Weekly campaign analysis with AI

A good weekly total can hide a weak ad

Imagine an ad spends £1,200 and brings in £3,600 of attributed revenue. A 3.0 ROAS might put it on the list for more budget. However, most of the revenue came from two days:

The useful question is whether those two days are repeatable: a promotion, a placement shift, delayed conversions or ordinary variation could lead to different decisions. Keep both views. Removing inconvenient days is not a way to discover the “real” result.

This is the kind of investigation the workflow should trigger before producing a confident recommendation.

Give each part of the setup a clear job

Start with platform exports, a spreadsheet or calculation script, and an LLM that performs well on campaign analysis. The export supplies the observations; the calculations establish what changed; the model proposes explanations and follow-up checks. You decide whether the evidence justifies an action.

Use daily rows with stable campaign and ad IDs, spend, conversions and conversion value, plus the preceding comparison period. Keep the conversion event, attribution window, currency and reporting timezone with the file. Add placement or other breakdowns when the question requires them and the platform supports that combination.

Post-click and post-view reporting can help investigate an attribution shift where those measures are available. Keep their definitions visible; do not add overlapping reporting windows or pretend a missing breakdown is zero.

Run the analysis in two passes

First, establish what the report can support. Check duplicate IDs and dates, missing days, changed conversion definitions and totals against the source report. Calculate CPA from total spend divided by total conversions, and ROAS from total conversion value divided by total spend. Averaging row-level ratios gives small rows too much influence; zero denominators should stay undefined.

Check reporting freshness too. Google notes that conversion data is not available instantly and that some fields are incompatible with some reports. A recent deterioration may need another export before it warrants a decision. Google’s conversion-reporting guidance.

Then investigate the changes that could affect a decision. Give the model the checked summary and the relevant rows. Ask it to identify the day, ad or placement responsible for a change, and to request a missing breakdown when competing explanations remain. In the example above, it should ask what happened on those two days before suggesting more budget.

Return a decision, with the evidence beside it

Keep the weekly output short. For each material finding, require the exact rows or calculation, the proposed action and what could change the recommendation:

• Act: the supplied evidence supports a specific change for your review.
• Investigate: name the competing explanations and the next report needed to distinguish them.
• Hold: there is not enough evidence to justify a change. An ordinary week is a valid result.

Save the reporting settings and instructions once. Before automating the export, check that repeated runs reproduce the same totals and retain the breakdowns the investigation needs. Keep budget changes as a separate decision.