPT lists what the table shows (the Retailer B gap, the media overrun, the print saving) and what it does not (why Retailer B fell short, why fulfillment cost more). It proposes questions. You come back with notes: "Sales: Retailer B's feature ran one week later than planned." "Agency B: the media overrun was an added placement approved mid-flight." "Vendor D: stock shipped in two lots, and the second lot cost extra to expedite."
Prompt 3: Repeat: the print approach, which came in under plan (the print row). Change: schedule the feature date with Retailer B in writing earlier (the Retailer B row and the sales note). Drop: nothing, since no row or note supports dropping anything. A "drop the Retailer B placement" suggestion would have no support, and you would delete it.
Prompt 4: A half-page entry in your headings, with the cost per redemption figures exactly as you verified them.
Time saved: The first draft of the review exists the same day the results land. Your time goes to checking figures and asking the three questions, not to assembling a document from memory.
What to Do When It Breaks
- A calculation is wrong → Point at the line, give the right figure and ask for the step to be redone. Large tables are where errors appear, so recompute every number anyway.
- A benchmark or industry comparison appears → Delete it. Repeat the instruction that only the table and your notes may be used, and rerun that prompt.
- ChatGPT states a cause as fact → Replace it with "not known" unless someone gave you the cause in your notes.
- The conversation gets long and ChatGPT starts mixing figures → Open a new conversation and paste the verified table and your notes there. Then run prompts 3 and 4.
- The entry will not match your library format → Paste an existing entry as a model and ask for the same headings, using only this conversation's content.
Variations
- Simpler version: Run prompts 1 and 2 only and use the questions in your review meeting.
- Extended version: Run the chain for several promotions and ask a fresh conversation to compare their filed entries, checking each comparison against the entries.
What to Do Next
- This week: Run the chain on the last promotion that finished.
- This month: File the entry in your learnings library and open it when you brief the next promotion.
- Advanced: Add a standing "questions for sales, agency and vendor" step to your promotion checklist.
Advanced guide for Advertising & Promotions Manager professionals. These techniques use more sophisticated AI features that may require paid subscriptions.