AI can summarize your channel numbers in seconds. Here's how to run the monthly review so the arithmetic and the explanations stay checked by you.
Use AI to read, group and draft. You recompute every number you’ll quote, label every cause as a hypothesis with a test, and score the summary on eight questions before it goes anywhere. The prompt, ledger, checklist and one-page template are all below.
AI is good at the part of the monthly marketing review that people put off, which is turning a pile of exports into readable paragraphs. It’s less dependable at the two parts that make the review worth holding: getting the arithmetic right, and not mistaking a guess about why for something the numbers show. Both failures tend to arrive in the same confident tone as the correct answers, so you can’t catch them by how the summary reads.
A marketing manager pastes the month’s channel export into a chat assistant and asks what changed. The summary comes back in seconds with a total spend figure at the top. It doesn’t match the total she has open in the ad platform, and it takes a few minutes to find out why: the export carried its own totals row, and the summary had counted it as one more campaign. Everything else in the summary was reasonable, which is what made the one wrong number believable.
The better move is to split the work. AI reads, groups, compares and drafts. You own every number that gets quoted and every claim about cause. The rest of this guide is the version of that split I’d run each month: a clean export, a prompt that forces the working into view, a ledger of what to recompute, a way to keep what happened apart from why, a scored check, and a one-page review to send.
Microsoft says something similar about its own tool. Its support page for the COPILOT function in Excel states that the function can give incorrect responses, and tells people to use native Excel formulas such as SUM, AVERAGE and IF for any task requiring accuracy or reproducibility.[1] That is one vendor’s guidance about one feature, read on October 9, 2026, but the habit travels: let the assistant draft and spot patterns, and let a formula or your own recalculation settle the number.
This guide covers the recurring monthly review. For a standing dashboard, see our KPI dashboard guide; for a one-off file, analyzing a spreadsheet with AI; for the attribution-model debate, AI for marketing attribution. The checks below work with any assistant.
A lot of wrong numbers in an AI summary start in the export, before the model reads anything. Totals rows, merged header cells, numbers stored as text and two periods stacked in one column are hard for a person to notice and a trap for a model. Ten minutes of cleaning before you paste saves more than ten minutes of arguing with the output later.
That last point catches more people than the formatting does. Google says data processing in Google Analytics can take 24 to 48 hours, and that some data can arrive late, up to seven days afterwards.[2]
Google Ads has its own version. Its help page says conversion data can take up to seven days to fully appear, and that Google Ads reports conversions against the date of the ad, where other reporting tools use the date of the conversion.[3] So a first-of-the-month export can understate the last few days, last month’s figure can move after you’ve reported it, and two tools can both be right while disagreeing. I’d run the review a few days into the new month and put the export date on the page.
Don’t compare two numbers until you know they were counted the same way.
Write the basis next to each source in your prompt: the date logic, the attribution window, the filters. Don’t ask the assistant to reconcile two platforms that count differently; report each on its own basis and say so. If GA4 is one of your sources, our GA4 reporting guide covers getting clean exports out of it.
Here’s the prompt I’d start from. Two features matter more than the wording: it makes the model show the formula behind every figure, and it keeps “what the data shows” and “what might explain it” in separate sections.
| I’m preparing a monthly marketing performance review. Below is an export with one row per [campaign / channel] per month. Period under review: [month]. Comparison period: [prior month / same month last year]. Definitions: conversion = [your definition]; attribution window = [per source]; spend = [currency, tax in or out]. Data exported on [date]. Work in this order. 1) Restate the columns, row count and periods you see. Flag anything that looks like a totals row, duplicate, blank, or number stored as text. If periods or definitions are unclear, stop and ask me. 2) Build a table by channel with spend, conversions and cost per conversion for both periods, and the change. For every figure you calculate, show the formula in words with its inputs, for example: cost per conversion = spend / conversions. Calculate ratios from summed totals. Do not average ratios. 3) Write WHAT THE DATA SHOWS: up to five statements, each tied to specific rows. No causes in this section. 4) Write WHAT MIGHT EXPLAIN IT: for each statement, up to two explanations, each labelled a hypothesis, with one thing I could check to test it. 5) List anything you could not calculate or had to assume. Rules: use only the numbers I pasted. Do not estimate missing values. If something is not in the data, write [NOT IN DATA]. Do not recommend budget changes yet. |
Fill the brackets once and keep the prompt in a shared doc. Next month only the export and the period should change, so a different answer is down to the data. Our 4-part prompt formula explains why each part earns its place.
