The big platforms have spent billions making their algorithms smarter. The promise: hand over your goal, the algorithm finds the customers. The reality: it works if you feed it good signals and you don't get rolled by the parts it does worse than you would.
Where the algorithm wins
Bidding
Manual bid management is dead for any account with conversion data. Auto-bidding (Target CPA, Target ROAS, Max Conversions) outperforms manual bidding once you have ~30 conversions in a 30-day window. Don't fight this.
Audience expansion
Lookalikes and "advantage+ audiences" use signal you couldn't replicate. They find people who look like your converters across thousands of dimensions you'd never think to target.
Creative testing
Modern campaigns let you upload 5 headlines + 5 descriptions and the algorithm tests combinations. It'll find the winners faster than you can A/B them.
Where the algorithm loses
Performance Max — the controversial one
Google's Performance Max ("PMax") combines all networks (search, display, YouTube, Gmail, Discover, Maps) into one campaign and lets the algorithm choose where to spend. Pros: simple, often hits CPA targets. Cons: the inventory it picks is opaque — much of the spend goes to display and YouTube where attribution is shaky. If your branded search is included in PMax, your reported CPA is artificially low.
Best practice: exclude branded search keywords from PMax with a negative list, run a separate branded search campaign, and watch placements with a script that pulls them via API.
Aggressive automation on small accounts
Auto-bidding needs ~30 conversions / 30 days to work. Below that, the algorithm has no signal and goes erratic. New campaigns should start manual until they get to threshold.
Generic AI-written copy
Auto-applied "asset suggestions" produce safe, generic ad copy that ranks fine but converts worse than human-written variants. Use them as starting points, not finals.
The four signals you should always control yourself
1. The conversion definition
Don't let the platform define a conversion as "page view of /thank-you OR scroll 75%." That's a vanity event. Define conversions as the actual business action — purchase, lead form submit, qualified booking — and only that.
2. Conversion value
If your conversions have different values, send the value with each one. $8 lead, $80 trial signup, $499 paid customer. Without value data, the algorithm optimizes for any conversion equally — including the cheap, low-quality ones.
3. Negative keywords / negative audiences
Tell the algorithm what to avoid. Brands you don't want associated with. Search terms that are technically related but never convert ("cheap," "free," "competitor name"). Existing customers who shouldn't see acquisition ads.
4. The brand voice in copy
The algorithm doesn't know what your brand sounds like. Your headlines and descriptions are still your job. Write them well; the algorithm will mix and match.
The right mental model
Think of the algorithm as a brilliant junior media buyer who joined yesterday. They're faster than you. They never sleep. But they don't know:
- Your customer's real lifetime value
- Your strategic goals
- Your brand's voice and positioning
- Which "conversions" are real and which are noise
Your job: feed them clean signals (good conversion data, value data, exclusion lists, brand-aligned creative). Their job: optimize within those constraints. Anyone who tries to do their job manually loses. Anyone who hands over the constraints loses too.