AI Max Will Buy the Leads You Count, Not the Leads You Want

Say you spend $20k a month on Search and tell AI Max to get you form fills. Three weeks later, the dashboard shows 184 conversions and a tidy CPA. Your techs are driving to empty houses and calling spam.

I used to tell clients in that position they needed better keywords. I was wrong. AI Max and Smart Bidding do not understand your business. They predict the probability and value of the conversion action you told them counts, then spend toward it. If every raw form submission counts as success, the system will faithfully buy more raw submissions. Junk included.

That is the one fact the rest of this setup follows from. The Leads objective is a sensible starting point when you want prospects through forms. But the objective does not inspect each lead and decide whether you would pay to get another one. The conversion goal underneath it supplies that instruction.

The objective starts the campaign; the goal steers it

Think of the campaign objective as the preset Google loads during setup. Choose Leads and Google preselects account-default lead conversions, suggests Maximize Conversions or Maximize Conversion Value, and points the interface toward lead volume. That preset helps you get started. After setup, the conversions active for the campaign tell Smart Bidding what to pursue.

I learned the distinction on a home services account that had picked Leads and still bought garbage. It counted a quote-request submission and a newsletter signup as primary actions. The system had no reason to treat one as a sales opportunity and the other as someone who wanted occasional emails. As one agency setup guide puts it, tell Google a newsletter signup is success and it will bring you thousands and consider the job done.

Keep a garage door company in mind for the rest of this piece. It spends $20k a month, and the outcome it wants is a qualified quote request. If its goal counts every form equally, reported CPA can fall while booked jobs fall with it. The objective points the car. The conversion goal is the address you typed in. The analogy stops there: the system is not reading a map of your business. It is predicting which auctions are likely to produce the event you labeled.

Primary and secondary are not two names for the same signal

A conversion action marked primary feeds the Conversions column and is used for bidding. A secondary action remains visible under All conversions for observation. Google’s help describes primary actions as used for bidding and secondary actions as observation-only. That distinction is a bidding decision, not a preference for how tidy your report looks.

For the garage door company, a newsletter signup can still be worth watching. So can a raw quote form, especially while the team works out which submissions are worth a callback. Neither has to be a primary bidding signal merely because someone wants it on a dashboard. Keep the action closest to a real customer primary; observe the rest until it earns a place in bidding.

Every auction takes its cue from what you counted

At an auction, Smart Bidding predicts how likely a click is to produce your counted conversion and sets a bid accordingly. A stronger predicted chance can justify a higher bid; a weaker one gets less. It learns from the clicks that previously carried the conversion label. Nothing in that label, by itself, says whether the garage door company serves the caller’s address or whether anyone answered when the office called back.

Leave spam in the training data and the model can learn to favor the pattern that produces it. That is how a late-night, one-word inquiry from outside the service area stops being merely annoying. If it counts as a conversion, it becomes evidence for the next auction.

AI Max gives that bidding signal more room to work. With conversion-based Smart Bidding, it can reach beyond your keywords through broad match and keywordless matching, customize headlines from your pages and assets, and send a click to a different URL it considers more relevant. A testing breakdown separates those capabilities into additional searches, text customization, and final URL expansion.

Expansion still takes its cues from what you provide. AI Max draws on landing pages, ad assets, and ad-group keywords as it looks for coverage. A generic quote page gives it less to distinguish a useful garage door inquiry from generic form-filling traffic. More reach does not repair a bad definition of success. It gives that definition more places to spend.

Fix the goal and the page before you blame the match type.

Conversion count and conversion value give different instructions

Under Maximize Conversions, each primary conversion adds one to the count. A quote request and a newsletter signup are both one conversion if you make both primary. Under Maximize Conversion Value, the system pursues the values assigned to the primary actions instead. That lets the value signal distinguish outcomes the count treats alike.

For the garage door company, a $4,000 replacement inquiry and an $89 repair inquiry are not commercially identical. Yet a goal that simply counts both as quote requests gives bidding no difference to pursue. That is how 40 cheap repair leads can crowd out four replacement jobs that would have paid for the month. If the business cannot pass real revenue yet, proxy values based on what those lead types tend to be worth can provide a better instruction than equal counts.

There is an important limit: a value assigned to a secondary raw submission does not make that submission a bidding target. You can track its $20 proxy value for observation while a qualified quote, valued at $200, stays primary. If you intend the system to bid toward different values, the valued actions must be included in the bidding goal. Do not move raw fills to secondary and expect their assigned value to steer spend. Count, value, and primary status each do a different job.