Assistants differ in how they handle an uploaded file, and I haven’t verified how each one does it. You don’t need to: the checks in the next section test the output, not the method. The prompt holds back recommendations on purpose, because a budget suggestion built on an unchecked total is the costliest place for an error to land.
Not every number in a summary deserves the same scrutiny. A row count is cheap to confirm, while a ratio that leads your headline is where a slip costs you a decision. The ledger sorts the usual figures by what AI computes, what you verify, and how.
| Figure in the summary | What AI computes | What you verify | How you verify it |
|---|---|---|---|
| Totals (spend, conversions, revenue) | Adds up the columns | That the sum matches the source and nothing was added twice | Sum the column yourself; compare with the platform’s total; look for a totals row counted as a campaign |
| Ratios (cost per conversion, click-through rate, conversion rate, return on ad spend) | Divides one figure by another | That the ratio comes from summed totals, not an average of row ratios | Recompute from your summed numerator and denominator; check two rows by hand |
| Change versus last period | Percent change between two figures | Right base, right periods; points versus percent for rates | Recompute (new minus old) divided by old on the three biggest lines |
| Rankings (top and worst campaign) | Sorts by a metric | That the metric is the right one and the volume is big enough to mean something | Sort the sheet yourself; look at the denominators of the top three |
| Anything from outside the file | May add context or benchmarks from memory | That it’s real, sourced and wanted | Ask for [NOT IN DATA] markers; delete or source anything else |
| A sample of single rows | Quotes values back | That the values match the platform screen | Pick five rows yourself, not the assistant’s choice, and compare each with the source |
A Future Factors working method based on our own judgment, not a validated model. Add a final column, “Caught this month”, and keep a tally.
Two traps are worth seeing in numbers. First, averaging ratios. In the sample review later in this guide, the plain average of three channels’ cost per conversion falls 6.5% from one month to the next, while total spend divided by total conversions falls 4.8%. Both are arithmetic, but only the second answers the question “what did a conversion cost us?”
Second, points versus percent. A conversion rate that moves from 2.0% to 2.5% is up half a percentage point, or a quarter in relative terms, so ask for the unit whenever a rate appears.
The fifteen-minute check, in order:
| For every figure in the table from step 2 of your summary, give me a table with these columns: the figure, the formula in words, the exact inputs you used, which rows those inputs came from, and the result. Then recalculate each result a second way (for example, sum the rows in a different order or by channel) and flag any difference. Do not change any figure silently. If you change one, tell me which and why. |
Treat the output as a map of where to look. A model re-checking itself can repeat the same misreading, so it supplements steps 1 to 4 and doesn’t replace them.
Any number you’ll quote gets recomputed by you, or by a formula, and not by the assistant that produced it.
Our fact-checking workflow applies the same instinct to claims in text, which helps if your reviews also quote industry figures.
Summaries drift from description to explanation without changing gear. “Paid search got more expensive” and “paid search got more expensive because competitors increased bids” read as the same kind of sentence, and only the first is in your data.
A social lead reads an AI summary saying engagement rose because short-form video is performing better. The first half is visible in the export: the video posts have higher engagement than the rest. The “because” isn’t anywhere in the file. When she sorts the post list, most of the month’s engagement sits on a small number of posts, and a couple of those were promoted, which the export marks in a column the summary didn’t mention. The finding that survives is narrower than the one on screen.
| What the data shows | What might explain it (hypotheses) | A test before you say it |
|---|---|---|
| Paid search cost per conversion rose 11.1% as spend rose 10.0% | A bidding or targeting change; more competition; seasonal demand | Ask the account owner what changed; compare with the same month last year |
| Paid social cost per conversion fell 20.0% on 12.5% less spend | The weakest ad sets were cut; new creative | Compare cost per conversion by ad set before the cut |
| Email conversions rose 14.0% on flat spend | A stronger send calendar; an offer in the period | Check which sends and offers ran |
Figures are placeholders from the sample review later in this guide. The middle column holds hypotheses only, and nothing in it should reach a slide as a finding until the right-hand test has been run.
