Where AI in Google, Meta and LinkedIn advertising earns its keep, where it quietly wastes budget, and what to check before you give it more of your money.
Every ad platform now sells automation as the answer to weak results. Google has Performance Max, Meta has its Advantage+ campaigns, and LinkedIn builds predictive audiences from your own lead and conversion data. The pitch is the same everywhere: give the system a budget and a goal, step back, and it will find your customers.
What actually works is narrower than the pitch. AI clearly beats people at setting bids auction by auction, and it has made producing ad variations far faster and cheaper. It fails when it is told to chase the wrong result, such as form fills that are mostly spam, because it then spends your budget very efficiently on the wrong thing.
The difference shows up directly in what each new customer costs you. This article covers where automation earns its place, where it quietly wastes money, how to check that your accounts feed it the right information, and which decisions should stay with you.
Google's automated bidding, known as Smart Bidding, sets a separate bid for every auction. It weighs signals such as the device, the location, the time of day, the exact search and whether the person has visited your site before, then bids higher where a conversion looks likely and lower where it does not. No person can do that across thousands of auctions a day, and manual bidding can only adjust for a few of those signals, in broad steps.
You still choose the goal. With a target CPA, short for cost per acquisition, you tell Google what you are willing to pay for each conversion. With a target ROAS, short for return on ad spend, you tell it how much revenue each euro of advertising should bring back. Meta's delivery system works on the same principle, which is why Meta now steers advertisers towards broad audiences and lets the system find the buyers within them.
The same logic sits behind broad match keywords paired with Smart Bidding, and behind Performance Max, which runs one campaign across Google's channels from Search and Shopping to YouTube and Gmail. Both hand the system decisions you used to make yourself, such as which searches and placements are relevant. Google has also been phasing out Enhanced CPC (cost per click), its halfway house between manual and automated bidding, so if any of your campaigns still rely on it, plan the switch.
The condition is volume. Automated bidding learns from conversions, and a campaign that records five a month is guessing. As a working rule, aim for at least 30 conversions a month per campaign before you rely on target-based bidding, and merge thin campaigns rather than splitting the budget further. If you still cannot get there, bid for an earlier step that happens more often and still predicts sales, such as a booked call rather than a signed contract.
Say you spend €6,000 a month on Google Ads across six Search campaigns and get 60 enquiries. That is ten per campaign, too few for the bidding to learn anything reliable. Merged into two campaigns, each sees around 30 enquiries a month and the system has something to work with. From there, change targets in small steps of 10 to 15 per cent and give each change a couple of weeks before you judge it.
Automated bidding works best when you can trust the result it is optimising for, and many businesses have never checked theirs. If your campaigns treat every form fill as a success, the system will find more people who fill in forms. It cannot tell a buyer from a bot, a job seeker or a competitor checking your prices.
Say 100 forms arrive in a month and 35 of them are spam or irrelevant. The bidding counts all 100 as wins and learns to find more of the cheap, junk end. Your cost per lead looks fine, perhaps better than last quarter, while your sales team says the leads have got worse. Both are true.
The same thing happens further along the sales process. A software company that optimises for demo requests, when only one demo in five becomes a real sales conversation, is spending efficiently towards the wrong outcome. An online shop that values every order equally drifts towards whatever sells most easily, which is often the discounted, low-margin product with the most returns.
AI makes good strategy faster and bad strategy more expensive. The answer is not to switch automation off, but to tell it precisely what a good result is worth to you.
AI makes good strategy faster and bad strategy more expensive.
Before you raise budgets or hand more campaigns to automated bidding and targeting, check what the platforms are actually learning from. Work through the checks below in order; the first two need nothing more than an afternoon with your account settings.
Spam deserves its own fix. A hidden spam-trap field on your forms, email verification and one short qualifying question cost little and protect your bidding as much as your sales team's time. None of this is glamorous, but it decides whether the AI in your ad accounts learns from your best customers or from your worst form fills.
The most useful AI in paid media is often not in the bidding at all, but in making the ads. Image and video generators, synthetic voice-over, automatic resizing and the creative tools built into Meta and Google have cut the time and cost of producing variations to a fraction of what they were. For a business without an in-house studio, that changes what is possible each month.
Say your team used to brief and produce five new ads a month at around €400 each. With AI handling resizing, background changes, voice-over and first drafts of copy, the same €2,000 can produce 20 or more variations of those five ideas. That matters because ads wear out: the same people see the same ad too often and stop responding.
The constraint has moved from production to testing and briefing. Every ad needs enough spend to show whether it works, and a €3,000 monthly Meta budget cannot fairly test 40 ads. Test three or four genuinely different ideas (a different problem, promise or proof) and use AI to make variations of the ones that win.
If you sell in several European markets, AI translation and dubbing make it realistic to run ads in German, Czech or Hungarian without a separate production for each. Have a native speaker check every version before it goes live, because humour, idioms and product claims are exactly where machine translation goes wrong without anyone noticing.
One line not to cross: do not use AI-generated people posing as customers giving testimonials. Presenting invented endorsements as real misleads buyers, and EU consumer law treats fake reviews and endorsements as a banned practice.
AI can execute a brief well, but it cannot write one. It cannot decide which customers you should be winning, what your offer does better than the alternatives, or whether you want to be the affordable option or the dependable one. Those choices shape every campaign, and a system optimising last week's conversions will not make them for you.
People also need to spot when a campaign is working for the wrong reasons. Automated campaigns are good at finding conversions that would have happened anyway: people searching for your brand name, existing customers clicking a retargeting ad before reordering, or visitors who were already halfway through buying. The reported results look strong, but little of it is new business.
A short monthly review catches most of this. Look at the search terms and placements your money actually went to, the share of spend on your own brand name, the split between new and returning customers, and the ads the system generated in your name. If you only read the headline cost per conversion, you are letting the system mark its own homework.