AI Max is Google's extension for search campaigns: the campaign may serve beyond your booked keywords, a little like Performance Max but confined to search. The obvious question is where to try it safely. A brand campaign looks like the safe place — closed search volume, high conversion rate, limited downside. That is exactly where we tested it in spring 2026. The result was negative, and the reason is more useful than the numbers.
The setup
The client is a German manufacturer of heating technology. The brand campaign runs on exact brand terms with a target CPA strategy and a fixed daily budget that was fully spent throughout. We enabled AI Max search term matching only. Automatic text customisation and final URL expansion stayed off, so AI Max wrote no ad copy and chose no landing pages of its own. A brand list was added four days after launch, and the first negative keywords a day later.
That constraint matters for how you read this: what we measured is the most cautious possible version of AI Max. Let it write text and pick landing pages too, and you are measuring something else.
The method: why most AI Max tests prove nothing
The most common mistake is comparing the weeks before and after activation with the learning period included. Then you are mostly measuring how long the algorithm took. The second most common is changing budget or target CPA at the same time. After that, nothing can be attributed to anything.
- Before: six full weeks prior to the switch.
- Switch week and learning period: roughly two weeks, excluded entirely.
- After: the three weeks that followed.
- Held constant: budget, target CPA and keyword structure across both windows.
A year-on-year comparison was impossible: the campaign had a different budget, bidding strategy and structure the year before. We deliberately avoided an A/B test with a duplicated campaign, because two campaigns bidding on the same brand terms cannibalise each other in the auction. What remains is a clean pre/post comparison with every lever frozen.
The results
All figures are weekly averages, after versus before. Spend stayed flat because the budget was fully used in both windows — the campaign spent the same money differently.
| Metric | Change |
|---|---|
| Cost | −2% |
| Impressions | +37% |
| Eligible auctions | +80% |
| Clicks | −13% |
| CTR | −11.6 pts |
| Conversions | −9% |
| Cost per conversion | +7% |
| Lost impression share (rank) | 10.7% → 44.8% |
The line that matters is the last one. The share of auctions lost to ad rank quadrupled. The campaign suddenly qualified for far more auctions (+80%), yet in a growing share of them it was no longer bidding competitively.
What AI Max did well
- It stayed on brand. 98% of AI Max spend went to queries containing the brand name. Almost nothing reached competitor terms. The brand list works.
- It found model queries. AI Max surfaced searches for specific product names that were not booked as keywords. That traffic converted well and was later added as keywords — the lasting gain from the test.
- More presence on exact brand searches. The share of exactly matching queries rose from 69% to 82%.
What AI Max cost
The new traffic was more expensive: 1.4 to 1.9 times the cost per conversion of the exact brand terms. On its own that would be tolerable. The real problem lies elsewhere.
We grouped the visible AI Max queries by spend. The picture is unambiguous:
| Type of query | Share of AI Max spend | Return |
|---|---|---|
| Brand + product or model | 44% | converts well |
| Service, contact, dealer search, locations | 26% | almost no conversions |
| Product research (reviews, comparisons) | 11% | converts well |
| Other product categories of the manufacturer | 8% | no conversions |
| Brand name only, miscellaneous | 8% | almost no conversions |
| Model name without brand | 2% | no conversions |
| Competitors and non-brand queries | 0.4% | no conversions |
45% of the AI Max budget went to queries with almost no conversions — people looking for customer service, a local dealer, or an entirely different product from the same manufacturer. All of them landed on the same page, a quote form for the main product. For a service enquiry that is the wrong page, and no bidding strategy repairs that.
There is also a loss of transparency that rarely makes it into reports: around 60% of AI Max traffic appears in the search terms report only as “other search terms”. The visible queries are a sample — we analysed the smaller half and extrapolated to the larger one.
The real effect: a side effect on the brand term
The campaign's core query — the plain brand name — moved like this:
| Core brand query | Change |
|---|---|
| Impressions | +18% |
| Clicks | −21% |
| CTR | −11.2 pts |
| Conversions | −29% |
That single query accounts for the entire conversion loss of the campaign. The obvious explanations can be ruled out, and each one is checkable:
- Demand was stable. Impressions divided by impression share gives the same number of searches in both windows.
- Ad quality was unchanged. Quality Score stayed at 10 out of 10 throughout.
- No new competitor appeared. The same five domains in auction insights, before and after.
When demand, quality and competition hold steady but your own position falls, the cause is inside the account. The most plausible mechanism: since AI Max, the target CPA strategy has been optimising across a mix that includes lower-converting additional traffic. To hold the target on average, it bids less on the single most valuable term. Competitors appear above the ad more often, CTR falls, and conversions follow.
To stay honest: that is the most plausible explanation, not proof. The reverse test — switching AI Max off and measuring three unchanged weeks — is the step that confirms or refutes it. We set the success criteria before running it, not after.
What we recommend
- Do not run AI Max in pure brand campaigns for now. With closed search volume and a high conversion rate there is little to discover and a lot to lose.
- Keep the model queries it found. That is the lasting value of the test: book searches for specific product names as their own keywords.
- Maintain negatives regardless of AI Max. Service, support, contact, dealer search, locations, careers and unrelated product categories belong on the negative list — phrase ad groups reach those queries too.
- Report the core term on its own. CTR, absolute top impression share and conversions for the most important brand term belong in the weekly report. At campaign level this effect would have stayed invisible for months.
- Before testing in generic campaigns: exclude the learning period, freeze budget and target CPA, leave text customisation and URL expansion off, exclude the brand in generic campaigns, and write the success criteria down first.
What this means for other accounts
One test, one account, one industry — that is not a law of nature. What transfers is the mechanism: as soon as an automated bidding strategy receives additional traffic with a different conversion rate, it redistributes bids. In campaigns with heterogeneous traffic that barely shows. In a brand campaign whose value hangs on a single term it shows immediately — you just will not see it if you only look at campaign-level metrics.
We run tests like this regularly, both in Google Ads management and across digital advertising. For companies in the energy sector we described our approach in the B2B SEO case for energy suppliers and on SEO for energy companies.