AI Marketing in Practice
How to Use AI to Draft Commentary for a Monthly Marketing Report
Draft monthly marketing commentary from a controlled export by separating checked observations, hypotheses and questions for action.
Use AI to help structure monthly report commentary, not to invent a reason for every change. Start with a controlled export, checked metric definitions, a date range and independently checked arithmetic. Then separate what the figures show from what they might mean and what you should investigate next.
This method drafts commentary from supplied data. It does not provide live analytics access, select a metric definition, prove causation or claim a business outcome.
Give the numbers a definition first
A number is only useful when the reader knows what it counts. Before drafting, record:
- metric name and definition;
- source and export date;
- date range and comparison period;
- filters, segments or attribution notes; and
- any missing or changed data.
“Traffic” could mean sessions, users, page views or something else. “Leads” could mean every form submission or only enquiries that meet a stated qualification. Use the source's definition rather than a familiar label.
Check the arithmetic independently
Recalculate totals, differences and percentage changes before asking AI to write commentary. A simple formula is enough for a first pass:
Percentage change = (new value - comparison value) / comparison value × 100
Record the comparison value and rounding. If the comparison is zero, missing or not like-for-like, do not force a percentage. Explain the limitation instead.
A controlled illustrative export
ILLUSTRATIVE EXAMPLE: The following table is fictional teaching material. It is not live analytics access, a client report, an employer result or a measured business outcome.
| Metric | Definition | April | May |
|---|---|---|---|
| Article page sessions | Sessions on the selected article page | 120 | 150 |
| Resource-link clicks | Clicks on the selected resource link | 24 | 30 |
| Qualified enquiries | Enquiries marked qualified under the team's stated definition | 8 | 9 |
The arithmetic is checked for this example. Sessions increased by 30, which is 25% of April's 120. Resource-link clicks increased by 6, also 25%. Qualified enquiries increased by 1, which is 12.5% of April's 8.
Those calculations describe the table. They do not explain why the values changed or prove that an AI-assisted activity caused the change.
Keep observation, interpretation and question separate
Draft the report in three columns. Do not make the interpretation do the job of the evidence.
| Observation | Possible explanation | Question for action |
|---|---|---|
| Article page sessions rose from 120 to 150, a checked 25% increase. | Hypothesis: more relevant visitors may have reached the page, or the source mix may have changed. | Check source/medium, campaign dates and any tracking change before explaining the increase. |
| Resource-link clicks rose from 24 to 30, a checked 25% increase. | Hypothesis: the page or call to action may have been more visible to May visitors. | Compare page version, placement and segment before changing the call to action. |
| Qualified enquiries rose from 8 to 9, a checked 12.5% increase. | Hypothesis: the mix or follow-up process may have changed; the two-month table cannot identify a cause. | Check the qualification definition, enquiry source and sales notes. |
The first column is what the export supports. The second is a hypothesis, not a conclusion. The third turns uncertainty into a next check.
Ask AI for bounded commentary
Give the tool the checked table, definitions, date range and the three-column output shape. Ask it to preserve the numbers, show arithmetic, label hypotheses and write questions where the export is silent. Tell it not to add causes, figures, customer stories or outcomes that are not supplied.
AI can help turn a working record into readable commentary. A person still checks every number, metric definition, qualification and proposed action against the source.
Keep attribution and outcome claims modest
Two months of movement can be worth reviewing, but it is not enough to assign a cause automatically. Audience mix, campaigns, seasonality, tracking changes, page edits, sales follow-up and random variation may all matter.
Do not write “the AI workflow increased enquiries” when the export only shows that enquiries rose from 8 to 9. A safer line is: “Qualified enquiries rose by one in this illustrative comparison. Check the definition, source mix and follow-up record before interpreting the change.”
If a business outcome matters, define the measure, comparison and context before the reporting period. Keep the explanation as a hypothesis until evidence supports a stronger statement.
Check freshness and ownership
Metric definitions and data pipelines can change. Record who owns the source, when the export was produced and which definition or filter was used. If a dashboard changes its naming or attribution, flag the break rather than comparing unlike periods silently.
Ask the responsible owner to confirm unusual values, missing rows or a changed qualification rule. The report writer can describe the observation without pretending to resolve a data-quality issue.
Protect data and access
Use the minimum data needed for the commentary. Remove names, email addresses, private customer notes, confidential employer material and credentials before sending a table to an AI tool unless an authorised process explicitly permits them.
This article uses a fictional export. A real report may have additional privacy, security, commercial or access controls. The same observation/interpretation/question structure does not replace those controls.
Review the drafted commentary
Before the commentary enters a report, ask:
- Does every number match the controlled export?
- Are definitions, date range and comparison visible?
- Is each explanation labelled as a hypothesis or supported by separate evidence?
- Does each question for action follow from the observation?
- Have causation, ROI, ranking, revenue or performance claims been removed unless appropriately supported?
- Is any private or restricted information present?
If a question cannot be answered from the export, keep it as a question. A clear unknown is more useful than an invented conclusion.
What this method does not do
This method does not provide analytics access, choose attribution rules, perform a full data-quality audit, establish consent, prove causation or replace a reporting owner. It makes a small set of definitions, observations and next checks easier to review.
Your next step: write three observations
Choose one controlled monthly export. Check its definitions, date range and arithmetic. Write three observations without causes. Label one possible explanation as a hypothesis, then write one question that could test it.
Ask an authorised reviewer to confirm the definitions and action questions. Keep any missing data visible and do not turn a short comparison into a business conclusion.
Further Reading
You Might Still Be Wondering...
Frequently asked questions
Give it a controlled export, checked definitions, date range and a required structure that separates observations, hypotheses and questions. Ask it not to add causes or outcomes, then have a person verify the result.
Show the comparison values and calculation, then write the change as an observation. Put possible reasons in a hypothesis field and add a question that could test them.
Not by themselves. Two months may show a change worth investigating, but audience, source, timing, tracking and follow-up changes can offer alternative explanations. Use appropriate comparison and context before making a causal claim.
Mark the comparison as limited or not comparable. Record the old and new definitions, ask the data owner what can be compared and avoid presenting the difference as a clean trend.
Only when it is necessary, authorised and covered by the relevant process. Otherwise aggregate or generalise the export and remove identifying or confidential information.
End with a proportionate question or next check that follows from the observation. The question may be to validate a definition, inspect a source, review a page change or collect more evidence.