top of page
Nov 13
5 min read

AI Email Marketing Analysis: From Messy Reports to a Decision You Can Check


By Trayan Hristov · Thrivelia · Published 13 November 2026


AI email analysis workflow: define comparable metrics, show calculations, and investigate explanations before making a decision.

An AI assistant can produce a convincing explanation for a falling revenue chart before it has checked whether the two reports use the same definition. I would make the first job smaller: establish what the numbers actually say.


AI email marketing analysis works best when the question, source and calculation are explicit. Give the assistant a bounded dataset, ask it to show the arithmetic and separate the observed change from possible explanations. Then use that output to choose an investigation.


Build an analysis packet before asking why


Include the two reports, their date ranges and the definitions behind them. State whether the revenue is store sales, email-attributed revenue or retained revenue after refunds. Add currency, time zone, conversion metric, attribution settings and what the rows represent.


For a file, name the tabs and relevant columns. For a connected account, ask the assistant to confirm which account and reports it can access. If data is paginated, ask how many records it read and whether that matches the available total. An incomplete first page is not a complete account analysis.


Inspect these sources before analysing performance: [files or account reports]. List the date range, row count, currency, metric definitions and missing fields for each. Check whether the comparison is like-for-like. Do not explain a performance change until you have identified differences in definitions or coverage.

Klaviyo's AI analytics overview describes answers with source and query visibility. Use those details to inspect the answer. A citation identifies where a number came from; it does not automatically make the interpretation correct.


Separate sending volume from yield


The following dataset is fictional. Both periods use the same definition of email-attributed revenue and delivered messages. It shows why a lower revenue total does not automatically mean worse message performance.


Measure

Period A

Period B

Delivered messages

1,000,000

800,000

Email-attributed revenue

$200,000

$184,000

Attributed revenue per 1,000 delivered

$200

$230


Revenue is down 8%. Delivered volume is down 20%. Attributed revenue per delivered message is up 15%. Those observations can all be true at once.


One exact arithmetic decomposition is: volume effect = change in deliveries × Period A revenue per delivery. Here, negative 200,000 × $0.20 = negative $40,000. Yield effect = Period B deliveries × change in revenue per delivery. Here, 800,000 × $0.03 = positive $24,000. Together they reconcile to the $16,000 decline.


That decomposition uses a stated calculation order. Other decompositions can allocate the interaction differently. More importantly, these are arithmetic components, not causal findings. 'Yield' can change because the audience, offer, product, timing or attribution changed.


Calculate the absolute and percentage changes in delivered volume, attributed revenue and revenue per 1,000 delivered. Show the formulas. Reconcile the revenue difference using a volume component and a yield component, stating the calculation order. Then list explanations that require additional evidence. Do not attribute the change to creative quality from these totals.

Ask the question the headline cannot answer


The next question is why delivery fell. Did fewer people enter the journey? Did exclusions improve? Did an integration stop passing events? Did the team send fewer campaigns? Did a delivery problem reduce successful sends? Those possibilities require different actions.


Split campaigns and flows. Within flows, inspect entries, eligibility, attempted sends, deliveries and the relevant outcome. Within campaigns, compare audience composition, offer, product and timing. Avoid combining unlike messages into a single performance verdict.


If a flow's revenue per recipient rises after its audience becomes more selective, that can be a useful observation. It does not tell you whether the removed recipients would have produced profitable orders. Ask what the decision was designed to improve.


Make customer retention comparisons at the same age


For second-purchase analysis, a customer acquired in January has had more time to return than someone acquired in October. Choose a fixed observation window before asking AI to compare cohorts.


Compare first-to-second purchase within [90] days for these first-order cohorts. Include only customers with a complete [90]-day observation window. State how customer identity, cancellations, refunds, exchanges and replacements are treated. Show eligible first-time customers and qualifying repeat purchasers separately. If the source cannot support this calculation, tell me which fields are missing.

A fictional cohort of 1,000 eligible first-time customers with 180 qualifying second purchasers has an observed 90-day rate of 18%. If another cohort has only 30 days of follow-up, do not label its customers as failed 90-day repeat purchasers.


Keep the first product, acquisition offer and customer mix visible. A change in retention can reflect who was acquired as well as what happened after purchase. Our retention measurement guide covers that wider interpretation.


Turn findings into a review memo


Part of the memo

What a useful answer contains

Observation

The metric, comparison, calculation and source

Interpretation

A plausible explanation with its assumptions

Missing evidence

The source or test needed to distinguish explanations

Next action

A named owner role and a bounded check


For the worked example, a useful memo might say: 'Revenue declined $16,000 while delivery fell 20% and yield rose 15%. Inspect the delivery reduction before commissioning a creative rewrite. The campaign and flow split is missing, so we cannot yet identify the affected journey.'


That is more useful than a confident list of five reasons customers stopped buying. It narrows the work without pretending to have completed the diagnosis.


Produce a one-page review memo from the verified calculations. Separate observations, hypotheses and missing evidence. For each recommendation, identify which finding supports it and what would change your view. Give me no more than three next actions. Do not forecast revenue uplift without a documented model.

Check the answer in the source report


  1. Recalculate one headline number independently from the original totals.

  2. Check that percentages use the intended denominator and percentage-point changes are labelled correctly.

  3. Confirm the complete period and the account time zone.

  4. Inspect a sample of the records behind any segment or cohort claim.

  5. Check that missing values were not silently converted to zero.

  6. Keep attributed performance separate from evidence of additional sales.


The last distinction matters when an assistant writes 'AI generated an extra $X'. A report can show attribution under chosen rules. A suitable experiment is needed to estimate what changed because of the intervention. See our attribution and incrementality guide.


If the answer is wrong, ask for the exact calculation or source selection to be corrected. Generating a fresh narrative without correcting the underlying work can leave the same error hidden in better prose.


Where the analysis should lead


The result should help decide whether to repair a route, improve a product explanation, change a contact policy or run a focused test. At Thrivelia, I would connect the analytical finding to a journey brief and an owner, then revisit the original question after the change.


Use the daily workflow for CMOs and founders to place this review inside the week. If your team needs help getting from reports to a prioritised programme, explore our services.


 
 
 

Recent Posts

See All

Comments


Commenting on this post isn't available anymore. Contact the site owner for more info.
bottom of page