AI Max is not new. It has been an opt-in setting for about a year. What changes on 1 September 2026 is that Google switches it on for you, by default, on two types of Search campaign, its broadest automated matching. Google emailed advertisers on 5 August. The wider net will scale whatever your conversion signals reward, and if you are counting the wrong thing it will buy more of the wrong thing, faster, while the report looks fine and the money leaks.
The standard advice, harden your conversions before automation scales, is only half right. A soft signal buys cheap traffic beautifully, so hardening it matters. But harden to a sale that fires only a few times a month and you starve Smart Bidding of the volume it needs to learn. In the accounts we have taken over, hardening the signal has more often exposed reported gains as noise than confirmed them. Choose the signal for both truth and volume, then prove the uplift with a holdout before you trust it.
The upgrade hits two campaign types, not your whole account
Only two groups of Search campaign are affected: those using automatically created assets, and those using the campaign-level broad match setting. Campaigns with automatically created assets get search term matching and text customisation switched on by default; broad match campaigns get search term matching. Search term matching lets a campaign spend against searches beyond your chosen keywords. Text customisation writes fresh headlines and descriptions from your ads and landing pages. Both Search Engine Land and Search Engine Roundtable covered the email. You keep control: AI Max can be switched off in campaign settings, individual features opted out of, and your brand inclusions, brand exclusions and negative keywords stay exactly as they were.
Treat Google's 14% uplift as a claim to test, not a number to bank
Google reports that advertisers who turn AI Max on see an average of 14% more conversions or conversion value at a similar cost per acquisition or return on ad spend, rising further for accounts previously reliant on exact and phrase match. That average and a disappointing result can come from the same account, because the conversions column reflects only what you told Google to count. To tell which one is yours, run a holdout: a geo split or a drafts-and-experiments test. Judge it on cost per genuine sale, not on the conversions column, sanity-checked against the search terms report, which now shows the headlines AI Max served and the pages it sent clicks to. If reported conversions climb while qualified pipeline stays flat, your uplift is a measurement artefact, not incremental demand.
Count the wrong thing and the automation turns against you
A soft conversion signal turns the automation against you, because it optimises impeccably towards the wrong target. One account we inherited counted a page-view-style signal as a conversion rather than a qualified enquiry. On paper it worked. Then broad matching widened the searches it bid on. Recorded conversions climbed steeply while genuine enquiries stayed flat, so the true cost of a real enquiry climbed sharply even as the reported cost per conversion held steady. That signal was trivially easy to trip, so the system bought more of the cheap, loosely related traffic that tripped it. Once we set the primary conversion action to a genuine enquiry, the spend reconciled to real business.
Hardening a rare signal can starve the algorithm, so choose it for volume too
Hardening to a signal that is real but too rare is the opposite mistake, and it is the objection most tracking advice ignores. Smart Bidding learns best with a steady flow of conversions. If your true sale fires only a handful of times a month, optimising to it directly leaves the algorithm guessing. The answer is not to fall back on a page view. Count the earliest action that is genuinely qualifying and clears the volume, a booked call or a vetted enquiry, then feed the eventual sale value back through offline conversion import so bidding still points at revenue.
On demotion, be precise. Secondary conversion actions are excluded from bidding by default, but influence leaks. Confirm that a soft action is not still marked Primary at account level, and check that no value rule or shared portfolio strategy is quietly reintroducing it. A proxy demoted in one campaign and left Primary in the goal set still drives your bids.
The inflated number reaches your Q4 budget committee in October
The real danger is not the September report but what a budget committee does with it. September is AI Max's first full month under the wider settings, and those numbers get read around 1 October, exactly as Q4 plans are set. A soft signal shows a conversion count that looks healthy, even up, while genuine pipeline has not followed. A committee reading that column pours more budget into the channel and scales the artefact, so the gap between reported conversions and real revenue widens with every extra pound. Reconcile the September number to actual sales before anyone allocates against it.
Harden the signal in an afternoon before the switch
Once your Primary conversion action is a qualified enquiry or sale, four operational steps make the wider net pay:
- Attach a value. Give conversions a value where you can, so the automation optimises towards revenue, not volume.
- Feed back what lands later. Where the true value of a sale arrives after the click, return it through offline conversion import.
- Strengthen the lead signal. Turn on Enhanced conversions for leads so the signal you send is accurate and complete.
- Set the guardrails. Use AI Max's brand controls, location settings and final URL expansion, and keep sensible negatives.
None of this delays AI Max. It is how the day it turns on widens the net towards real customers rather than noise. The fix costs an afternoon and it is available now, before the switch. Choose the signal for truth and volume, prove the uplift with a holdout, then let the automation work.
