You don't have to give up AI Max to stop being blind. This lesson brings together the study's principles —in order of return— and the controls Google does offer, and it ends with the decision by scenario.
The four principles
1. Measure by source, not by campaign
The number that matters is not the campaign's CPA, but the CPA of the keyword baseline kept separate from AI Max's. If you don't have them separated, you are not measuring AI Max: you are measuring a blend.
2. Set the baseline before switching it on
The only honest comparison is against the performance of your own account before activation. Save that PRE period (per campaign: cost, conversions, CPA, CTR, the weight of the "no breakdown" bucket); without it, any "it's going well" is an impression. If AI Max was already on when you arrived, reconstruct the PRE from the detected activation date.
3. Watch whether the «no breakdown» bucket grows after activation
It is the most valuable signal and the one nobody looks at. The opaque bucket, on its own, does not accuse AI Max of anything. But if after switching it on it grows in weight or its CPA worsens far more than your baseline, that excess over the control is attributable: AI Max is hiding bad traffic in the blind spot. Look at it before/after, not in absolute terms.
4. Watch the drift, not just the launch
AI Max evolves: what is profitable incremental traffic today can turn into unqualified traffic in weeks (the model explores, the competition changes, the long tail moves). The useful control is continuous (alerts), not a one-off snapshot.
And a fifth, operational one: treat the long tail as permanent hygiene — systematic exclusions of unqualified terms, precisely in the band that is hardest to see.
The controls Google does give you
| Control | What it allows |
|---|---|
| Components separately | Switch on matching only, or text customization only, or final URL expansion only; turn the rest off |
| Brand exclusions (or inclusions) | Stop keywordless traffic entering searches for brands you don't want (or restrict it to the ones you do) |
| URL exclusions | Pages URL expansion must not send traffic to |
| Negatives | They work on AI Max; account-level ones too |
| Locations of interest (in some cases) | Restrict keywordless traffic by area |
| Reports by source | Terms, generated headlines and URLs by campaign |
| AI Max experiments | Compare the campaign with and without AI Max with the traffic split (the clean incrementality test) |
| Switching it off | Per campaign, at any time |
The AI Max experiment
Google offers specific experiments: a copy of the campaign with AI Max against the original without it, 50/50, several weeks, with significance. It is the official way of answering "does it contribute?" — with two caveats: it measures the whole campaign (not the opaque bucket separately), and a positive result today does not guarantee the one in three months' time (drift). Use it to decide on activation; use the module's method for monitoring afterwards.
Decision by scenario
| Scenario | Decision |
|---|---|
| Visible sources 🟢, Pre/Post bucket 🟢, acceptable observability | Keep AI Max; continuous monitoring; long-tail hygiene |
| Visible sources 🟢, bucket 🔴 in some campaigns | Rein it in: in those campaigns, switch off the keywordless component (or AI Max entirely), negatives, URL exclusions; keep it in the rest |
| Visible sources 🔴, bucket 🔴 | Switch off AI Max in that campaign; if you want volume, try dynamic broad matching only with strong negatives; or run a formal experiment |
| Very low observability (< 10%) with any result | Treat the verdict as provisional; lean on Pre/Post and on the account total; consider reining it in as a precaution |
| You don't know when it was switched on | The date first (change history, first AI Max activity); with no PRE there is no verdict |
What not to do
- Switching AI Max off "as a matter of course" without measuring: sometimes it genuinely contributes (case B).
- Leaving it on "because Google recommends it" without measuring: sometimes it degrades things quietly (case A).
- Judging it by the campaign's CPA.
- Looking at it once and forgetting about it.
💡 Ninja trick: the scenario decision can be taken every week if the table exists every night. AI Max Analyzer produces that table — per campaign, with a verdict and a recommendation — and emails you when a campaign changes colour. It is free in the panel because measurement, unlike the rest of the Suite, is not a luxury: it is the condition for AI Max being your decision and not Google's.
What you should remember
- Four principles: by source, baseline first, watch the bucket after activation, watch the drift (+ long-tail hygiene).
- Controls: components separately, brands, URLs, negatives, experiments, switching it off.
- Decide by scenario (visible sources × bucket × observability), campaign by campaign.
- Neither switch it off as a matter of course nor leave it on out of faith: measure.