Bad lead data becomes a bidding problem

When every raw submission is primary, bidding learns to find more submitters. It cannot infer from a thank-you page that sales later rejected the lead. A lead-fraud breakdown recommends making the qualified-lead goal primary, feeding qualified outcomes through Enhanced Conversions for Leads, and keeping raw fills secondary.

I have seen accounts where 30% of reported conversions were bot fills from the same two spam phrases. The account was not broken. The counting was. Moving those fills out of the primary goal changes the instruction from “find more of these forms” to “find more of the leads we actually accept.” It does not undo the wasted spend, but it stops presenting that waste as a success the system should repeat.

The quieter version shows up in B2B SaaS. Say an account makes demo requests and gated ebook downloads primary. The ebook converts at 11%; the demo converts at 1.4%. Within ten days, Smart Bidding has a lesson it can act on: ebook seekers are easier to find. Budget drifts toward queries for templates and examples. Demo volume flatlines while reported conversions climb. The dashboard calls it growth.

I would keep the demo request primary and move the ebook, newsletter, and pricing PDF to secondary for observation. If the business needs softer actions in its bidding goal, their values have to reflect their different worth rather than treating every download like a sales conversation. A demo that closes 18% of the time at $9,000 ACV should not receive the same instruction as an ebook that almost never closes. This account of a B2B team moving from demo requests toward pipeline shows why the definition of success matters beyond the form.

If two actions do not deserve the same bid, do not put them in an equal-count primary bucket.

Clean signals still need time to accumulate

A better goal is not an instant fix for a small account. Google describes learning in two layers: the visible Learning label clears in at most seven days, while calibration behind it may take roughly 50 conversion events or about three conversion cycles. That 50 is an upper benchmark, not a requirement to switch the system on.

A garage door account getting 12 qualified quotes a month will take calendar time to build that signal. AI Max has no minimum conversion requirement to turn on. It can run on a thin trickle of qualified leads; it just has less evidence from which to learn. Splitting the same small budget across several near-identical campaigns makes that problem harder to manage.

Changes to the goal matter, too. Creating a new strategy or significantly changing which conversions feed it can put bidding back into Learning. Budget, target, creative, and targeting edits do not wipe its history or retrigger that status under the October 2026 help rewrite. I used to pause and rebuild small campaigns every two weeks when results looked soft. That habit gave the system a moving target just when it needed a consistent one.

For the garage door company, I would start with one primary qualified-lead goal, a geography it can actually serve, and a structure that lets that signal accumulate. Consolidation is useful when it gives the same clean outcome more chances to teach bidding. It is not a substitute for defining that outcome correctly.

Set the instruction before you adjust the bids

For a form-led AI Max campaign, I would make these decisions in order:

1. Choose the primary money action. Use a qualified quote, qualified demo, or booked job. Keep raw submits and softer actions secondary while you assess them.
2. Decide whether values should steer bidding. With Maximize Conversion Value, assign values to the actions in the bidding goal according to their worth. Use proxies until offline outcomes can replace them; do not mistake a secondary action’s recorded value for a bid instruction.
3. Keep junk out of the primary signal. Review spam, out-of-area inquiries, and short calls rather than treating every recorded interaction as a lead worth repeating.
4. Give the clean goal time. Hold a consistent account-default goal and avoid splitting a small signal across near-identical campaigns. A budget or creative edit is not the same as swapping the outcome bidding learns from.

In testing across 250-plus campaigns, AI Max delivered a median 13% more revenue at 16% higher CPA, with ROAS ranging from plus 42% to minus 35% when guardrails were loose. That spread is a useful warning against judging expansion by conversion count alone. At groas, the goal is the first check before a bid or budget change. Continuous execution against a sloppy goal just buys junk faster.

Put the model to work on a harder account

Now take a local group with 12 locations. Each has its own small campaign, $25 a day, and a quote form. None collects qualified quotes quickly. Four locations share a generic contact page, so even a relevant click can land on a page that does little to identify the right shop.

Start with the signal, not the campaign count. If every submitted form is primary, each location teaches bidding to find form submitters, including people the business cannot serve. Marking a qualified lead as primary fixes the instruction, but each tiny campaign still has few examples from which to learn. Consolidating around one account-default qualified-lead goal gives that signal room to accumulate. Specific, current location pages then give AI Max clearer material to use when it matches location-intent searches and considers a final URL.

That does not mean final URL expansion knows which shop should take every call. The pages and the conversion data still have to say what the business means. Twelve separate campaigns do not automatically give you control; they can give you twelve slow learners. Pick Leads, then prove you meant it by counting the forms you would pay to receive again.