How to Set Up Google AI Max for Maximum Results

1.Conversion tracking — the most undervalued lever

Most accounts get this wrong by counting too shallow. They wire up "form submitted" or "thank-you page viewed" as the primary conversion, feed that to Smart Bidding, and wonder why the leads are awful. AI Max — like every modern bidding system — buys you exactly what you ask for. Ask for cheap form-fills, get cheap form-fills.

The fix is to optimise toward the deepest stage your stack can reliably report. Not the deepest stage that exists in your funnel — the deepest one where the data lands in Google Ads cleanly, on time, every time.

If you're lead-gen

• Identify your funnel stages: lead → qualified lead → booked meeting → sales-accepted lead → closed-won revenue.
• Pick the stage where (a) volume is high enough to feed Smart Bidding (broadly: 30+ conversions per month per campaign) and (b) the signal is unambiguous.
• Connect it via offline conversion imports — feed back CRM or sales-system data using GCLID or enhanced conversions.
• Set conversion values, not just counts. A booked meeting from an enterprise lead is worth more than one from a student.

If you're e-commerce

• Send revenue per transaction, not just transaction count.
• Differentiate ROAS targets by margin tier or product category. Catalogue-wide ROAS averages a high-margin and low-margin product into the same number — and Smart Bidding optimises for the average.
• Use enhanced conversions for purchases to recover lost match rate from cookie loss.
• For high-LTV businesses (subscription, repeat-purchase), feed LTV-adjusted values, not first-purchase value.

2.AI Brief — write the brief the agency would write

AI Brief is Google's most useful new control. It's the closest the platform has come to letting you give Gemini a creative brief instead of a keyword list. Three guideline categories: messaging, matching, audience. Most accounts leave it blank — which is the same as telling the agency "do whatever you think."

What goes in each category:

Messaging

• Brand voice descriptors — tone (warm, technical, irreverent), reading level, sentence structure preferences.
• Words to favour — the language you actually use about your category.
• Words to avoid — corporate filler, banned superlatives, anything off-tone ("game-changing", "revolutionary", "best-in-class" — the usual suspects).
• Claims you can defensibly make — quantified outcomes, named methodologies, geographic specificity.
• Claims you cannot make — anything legal, compliance, or factual won't let you say. Especially important in healthcare, finance, education, insurance, legal.

Matching

• Topics to target — the commercial intent clusters you actually want.
• Topics to exclude — adjacent searches that look relevant but aren't (job seekers, students, hobbyists, competitor employees researching).
• Geographic constraints — if your product is country-specific or city-specific, state it explicitly.

Audience

• Who you want to reach — job titles, life stages, intent signals, customer profile.
• Who you don't want to reach — competitors, employees, existing customers (where retargeting handles them separately), low-fit segments.

3.Negative signals — define what you DON'T want first

AI Max's expansion is broad by design. The only way to keep it controlled is explicit exclusions. Build the negative system before flipping toggles, not after the budget bleeds.

What to define up front:

• Negative keyword list at account level and campaign level. Mine your last 90 days of search-term reports for obvious waste — job queries, free-tier seekers, students, competitor employees, geographies you don't serve.
• Audience exclusions — your own customers (if retargeting handles them separately), existing leads in nurture, competitor employees if Customer Match is set up.
• Geo exclusions — anywhere you don't ship, don't serve, or can't legally operate.
• Device exclusions — rare but sometimes needed (e.g., B2B SaaS where mobile converts at zero).
• Brand-bidding rules — who else is allowed to bid on your brand terms, and your defensive response.

4.Brand controls and text restrictions

With text customisation on, AI Max generates headlines and descriptions from your assets, pages, and AI Brief. Brand controls and text restrictions set the boundaries on what it can output.

What to set:

• Text restrictions in the account — banned words, required phrases, claim governance. Write them down in the platform, not just in policy docs.
• For regulated industries — map every disallowed claim to an explicit restriction. Healthcare: no outcome guarantees. Finance: no guaranteed return language. Education: only accredited claims. Insurance: only approved benefit language.
• For trademark and legal — protect product and brand names with required exact phrasing. Stop Gemini reformatting "iPhone 17 Pro Max" into "iPhone 17 ProMax" or worse.
• For tone — banned filler phrases, banned superlatives, banned competitor name-checks if your legal team won't allow them.
• Which campaigns get AI Brief vs which leave the defaults. Not every campaign needs the full brief — but the ones with the most variable copy generation do.

5.Smart Bidding — value, not volume

AI Max bids using Smart Bidding. The bid signal is downstream of conversion tracking. If conversion equals form-fill, Smart Bidding chases cheap form-fills. If conversion equals qualified-lead value or revenue, Smart Bidding chases higher-value users. Garbage in, garbage out is the rule.

The right bid strategy depends on what you actually want:

• Max Conversions — useful only when conversion = the real outcome (qualified lead, sale, booked meeting). Useless if conversion = vanity event.
• Max Conversion Value — better. Optimises toward the value you assign, not just count. Requires conversion values to be set accurately.
• Target CPA — works if your CPA target reflects the true cost of acquiring a customer at the deeper stage, not the cheap stage.
• Target ROAS — best for e-commerce, broken down by margin tier or category. A catalogue-wide ROAS target averages products that need different treatment.

For seasonality (Black Friday, end-of-financial-year, planned promotions), use Smart Bidding's seasonality adjustments rather than panic-overriding bids. For high-LTV businesses, feed LTV-adjusted values — first-purchase value undersells repeat-purchase users.

6.Campaign structure — split by economics, not syntax

The architecture point in one line: don't split campaigns by keyword match-type or syntactic cluster. Split them by the business logic that actually requires separation — different budgets, different ROAS targets, different conversion goals, different page sets, different compliance rules.

Brand stays separate because brand demand is already qualified. Competitor stays separate because copy, legal risk, and page strategy differ. Problem-led discovery isn't judged on the same CPA as bottom-funnel intent. Generic high-intent gets the most AI Max freedom. Location-led gets separate budgets where geo economics matter.

The full structure-by-layer breakdown is in the explainer. The setup point here: decide your structure before you turn AI Max on. Restructuring an already-learning AI Max account is more expensive than getting the structure right at launch.

7.Experiment design — use Google's test tool properly

Google built an experiment tool for AI Max specifically. Use it. Don't flip account-wide on day one. The experiment splits traffic 50/50 between your current setup and the AI Max version, runs for a defined period, and reports on conversion lift.

The trap is reading the result through the wrong lens.

What to do:

• Set up an AI Max experiment at the campaign level — not the account level for the first run.
• Run for at least 2-4 weeks, depending on conversion volume. Statistical significance needs the conversions to land.
• Hold every other variable constant during the test — no major budget changes, no new ad copy, no new landing pages.
• Stage experiments — run on generic search first, then category, then problem-led demand. Brand campaigns last, if ever.
• Write the test plan and stopping rules before launch. "We'll roll out if X happens, hold if Y, abort if Z."
• Measure on the same conversion event Smart Bidding is optimising toward. Cross-reference with downstream quality.