Say what the data shows before anyone says why.
AI is a useful source of hypotheses, because it can list candidates you hadn’t considered. They become findings when a test confirms them or the person who ran the campaign says what changed. A follow-up prompt helps: “For each hypothesis, what data would confirm or rule it out, and is it already in the export?” This matters most for whoever presents, since a head of marketing in a leadership meeting will have any “because” heard as a decision.
Before a summary leaves your hands, run it through eight yes-or-no questions. A “yes” means you did the check, not that the assistant says it did.
| # | Question | Point |
|---|---|---|
| 1 | Do the periods and definitions in the summary match what I meant? | 1 |
| 2 | Did I sum the totals myself and match them to the source? | 1 |
| 3 | Did I recompute every ratio the summary leads with, from totals? | 1 |
| 4 | Did I compare five rows I chose against the platform screen? | 1 |
| 5 | Is every number in the summary traceable to the data I pasted? | 1 |
| 6 | Is every cause labelled as a hypothesis, with a test next to it? | 1 |
| 7 | Do the biggest changes rest on enough volume to mean something? | 1 |
| 8 | Does the page state the export date and any late-data caveat? | 1 |
A Future Factors working checklist based on our own judgment, not a validated instrument.
| Score | What to do |
|---|---|
| 7 to 8 | Ready to share. Note the one gap, if there is one, on the page. |
| 5 to 6 | Fix the missing checks first and rescore. Name the gaps if you share it anyway. |
| 3 to 4 | Use it as a draft for your own thinking. Don’t forward it. |
| 0 to 2 | Go back to the export, or do the key numbers by hand. |
Questions 2, 3 and 6 are the ones I’d refuse to skip. A summary can score a 5 and still be unsafe if those are the three gaps.
The checklist covers the page. This table covers the split of work behind it.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Reads the export, groups rows and flags oddities | Deciding the export is clean and the definitions are right | Read the prompt’s first-step output before anything else |
| Calculates totals, ratios and changes, and shows its formulas | Every number that gets quoted | The ledger: recompute from totals, spot-check five rows, tally what was caught |
| Drafts hypotheses for what moved | Which hypotheses are credible and what gets tested | Each carries a named test and owner before it’s mentioned in a meeting |
| Writes the first draft of the narrative | The decision, the recommendation and the sign-off | The reviewer reads the page against the trust score and data notes |
The split is a Future Factors working view. If your organisation has an approval process, add the approver’s name to the third column.
Whether the review ran is quick to measure. Whether it’s working takes more, and the lens I’d use is use, then persistence, then impact. Use means the ledger got filled in. Persistence means it’s still being filled in three months later without anyone chasing it. Impact means a decision changed, or an error was caught before the room saw it. The tally of things caught is the number to keep, because it shows whether the check earns its minutes.
This is the page that goes to the room. The figures are placeholders chosen so the arithmetic stays clear, and every number can be recomputed from the first table. Replace all of it with your own.
| Channel | Spend Aug | Spend Sep | Conv. Aug | Conv. Sep | Cost per conv. Aug | Cost per conv. Sep | Change |
|---|---|---|---|---|---|---|---|
| Paid search | 12,000 | 13,200 | 300 | 297 | 40.00 | 44.44 | +11.1% |
| Paid social | 8,000 | 7,000 | 160 | 175 | 50.00 | 40.00 | -20.0% |
| 1,000 | 1,000 | 200 | 228 | 5.00 | 4.39 | -12.3% | |
| All channels | 21,000 | 21,200 | 660 | 700 | 31.82 | 30.29 | -4.8% |
The last row is total spend divided by total conversions, not an average of the three rows above it. Spend and conversions are placeholders in one currency, not results from a real account.
