The method from the previous lesson was applied to two real accounts in different sectors, over the same months of 2026. The results were opposite. Together, they teach more than either of them on its own.
Case A · Real estate: when AI Max degrades
Bulk activation. The first finding, before looking at performance at all: of the 21 campaigns, 14 started generating AI Max traffic on the same day (12 May 2026) and 20 of 21 within a nine-day window. That pattern is not the trace of a campaign-by-campaign decision: it is the signature of an account-wide rollout by default. The decision was not the advertiser's.
The central finding: 99% cannot be traced. The 200 highest-spending combinations generated by AI Max were taken and each one was matched against the search terms report to recover its conversions. Only 2 of 200 (1%) could be reconciled; the other 198 returned "0" because that view applies a different privacy threshold. It is not that they didn't convert: they cannot be attributed. Over 99% of the highest-volume spend, the advertiser could not say whether it converted well, badly or not at all. Any "optimisation" over that mass was, by construction, blind.
The long tail you never see. The analysis captured 2,582 distinct terms served by AI Max in the window; Google only allowed the ones with enough volume to be seen (the 500 with the highest spend). Below that, an invisible tail. As a matter of hygiene, the account was already carrying hundreds of terms blocked by negatives — and the system added 605 exclusions in a month in the band that is hardest to see.
And when it finally can be compared, it loses. In the visible part, the AI Max sources performed worse than the baseline; and the Pre/Post of the opaque bucket came out red in all 21 campaigns: the "no breakdown" bucket grew and worsened far more than the control after activation. The fingerprint of AI Max pushing low-quality traffic into the blind spot, campaign by campaign.
Case B · Training: when AI Max delivers
Level with the baseline. Aggregate verdict 🟢: AI Max performs in line with the keywords. Metric by metric: CPA €11.48 against €11.48, ROAS 26.1× identical, an AI Max CPC 4% cheaper; and dynamic broad even beat the baseline (€11.21).
The split of the cost (30 days): keywords 33.7%; AI Max dynamic broad 25.8%; AI Max keywordless 8.0%; no breakdown 32.5%. Even in an account that is doing well, one euro in three falls into the opaque bucket.
The opaque bucket under examination. Pre/Post with the baseline as control: in aggregate, the "no breakdown" bucket improved (its CPA fell by 17% while the baseline worsened by 12% due to seasonality): an excess over the control of −29.5%. Verdict 🟢: what is happening to the bucket is seasonality, not AI Max.
But —and this is what makes measuring worthwhile even in a healthy account— the same analysis flagged two campaigns where the bucket did go off the rails above the control: in one, the "no breakdown" CPA worsened 65 points more than the baseline; in another, its CTR fell 85 points more while its share of the spend grew by 92%. There the recommendation was explicit: investigate whether AI Max is masking low-quality traffic in those two campaigns.
The constant and the variable
| The constant · the opacity (in both) | The variable · the result (different) |
|---|---|
| A third of the spend in "no breakdown" | AI Max CPA vs baseline: 🔴 worse (A) / 🟢 the same (B) |
| The combinations never fully reconcile (1% in A, 36% in B) | Pre/Post: red in 21/21 (A) / stable (B) |
| AI Max gets switched on by default, in bulk | ROAS: degraded (A) / intact (B) |
| The keyword stops mapping to the traffic | Excess of the bucket vs control: present (A) / absent except in 2 campaigns (B) |
Google delivers the same blurred report for the account that is doing well and for the one that isn't. The difference is not in the report: it is in whether you measure it. Without instrumentation, A and B arrive as an indistinguishable "blended CPA". With the instrument, you know which is which — and even in the good one you know which two campaigns to look at.
Why this opacity won't fix itself
Google does offer controls and reports, and the privacy restrictions have a regulatory basis. There is no need to assume bad faith: the incentive is enough. Automatic, low-detail expansion increases the auction surface (more eligible queries, more bids, more Search spend), and Search is by an enormous margin Alphabet's main revenue line. A product that widens that surface while blurring traceability doesn't need a conspiracy to thrive: it is enough that the advertiser cannot prove where the marginal money went. And what cannot be proved cannot be negotiated. Nobody is going to audit your spend for you.
💡 Ninja trick: the full study, with tables, figures and the detailed methodology, is in the panel's Library («The black box of AI Max», free with your account). And the instrument that produced those figures is AI Max Analyzer, also free: the same measurement on your own account, every night.
What you should remember
- Case A: bulk activation, 1% traceable, 605 exclusions in a month, a red Pre/Post in 21/21.
- Case B: CPA and ROAS identical to the baseline, a green verdict… and 32.5% opaque, with two campaigns off the rails that only the measurement uncovered.
- The opacity is constant; the result is variable. The report is the same; the difference is measuring.
- The opacity won't fix itself: the incentive is on the other side.