AI in paid ads: what automation does and where you step in

What AI handles in advertising today

The big ad platforms have handed much of classic campaign work to their AI. What used to be set manually is now often decided by an algorithm trained on data from all advertisers.

The principle is the same everywhere: you set the goal, budget, creative and measurement; the platform decides the rest. The fewer signals you provide, the more the machine decides alone, and it optimises for whatever you report as success.

Meta Advantage+ in practice

• Check placements: automatic placements often include the Audience Network, ad slots in third-party apps. For leads on a landing page, check the placement breakdown regularly.
• Pick the optimisation goal deliberately: optimise for clicks and you get clicks. Optimise for leads or purchases and the campaign needs enough of those events to learn.
• Provide signals: set up the pixel and Conversions API properly, only after consent, so optimisation sees real events.
• Creative as the lever: when audience and placement run automatically, creative is the main lever left to you.
• Don’t bend things afterwards: optimisation goal and attribution window often can’t be changed on existing ad sets; create new ones instead.

Google: Performance Max and AI Max

Performance Max is a goal-based campaign type that uses all Google Ads inventory from one campaign. You set conversion goals and provide asset groups with text, images, logos and video; Google’s AI combines them into ads for the different channels. It’s available for objectives such as sales, leads or website traffic.

AI Max for Search brings similar automation to classic Search campaigns: search term matching is broadened via broad match and keywordless targeting, ad text is customised, and Google can send users to a different, better-fitting page on your site.

• Review search terms: more reach also means more irrelevant queries. Maintain negative keywords.
• Watch final URLs: if Google picks the landing page, make sure outdated or unsuitable pages can’t be chosen.
• Read vendor claims carefully: Google cites typical extra conversions for AI Max based on internal data. That’s an indication, not an independent measurement, and no promise for your account.

AI creative: making and testing it

“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.

AI makes ad creative cheap: image variants, copy, short videos, AI actors. That changes the way you work: fewer elaborate one-offs, more variants and faster tests.

• One variable per test: hook, image or offer, not all at once.
• Enough delivery per variant: with small budgets, test two variants well rather than ten badly.
• Record winners: document hypothesis, variants, result and decision.
• Watch for fatigue: if costs rise on the same creative, it’s time for a new variant.

More on AI actors and UGC ads in AI avatars and UGC ads, and on AI video in AI video for marketing.

Budget and the learning phase

AI campaigns need data to learn. With very small budgets, the algorithm gets too few results to optimise sensibly, and you get too few to learn anything.

• Match the goal to the budget: if a lead is expensive and the budget small, too few leads come in. An earlier funnel goal (e.g. landing page views) often makes more sense, knowing quality can vary.
• Don’t disturb the learning phase: frequent changes to budget, audience or creative reset learning.
• Fixed stop rules: decide in advance at what cost per result you’ll stop.
• Duration: at least a full week so weekdays even out.

My rule of thumb: if you let AI change your ads, platform terms make you solely responsible. Check AI features that are switched on by default and set approvals before the campaign starts.

Limits of measurement

Platforms measure their own success. That’s convenient, but not neutral. Three limits to know:

1. Consent: in the EU, pixels may only load after consent. People who decline don’t show up in measurement, or only in modelled form.
2. Attribution: every platform credits conversions by its own rules. The sum across platforms is often larger than real sales.
3. Click quality: not every click is an interested person. App placements or bots can inflate numbers.

Remedy: measure real outcomes where they happen (sign-ups in your newsletter tool, enquiries in your inbox, sales in your shop) and compare with platform numbers. UTM parameters on every ad help you see the source in your own data.

Labels and rules

For most businesses the TTPA is only marginally relevant. But note: Meta stopped political, electoral and social-issue ads in the EU from October 2025, and Google had announced a similar step. If you advertise on social issues, check whether your ad could fall under it.

Audience: the algorithm needs signals, not a wish list

With Advantage+ and Performance Max, precise targeting matters less than it used to. The algorithm finds who responds by itself. Your job shifts: you supply good signals so it learns what a valuable contact looks like.