| Field | Filled-in example |
|---|---|
| Headline (one fact) | Across all channels, cost per conversion fell from 31.82 to 30.29 in September (-4.8%), on spend up 0.95% and conversions up 6.1%. |
| What the data shows | Search cost per conversion rose 11.1% as spend rose 10.0% and conversions slipped 1.0%. Social fell 20.0% on 12.5% less spend and 9.4% more conversions. Email fell 12.3% on flat spend. |
| What might explain it (hypotheses) | Search: a bidding or targeting change, or more competition. Test: ask the account owner what changed. Owner: [search lead], [date]. Social: the weakest ad sets were cut. Test: cost per conversion by ad set before the cut. Owner: [social lead], [date]. |
| Decision or ask | Hold search spend flat until the test is back. Keep the social reallocation for one more month. Approval needed from [name] by [date]. |
| Data notes | Exported [date], a few days after month end. Google Ads conversions can take up to seven days to fully appear, so September may still move. Each platform reports conversions on its own attribution settings, so channel results may overlap. |
| Checks done | Totals summed by [name] and matched to platform totals. Ratios recomputed from totals. Trust score 7 of 8. Gap: only three of five rows spot-checked so far; two more before this goes out. Errors caught this month: [tally]. |
Square brackets mark the places where a name, a date or your own figure goes. Everything else is placeholder wording to show the standard.
Two lines get skipped. The data notes line protects you when a number moves after the meeting, and the checks line tells the reader how much weight the page can bear. Reports that leave both off tend to end up defending a figure instead of pointing to a caveat already printed on the page.
The structure is the same for everyone, but the checking effort goes to different places depending on what you own.
| Role | Check hardest | Let AI do more of | Watch for |
|---|---|---|---|
| Performance marketer | Cost-per-result ratios, date and attribution basis, row-level spot-checks against each platform | The first-pass anomaly scan across many campaigns, and the variance narrative | Late conversions revising last month, and platforms crediting the same sale |
| Content or social lead | Small volumes and single-post outliers; which denominator an engagement rate uses; promoted versus organic | Grouping posts by format or theme, and summarising comment themes in words | A causal story built on a handful of posts |
| Head of marketing | Cross-channel totals, which can overlap, and the link to finance or sales figures | The one-page narrative and the questions the room is likely to ask | A story that’s too neat, and a ‘because’ presented as a finding |
A Future Factors working view. The checks are the same; the amount of time each deserves is not.
The head of marketing’s overlap risk is real: platform-reported conversions can total more than your analytics or sales system shows, so the sign-off is on whether the story holds, not on any single figure.
Teams can set this up on their own: one owner for the monthly slot, the prompt and ledger in a shared doc, and a second person looking at the tally once a quarter. Where several people produce reports, practising on your own exports with an experienced trainer shortens the learning curve and gets everyone checking to the same standard; our corporate training is built around exactly that kind of work. In the terms we use for rollouts, training builds the skill, the monthly slot gives it somewhere to be used, and the second reviewer is the reinforcement that keeps it going after the first month.
If you only do one thing: take last month’s export, run the prompt, then do steps 1 to 3 of the fifteen-minute check and write down every difference you find. That list is your first ledger, and it will tell you which row of the table your own reports need most.
The fifteen-minute check is sized for one clean export. A multi-platform review takes longer, and your first month longer again. Time your own first run rather than trusting my estimate, and write it down so you can see it shrink.
Check your organisation’s AI policy and the tool’s data terms before you upload anything. Campaign-level spend and conversion totals are a common scope, while customer records with names or emails are a different conversation. If you aren’t sure, aggregate first and leave personal data out.
Treat the platform as the record for that platform and look for the reason in a set order: a totals row counted twice, different date logic, a different attribution window, late-arriving conversions, then filters.
With a handful of campaigns, do the arithmetic in the spreadsheet itself and use AI for the narrative and the hypotheses. The ledger shrinks to three checks: the totals, the headline ratio and the biggest change. The two-column habit costs almost no time, so keep it at any size.
It can draft it, and a draft saves real time. What stays with you is the sign-off, plus the hypothesis labels, the data notes and the checks line, which should survive into the final version. A report that’s lost those lines has lost the part that tells the reader how far to trust it.
Sources were read on October 9, 2026: Microsoft Support’s page on the COPILOT function in Excel, Google Analytics Help on data freshness, and Google Ads Help on conversion time columns. Software behaviour changes, so confirm vendor details on the vendor’s own page before you depend on them. I haven’t described how any specific AI assistant processes uploaded files, because I couldn’t confirm that from vendor documentation. The math-check ledger, the trust checklist and the one-page review are Future Factors working methods based on our own judgment, not research-validated models. The scenes in this guide are illustrations, not client stories, and the figures in the sample review are placeholders.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.