Anomaly root-cause explanations

When CPA, CAC, or ROAS moves outside an expected range, the hard part is rarely spotting the spike. It is assembling a credible explanation from spend logs, channel mix, creative changes, attribution windows, and CRM outcomes before the next planning cycle. This page describes how an LLM can draft anomaly root-cause explanations for marketing analysts, with analysts verifying every claim before anyone changes budgets or creative.

Related reading: Auto-generated review decks, Auto-narrative performance reports, and Conversational reporting in Slack.

Anomaly root-cause explanations turn a detected KPI shift into a structured draft that lists plausible drivers, supporting evidence from connected sources, and open questions that still need a human check. The model does not replace anomaly detection or budget decisions. It reduces the time spent stitching screenshots, export files, and Slack threads into a first pass narrative.

Typical inputs include paid media platform metrics, ad account spend and impression data, landing page or analytics events, CRM or pipeline fields that feed CAC, and a short change log of launches, bids, audiences, or tracking updates. Typical outputs are a dated explanation draft, ranked candidate causes with confidence notes, and citations back to the underlying tables or reports the model used.

The outcome is speed: analysts get a readable first draft sooner, then spend their time confirming or rejecting hypotheses instead of reconstructing the timeline from scratch.

Use this pattern when a monitored KPI moved enough to warrant investigation, and when the supporting data for that window is available. Empty or incomplete source pulls should produce empty output rather than a speculative story. If the anomaly window is missing, or required sources (for example spend by campaign and conversion counts for the same dates) cannot be loaded, the system should return no explanation and flag the gap.

Good fits include weekly performance reviews, post-launch checks after creative or bidding changes, and mid-funnel reviews when CAC drifts while top-of-funnel volume looks stable. Poor fits include one-off metrics with no baseline, channels with broken tracking for the period in question, or cases where finance and marketing definitions of CAC disagree and have not been reconciled.

A practical workflow starts after an anomaly rule or analyst selects a KPI, a date range, and a comparison baseline (prior period, same weekday average, or a campaign cohort). The system retrieves aligned slices from each connected source for that window and the baseline, then asks the model to propose causes that are consistent with the observed deltas.

The draft should separate observed facts from inferred links. Facts are measurable: spend up X%, conversion rate down Y%, a new creative set launched on date D, attribution lookback changed. Inferences are interpretive: “creative fatigue likely contributed” or “audience overlap may have raised CPA.” Inferences need explicit hedges and pointers to what an analyst should verify next.

Human-in-the-loop is mandatory. The model drafts explanations; analysts verify before acting. Verification usually means checking that the cited campaigns and dates match the source UI, confirming that conversion definitions match the KPI in the dashboard, and rejecting causes that ignore known offline events (promotions, outages, inventory limits) not present in the connected data.

Strong drafts cite multi-source agreement. For example, ROAS fell while spend rose and revenue stayed flat: the draft should show spend by channel, revenue attribution for the same window, and any change in discounting or product mix if those fields exist. For CAC, the draft should connect media cost to the CRM cohort that actually entered the pipeline, not only to lead form fills that never progressed.

Weak drafts invent single-cause stories from one chart. CPA rising after a bid increase might be expected efficiency decay, a tracking lag, a landing page regression, or a mix shift toward colder audiences. A useful explanation lists competing hypotheses, notes which data supports each, and marks which hypotheses cannot be tested with the current sources.

When creative or campaign change logs are available, include them as timeline anchors. When they are not, the draft should say so and avoid backfilling a launch narrative from metric movement alone.

Define the KPI formulas and attribution windows in configuration, not in free-form prompts, so the model reasons against the same definitions the dashboard uses. Keep a short allowlist of sources per KPI so CAC drafts never silently mix MQL counts with closed-won if that is not how CAC is calculated.

Require source freshness checks before drafting. Stale CRM syncs or delayed platform APIs should block output or label the draft as incomplete rather than present a finished explanation. Preserve the anomaly window, baseline, and query timestamps on every draft so later reviewers can reproduce the pull.

Do not auto-apply budget or bid changes from an explanation draft. Route the verified narrative into review decks, narrative reports, or Slack threads as context for a decision that a person still owns.

Return empty output when the anomaly window is unspecified, when baseline and comparison periods cannot be resolved, or when required supporting sources are missing or fail to join on campaign and date keys. Prefer silence over a polished paragraph that cannot be audited.

Common failure modes include attribution lag mistaken for performance collapse, currency or timezone mismatches across platforms, and double-counting conversions when the same conversion fires in ads and analytics. Analysts should treat model confidence language as a prioritization hint, not proof. If two sources conflict, the draft should surface the conflict instead of averaging them away.

Used this way, anomaly root-cause explanations shorten the path from “something moved” to “here is a verified shortlist of why,” without pretending the model has complete visibility into every marketing change.