• The right goal: optimise for the event that makes you money, not clicks. Click goals find people who like clicking.
• Clean events: pixel and Conversions API send the same event without duplicates; check it in Events Manager.
• Respect consent: without consent in the cookie banner the pixel must not load. Fewer signals are the price of staying compliant.
• Exclusions: exclude existing customers and your own test devices, or you pay for people you already have.
• Check placements: after every change, look at the breakdown by placement. Unsuitable placements can bring many cheap but worthless clicks.

My rule of thumb: an AI system is only as good as the goal you give it. Set the wrong goal and you get the wrong thing very efficiently.

The ad and the landing page belong together

Many campaigns fail not because of the ad but because of the page after it. Someone who taps an ad expects the same promise in the same words. If the page drifts, they leave, and the algorithm learns your ad doesn’t work.

• The page headline picks up the ad’s line.
• The main action sits in the first mobile screen.
• Check mobile load time before money flows.
• One goal per page, no menu full of distractions.
• UTM parameters per ad, so your own data shows what lands.

How to build and test landing pages with AI is covered in AI landing pages.

A simple four-week test plan

Before launch, decide at what spend without results you stop. Setting that limit in advance protects you from the feeling that you just need to wait a little longer.

Reading results properly

After the first weeks you have plenty of numbers, but only a few matter for the decision. Read them in this order: are events arriving cleanly? Which placements do results come from? What does a result cost? And how good are the contacts, measured in your own data, such as whether sign-ups turn into calls or purchases?

• A low cost per click isn’t success if nobody does anything afterwards.
• A high cost per result can pay off if the contacts are good.
• Day-to-day swings are normal; decide by weeks, not hours.
• Write down every change with a date, so you later know what worked.

Checklist before launch

• Goal, budget, duration and stop rule written down
• Optimisation goal matches the budget
• Placements chosen deliberately, Audience Network decided on purpose
• Pixel and Conversions API after consent, test event checked
• UTM parameters on all ads
• Creative: claims backed up, no invented testimonials, AI labelling where needed
• Landing page matches the ad, loads fast, checked on mobile
• Default-on AI features (text, image variants) reviewed
• A person has signed off ad and settings

Metrics

12 prompts for AI ads

My offer: [description]. Audience: [audience]. State the one benefit that’s understood in an ad in three seconds, without promises I can’t back up.

Write ten opening lines for a video ad about [offer], two each in the patterns problem, outcome, objection, curiosity, contrarian. Max 10 words.

Write five primary texts (max 125 characters) and five headlines (max 40 characters) for a Meta ad about [offer]. Honest, no superlatives, with one clear action.

Write 15 headlines (max 30 characters) and 4 descriptions (max 90 characters) for a Google Search ad for [query] and [offer].

Describe three image ideas for an ad about [offer] for [audience]: scene, person (no resemblance to famous people), light, framing for 4:5 and 9:16. No text in the image.

Plan a creative test with [budget] over [days] on [platform]: hypothesis, variants (one variable), success metric, stop rule. Say where the budget is too small for a conclusion.

Here are my campaign’s search terms from the last 30 days: [list]. Which are irrelevant? Suggest negative keywords, grouped by reason.

Here’s my Meta campaign broken down by placement: [table]. Where do clicks look like junk (many clicks, few landing page views)? What should I exclude?

Check this ad for risks: [copy + image description]. Misleading claims, health or income promises, missing AI labelling, possible political advertising. A first pass, not legal advice.

Here are my ad’s daily numbers: [table with cost, click rate, frequency]. Are there signs of fatigue? What should I test next?

Meta reports [number] leads, my newsletter tool [number] sign-ups with UTM [campaign]. How do I explain the gap, and which number should I trust for decisions?

This ad leads to my page: [ad copy]. Here’s the page: [text]. Does the headline match the ad? Suggest a headline and subhead that pick up seamlessly.

The most common mistakes

1. Trusting automation blindly. Review placements, search terms and final URLs regularly.
2. The wrong optimisation goal. Optimise for clicks and you get clicks, not necessarily customers.
3. Too many variants for the budget. Then neither the platform nor you learns anything.
4. Taking platform numbers as truth. Always compare with your own data.
5. AI creative without labelling. Disclose realistic AI people; no invented testimonials.
6. Constant changes. Every bigger change resets the learning phase.