AI Max, Performance Max and Advantage+ on a small budget: what to check before you let the machine drive

Background

The pitch for every AI feature in an ad platform is the same: give it your budget and your goal, and it'll find customers you'd never have found with keywords. On a big account with thousands of conversions a month, that's often true. On a small UK budget, a few hundred pounds a month and a dozen enquiries, the machine has very little to learn from, and every pound it spends finding out costs you.

Both platforms have moved the defaults, too. In Google Ads, AI Max sits on top of an ordinary Search campaign and switches on three things: search term matching (it goes beyond your keywords), text customisation (it writes ad copy from your site and existing ads) and final URL expansion (it chooses which page to send people to). Google is also moving Dynamic Search Ads campaigns over to AI Max. On Meta, campaigns for the leads objective are Advantage+ by default, and Advantage+ creative can adjust your images and text, with some enhancements on by default.

None of that is bad. It just means the defaults decide a lot unless you look.

My bias, for what it's worth: my early jobs included web marketing and search visibility, and I'd still rather see what's being bought before I hand over the steering wheel.

Setup

1. Fix the tracking before anything else

Smart bidding, AI Max and Advantage+ all learn from conversions. If your conversions are wrong, the machine optimises for the wrong thing very efficiently. Google's own setup guidance says AI Max's features need conversion-based Smart Bidding to work well, which is the same point from the other side.

Before any AI feature goes on, check:

• What counts as a conversion. A form submit, a call over a minute, a booking. Not a page view, and not "clicked the phone number" if most of those never ring.
• That each one fires once. Test it yourself and look in the platform.
• That consent is handled. Tracking pixels and remarketing need consent under PECR, and so does adding people to a mailing list (see mailing list consent).

Decide how you'll measure it first covers the planning side.

2. Google: switch AI Max on deliberately, one feature at a time

AI Max lets you control its parts separately, so use that.

• Search term matching can be turned off per ad group. On a small budget I'd start with it on in one ad group, with the search terms report open, rather than across the account.
• Final URL expansion can send people to pages you didn't pick. Add URL exclusions at campaign level for anything you'd never want as a landing page: careers, privacy notice, blog posts that don't sell anything.
• Text customisation writes copy for you. Read what it produces, because you're responsible for it (more on that below).
• Check your account's notifications for when any Dynamic Search Ads campaigns move to AI Max, rather than relying on a date you read somewhere.

3. Google: keep Performance Max on a lead

Performance Max spends across Search, Shopping, YouTube, Display and the rest, and shows you less. The controls that matter most:

• Brand exclusions so it doesn't spend on people already searching for you by name.
• Negative keywords at campaign or account level. They only cover Search and Shopping inventory; Display and video use the content suitability settings instead.
• The search terms report for Performance Max, which now exists. Read it weekly and add negatives.

4. Meta: check every toggle Advantage+ turned on

• Creative enhancements. Meta says some may be on by default and that you can turn them off at any time. Look at the previews of every format before launch. A cropped photo of a clinic or a filter on a product shot can change what the ad says.
• Placements. Advantage+ placements spread spend across every surface. If that's wrong for you, choose manual placements and accept that you're overriding the default.
• Audience. On a leads campaign, narrowing the settings too much switches Advantage+ off. Know which state you're in.

5. Write variations with AI, then check them as the advertiser

Using a model to write ad variations is fine. Publishing them unread isn't. This prompt keeps the claims tied to what you can prove:

You're writing Google responsive search ad assets for [BUSINESS], a
[TYPE OF BUSINESS] in [PLACE], UK.

Facts you may use (and nothing else):
- [SERVICE AND WHAT IT INCLUDES]
- [PRICE OR "FROM" PRICE, IF PUBLISHED]
- [PROOF POINTS WE CAN EVIDENCE, e.g. qualifications, years trading]

Write 10 headlines (max 30 characters) and 4 descriptions (max 90
characters), in UK English.
Rules:
- No claim that isn't in the facts list. No "best", "No.1", "guaranteed",
  "cheapest" or invented statistics.
- No urgency we can't honour ("today only") unless it's true.
- After the list, mark any line that relies on a fact I should double-check.

And a second one for Meta, where the image matters as much as the words:

Suggest 3 short primary texts (max 125 characters) for a Meta ad for
[OFFER], aimed at [AUDIENCE] in [PLACE].
Use only these facts: [FACTS].
Then list anything about the image I'm planning ([DESCRIBE IMAGE]) that
could make the ad misleading or imply something the text doesn't say.

Then a person reads every line against the facts list before it goes live.

6. Bring your own images, so the platform isn't inventing them

Both platforms will happily generate or adjust images for you. If you'd rather control what the ad shows, supply finished, on-brand assets and switch off the enhancements you don't want. (For property listings, my own Marketool, in beta, makes those images.)