Reading and interpreting analytics with AI · AI for Marketing & Content

The most common, the most fluent, and the most dangerous AI mistake in marketing is the answer to “why did our numbers change?” A model handed a vague question and a sketchy data picture will produce a confident, well-structured, three-paragraph explanation involving seasonality, audience shift, and a launch — none of which it has actually verified, all of which sound plausible, and which you will quote to your director in the Monday meeting before discovering, two weeks later, that the real cause was a tracking break on the checkout page. This section is about how to use AI to read GA4, Meta Business Suite, Mailchimp, and the inevitable mixed-source spreadsheet without falling into that trap.

The do-not-invent rule, ported to analytics

The rule that governed the management-report workflow ports here directly, with one extra clause. The spreadsheet owns the numbers; the model owns the prose around them; the model is explicitly forbidden from supplying numbers that are not in the data. The extra clause exists because analytics data is messier than a P&L — multiple sources, multiple time windows, gaps, anomalies — and the temptation for the model to “round out” a story with plausible figures is higher.

The prompt instruction that does the work, every time, is some version of this: “Use only the numbers in the data below. Do not state any figure that is not in this data. If a metric I have asked about is not in the data, say so explicitly. Do not estimate.” It costs you four lines. It saves you the embarrassment of presenting a metric in a board meeting that the model invented to make a paragraph flow.

The workflow

A practical end-to-end shape, refined over enough campaigns to know where it breaks, looks like this.

Step one — export. Pull the data from the actual tool. GA4’s explore reports, Meta Business Suite’s campaign export, Mailchimp’s campaign report, Search Console’s query export. Resist the temptation to screenshot a dashboard and paste the screenshot into the model; the model reads tables far better than it reads charts, and the chart often hides the row that would have answered your question. CSV or a clean table in the prompt, every time.

Step two — frame the question. Not “what does this data tell us?” — that produces a meandering paragraph. A specific question: “Sessions to our blog fell from 48,000 in March to 31,000 in April. The data below is the channel split, the top ten pages by traffic, and the search-console query export for both months. Use only these numbers. Produce a structured list of candidate explanations — ranked by how well each is supported by the data — covering at minimum: a tracking or measurement change, a seasonality effect, a change in channel mix, a change in audience, a launch or campaign change on our side, and an algorithm or platform change on the source side. For each candidate, state explicitly what in the data supports it and what would falsify it. Do not pick a single answer.”

Step three — verify against the source. The candidate-list output is a thinking tool, not a conclusion. Take the top one or two candidates and verify them against the source — pull the GA4 report directly, check the date range, look at the channel that the model flagged, see if the pattern is actually there. About a third of the time, the model has correctly identified the most likely cause. About a third of the time, the most likely cause is genuinely something the model proposed but which it ranked third or fourth. About a third of the time, the real cause is something the model did not propose at all, and the verification step is what surfaces that gap.

Step four — decide and write. Now the marketer writes — or asks the model to draft — the actual conclusion. The traffic fell because of X; the evidence for X is Y; we are doing Z about it. The model can draft this sentence cleanly once the marketer has decided what X, Y, and Z are. It should not be deciding X.

Where AI is genuinely useful

Once you have the discipline, the productivity gain in analytics work is large and worth being specific about.

Pattern spotting across many metrics quickly. A typical campaign report contains forty or fifty numbers: impressions, reach, frequency, CTR, CPC, CPM, landing-page views, scroll depth, time on page, conversions, conversion rate, cost per conversion, retention by cohort, channel split, device split, geographic split. The marketer can hold maybe a dozen of these in their head at once. The model can scan all forty and surface which three or four have moved the most relative to baseline. You decide whether those movements matter; the model has done the looking.

Plain-language description of what changed. Turning “CTR moved from 1.8% to 2.4%, CPC fell from NPR 32 to NPR 27, conversion rate held at 1.1%” into “the ad is being clicked more and costing less per click, but the same proportion of those clicks is converting — the gain is in efficiency, not in the funnel below the click” is exactly the kind of translation work a model does well. The marketer’s hour becomes ten minutes.

Candidate causes you might have missed. Asked for a comprehensive list of possible explanations, the model often surfaces the one you forgot — the bot-traffic spike that inflated sessions, the tracking-tag change the developer pushed last Tuesday, the holiday in a feeder country that shifted a small audience segment. About one campaign in five, this single use justifies the workflow.

Cross-source narrative. When the data is split across GA4, Meta, and Mailchimp — and the campaign ran across all three — the model is genuinely useful at producing a unified narrative that traces a cohort from email-open to ad-click to landing-page-visit to signup. You verify each leg; the model joins them.

What stays the marketer’s job

Stay equally clear about what the model cannot do, because the failure mode in analytics is letting the fluent paragraph substitute for the judgement.

Deciding which patterns matter. A 40% increase in sessions from a 2,000-person geographic segment may or may not matter to the business. The model has no way to know. The marketer does.

Deciding which candidate cause is real. The model can rank candidates; only the marketer knows that the developer pushed a tracking change on the 14th, that the competing app launched a referral promo on the 19th, that the sales team paused outbound for two weeks.

Deciding what to do about it. The model can draft a recommendation if you tell it what the recommendation is. It cannot, in any responsible sense, decide the recommendation. That is the job.

Two concrete Nepali scenarios

The Khalti referral campaign that drove low signup despite high traffic. A campaign brings 24,000 sessions and 312 signups in a week — a 1.3% conversion rate against a 4.1% baseline on the same landing page. The model, handed the channel split, the device split, the time-of-day split, and the funnel report, produces a ranked candidate list: the campaign skews to a younger audience that traditionally converts at lower rates here; the device split is 84% Android low-end where the signup flow has known performance issues; the time-of-day skew is heavy after 9 PM when our verification SMS sometimes delays. The marketer verifies the device-performance theory against the page-speed logs, finds a real degradation in the Android flow that week, and ships a fix. The model surfaced the right candidate; the marketer confirmed it; the data did the rest.

A fintech blog that lost organic traffic over a quarter. Sessions to the blog fall from 62,000 to 41,000 across Q1. The model, handed the GA4 channel data and the Search Console query export, produces a ranked candidate list: an algorithm update that hit financial-content sites in February; the loss of three queries that previously accounted for 28% of traffic; a sister site at the same firm that started competing for the same queries; a slowdown in publishing cadence from 8 to 3 posts per month. The marketer checks each candidate against the underlying data, finds that the queries-loss and the sister-site cannibalisation together account for most of the drop, and adjusts the editorial plan. None of the conclusions were the model’s; all of the candidate framing was.

Check your understanding

Quick check

—

You paste a GA4 channel export into the model and ask it to explain a traffic drop. Which prompt addition most reduces the risk of the model fabricating a metric or a cause?

Be creative and consider all possible explanations.
Use only the numbers in the data below; do not state any figure that is not in this data; if a metric I have asked about is not in the data, say so explicitly; produce a ranked list of candidate explanations with what supports and what would falsify each — do not pick a single answer.
Write a confident, executive-ready paragraph explaining what happened.
Use industry-standard benchmarks to fill in any missing context.

What comes next

Once you have a defensible read of what happened and why, the next problem is communication. A campaign-by-campaign analytical narrative is not, on its own, what a stakeholder reads on a Monday morning — and the gap between the analyst’s view and what a director can act on in thirty seconds is where many marketing reports quietly die. The next section is about how to use AI to write the stakeholder report itself, without letting it substitute fluent prose for the specific evidence the report has to carry.