None of the three policies that follow is, on its own, a scandal. Each has a justification (product, privacy, reporting architecture). But stacked on top of one another they produce a system in which the advertiser pays for traffic that, by design, cannot be traced back to the query that caused it. This lesson describes the mechanism.
Layer 1 · The keyword stops mapping to the traffic
In a classic campaign, every search term hangs off a keyword and a match type: the search terms report is a complete map of why you are competing. In AI Max, dynamic broad matching and the keywordless engine generate impressions that hang off none of your keywords. The search terms report still exists, but it now shows a sample of what the model decided, not the inventory of why you are competing. The query → keyword map disappears.
Layer 2 · The privacy threshold trims what you can see
Since September 2020, Google does not show queries in the search terms report that don't reach a "significant" volume of people, for privacy reasons. In practice, a fraction of the spend spreads across a long tail of terms that never appear in any report. In a curated keyword campaign, the impact is limited. In AI Max, whose engine lives precisely off that long tail of new, low-volume queries, the trimming bites exactly where the incremental traffic you were sold lives.
Layer 3 · The reports don't reconcile with each other
And here is the knot. Google does not expose AI Max in a single coherent table, but spread across different views —search terms, asset/ad combinations, campaign— and each view applies its own privacy threshold. The result: a combination that spent money and appears in one view can return "zero conversions" in another, not because it didn't convert, but because that second view's threshold censors it. The numbers exist; they simply cannot be cross-referenced. Attribution becomes, literally, impossible to reconstruct from outside.
The residue has a name: «no breakdown»
When you try to split a campaign's spend into its sources —classic keywords, AI Max dynamic broad, AI Max keywordless— whatever fails to match any of them because of the thresholds falls into a "no breakdown" bucket. It is not marginal: in the study's healthy account it was 32.5% of the spend (one euro in three), with its own CPA and its own value, but with no source. Google doesn't say whether that cost comes from your keywords, from AI Max broad or from keywordless.
Why this matters more than in PMax
PMax is opaque from the outset: nobody expects to see its detail. AI Max is opaque inside a Search campaign that used to be transparent: the same search terms report you have been reading for years is still there, looking exactly the same, and it no longer means the same thing. It is an opacity that gives no warning — and that, on top of that, affects the campaign where you had the most control.
What it is not
- It is not a bug: it is the sum of three features.
- It is not necessarily bad faith: each layer has its reason.
- It does not imply that AI Max performs badly: the opacity neither rewards nor punishes performance; it is simply there, in the good account and in the bad one.
Hence the only operational conclusion: the observability of AI Max is the advertiser's problem, and as such it has to be instrumented from outside — not to switch it off as a matter of course (sometimes it contributes), but so as not to let it run blind. The next lesson is the instrument.
💡 Ninja trick: one practical consequence of layer 3 worth taking to heart: never subtract conversions between two AI Max views (terms vs combinations, campaign vs terms). They don't add up by construction, and the difference is not "lost conversions", it is threshold censorship. AI Max Analyzer works with each view inside its own limits and explicitly labels what is "no breakdown" instead of inventing an attribution.
What you should remember
- Three layers: no anchoring keyword, a privacy threshold that trims the long tail, views that don't reconcile with each other.
- The residue is the «no breakdown» bucket: up to a third of the spend with no source.
- The opacity gives no warning and affects the most transparent campaign you had.
- It is not a bug and it doesn't imply bad performance: you have to instrument it from outside.