If the reports don't reconcile, the way out is not to give up: it is to design a measurement that doesn't need to reconcile them. Three steps, from least to most powerful, that turn a "non-attributable" figure into a decision.
Step 1 · Split by source
First you split each campaign's spend into its sources, exactly as Google labels them in the AI Max search terms report:
| Source | What it is |
|---|---|
| Your keywords (baseline) | Terms that hang off one of your keywords with its match type |
| AI Max · dynamic broad | Terms triggered by expansion of your keywords |
| AI Max · keywordless | Terms with no anchoring keyword |
| No breakdown | Whatever the thresholds prevent being assigned |
Each source with its cost, conversions, CPA, ROAS and value. With this alone you can see whether AI Max, in the part Google does attribute, costs more or less per conversion than your keywords. In the study's healthy account, dynamic broad (€11.21) even beat the baseline (€11.48); in the real estate account, every AI Max source was worse.
The limitation: step 1 only judges the visible part. The "no breakdown" third stays outside… until step 2.
Step 2 · Before and after: how the invisible becomes measurable
The "no breakdown" bucket looks impossible to judge: by definition you don't know whether it is baseline or AI Max. But there is a way out: compare it with itself before and after switching AI Max on, using the keyword baseline as a control for seasonality.
The reasoning: seasonality and the market affect everything in the account equally; that is why the baseline works as a witness. If, when AI Max is switched on, the "no breakdown" bucket shoots up —it grows in weight or its CPA worsens— far more than the baseline over the same period, that excess over the control is no longer explained by the market: it is AI Max's fingerprint, pushing traffic into the blind spot. If, on the other hand, the bucket rises and falls in line with the baseline, it is seasonality, and AI Max is acquitted.
It is a difference-in-differences design:
Excess = (Δ «no breakdown» POST vs PRE) − (Δ baseline POST vs PRE)
In one sentence: the "no breakdown" bucket on its own says nothing; its variation relative to the baseline around the activation date says almost everything.
Requirements: knowing the activation date per campaign (lesson 1), a PRE window and a POST window of similar size (2-4 weeks), and enough volume in both. It is calculated per campaign and in aggregate.
Step 3 · Reconciling the combinations
As a direct test of traceability: take the highest-spending AI Max combinations (term + generated headline + URL) and try to match each one with the search terms report to recover its conversions. The percentage that reconciles does not measure performance: it measures observability — how much of the spend you can say anything about. In the study: 1% in account A, 36% in account B. The lower it is, the more your verdict depends on step 2.
What the method produces
Per campaign, a verdict with three pieces:
- Visible sources: AI Max CPA/ROAS vs baseline (🟢 the same or better / 🔴 worse).
- Opaque bucket, Pre/Post: excess over the control (🟢 absent / 🔴 present).
- Observability: % reconcilable (how reliable the above is).
And an action: nothing (🟢), investigate specific campaigns (🟡), or reduce AI Max (switch components off, exclusions, negatives) in the red campaigns.
The mistakes of a home-made analysis
- Comparing the campaign's CPA before and after AI Max with no seasonality control: it confuses the market with the tool.
- Judging the visible sources only and ignoring the "no breakdown" bucket.
- Adding up views that don't reconcile.
- Uneven PRE/POST windows, or windows with other changes inside them (bids, budgets, creatives).
- Looking once: AI Max drifts; the useful control is continuous.
💡 Ninja trick: AI Max Analyzer (free in the panel) runs the three steps every night: the split by source with CPA/ROAS, the Pre/Post of the opaque bucket with the baseline as control and the activation date detected per campaign, the reconciliation of combinations, and a traffic-light verdict per campaign with an explicit recommendation. What took weeks of analysis in the study, in a sheet every morning.
What you should remember
- Step 1: split by source (baseline, dynamic broad, keywordless, no breakdown) and compare CPA/ROAS in the visible part.
- Step 2: the opaque bucket is judged before/after with the baseline as the control — the excess over the control is AI Max's fingerprint.
- Step 3: reconciliation measures how much you can actually claim.
- Activation date, comparable windows, no other changes, and continuous control.