How to Test AI Max in Google Ads Without Guessing What Worked

Choose a campaign that can answer the question

A familiar campaign makes the result easier to interpret. Before testing, check whether its conversions mean what the team thinks they mean. A form submission and a sales-qualified lead are different outcomes; a purchase value and its margin are different numbers. Use the same definitions on both sides of the comparison.

Write a short hypothesis. For example: “AI Max could find more qualified installation enquiries without taking our cost per qualified enquiry above the amount the business can sustain.” This is an illustrative business question, not a promise about the feature. Decide what change would be large enough to matter, what downside you can afford, and how long sales normally needs to assess an enquiry.

Keep major changes to the offer, conversion setup and website out of the test where practical. If a promotion has to launch halfway through, record it and reconsider what the experiment can tell you. Good experiment notes explain the unusual events; they do not need to narrate every day.

Before enabling URL expansion, test your tracking template with a dynamic landing page. Google documents how incompatible templates can send visitors to a fixed address or a broken page. A measurement problem is worth fixing before it becomes an experiment result.

Use the current experiment setup

Google's native AI Max experiment divides traffic and budget inside one campaign: control has AI Max off, treatment has it on. Custom experiments remain a separate, copied-campaign option.

In Campaigns → Experiments, add Campaign features and settings → AI Max for Search campaigns. Check the treatment's asset settings. Brand inclusions or exclusions added during setup affect both arms. Initial setup does not support ad-group changes; those can be edited later. Check Google's eligibility list, including restrictions on shared budgets, portfolio bidding and campaigns already using text customization.

The setup also offers automatic application of favorable results. Leave that option off if you want the team to make the final decision. Results appear in Experiment summary. Ending without applying restores the original AI Max settings.

Do not dismantle a well-functioning budget or bidding arrangement just to make a campaign selectable. Another campaign may be a better test candidate. The experiment is supposed to help the business make a decision, not become a reason to reorganize the account.

Choose the feature combination deliberately. AI Max has separate matching and asset controls; final URL expansion depends on text customization. A test of the full combination cannot tell you the isolated effect of every component. If the question is specifically about landing-page selection, make that limitation explicit before interpreting the result.

Separate the experiment result from the explanation

Begin with the control-versus-treatment outcome. Then investigate how it happened. AI Max reporting includes a combined search-term, headline and landing-page view, keyword totals for expanded and landing-page matches, and reports on assets and selected URLs. Filtering search terms for AI Max omits “Other search terms,” so that view is not a complete traffic total.

Those reports help explain the result; they do not independently prove incrementality. A conversion attributed to AI Max is not automatically a conversion the business would otherwise have missed. That is the question the experiment comparison is designed to address.

Consider an illustrative test where treatment produces more enquiries and a lower cost per form submission. Sales then finds that the extra forms are mostly requests for a service the company does not provide. The campaign metric improved, but the original hypothesis did not. Equally, a broader-looking query can produce a valuable customer. Judge the outcome, not just whether the wording resembles your keyword list.

Allow conversions and any downstream qualification to mature before comparing results. Google recommends a two-week ramp period before routine AI Max optimization changes; that is not a universal experiment duration or a guarantee of statistical certainty. A small sample may still leave the answer unresolved. Broken destinations or false claims deserve immediate attention, not two weeks of tolerance.

A useful test can end without a winner

Adopt the tested settings when the evidence supports a worthwhile improvement and the resulting customer experience is acceptable. Check the experiment's uncertainty as well as its point estimate. A large-looking lift with little data is a reason to be cautious, not a reason to write a bigger announcement.

If the result is inconclusive, decide whether more time is likely to resolve it at a sensible cost. If not, keep the existing setup. “We did not establish an improvement” is a legitimate finding.

If a particular URL or copy problem spoiled an otherwise promising test, fix that problem and decide whether a new comparison is warranted. Do not quietly remove the bad period and present the remainder as the original experiment's result.

After applying a chosen change, verify the actual campaign settings and destination behavior immediately. Evaluate its later performance separately, once enough new outcomes have arrived. Confirmation that a setting saved and evidence that it improved results are different things.