Automated bidding is no longer a choice you make campaign by campaign. On most platforms it is the default, and several campaign types have no manual option at all. The useful question is not whether to use ai bidding strategies but what you owe them in return. An automated bidder is a model that predicts conversion probability for each auction and prices accordingly. It is very good at that job and completely dependent on the signals you give it. Feed it bad conversion data and it will pursue bad outcomes with impressive efficiency.
We see the same failure repeatedly. An account switches from manual CPC to a value-based strategy, performance improves for three weeks, then degrades. Nothing changed in the market. What changed is that the bidder learned, correctly, that a particular cheap conversion action was the easiest one to produce, and shifted the whole account towards it. The strategy did exactly what it was told. The instruction was wrong.
Decide what you are actually buying
Before selecting any strategy, settle the target. Target CPA optimises for a count of conversions at a cost. Target ROAS optimises for revenue against spend. Maximise conversions with no target spends whatever it is given. These are genuinely different objectives and they produce different accounts. A business with wide margin variation across products should not run target CPA, because the bidder will treat a low-margin sale and a high-margin sale as identical and buy more of whichever is cheaper to win.
If margins vary, send values, not counts. That means passing a value with every conversion, ideally gross profit rather than order total. This is the single highest leverage change available to most ecommerce accounts, and it is usually a two-hour job in the tag configuration. Until it is done, no amount of strategy switching will help.
Give the bidder enough data to learn from
All ai bidding strategies need volume. The commonly cited floor is something like thirty conversions in a thirty day window per campaign, and that is a floor, not a comfortable level. Below it the model is working from very thin evidence and its estimates swing. The remedy is structural: consolidate. Five campaigns with eight conversions each learn nothing. One campaign with forty conversions learns something.
This runs against the instinct of anyone trained on manual bidding, where fine segmentation gave you control. With automation, segmentation costs you accuracy and buys control you no longer exercise. Keep separation only where you genuinely need different budgets or different targets: distinct countries, distinct margin profiles, brand versus non-brand. Everything else should be merged.
Change targets slowly and one at a time
Every meaningful change resets some of the learning. Editing a target CPA by fifty per cent, swapping the conversion action and adding new audiences in the same afternoon leaves you unable to attribute the result to anything. Move targets in increments of roughly ten to twenty per cent and wait for a full conversion cycle before judging. If your typical lag from click to sale is nine days, a three day read is not a read.
The learning period is real but it is also used as an excuse. If a strategy has been in place for six weeks and is still underperforming, the answer is not more patience. It is usually a data problem: duplicate conversions, a missing value, an attribution window that does not match the sales cycle, or a conversion action counting an event the business does not care about.
Where ai bidding strategies still need a human
Automation prices the auction. It does not decide what the auction is for. These remain yours:
- Defining the conversion action and its value, including which actions to exclude entirely.
- Setting budget ceilings, since a maximise strategy will spend everything available.
- Deciding seasonality adjustments ahead of known spikes rather than letting the model discover them late.
- Excluding audiences and placements the business will not serve, which no bidder can infer.
- Judging whether a cheaper conversion is genuinely a better one, which requires knowing the business.
That last point deserves emphasis. Bidders optimise the metric, not the outcome. If lead quality drops while cost per lead improves, every dashboard will look better and the sales team will be the first to notice the problem. Close the loop by importing offline conversions where you can, so that the quality signal reaches the model rather than living in a CRM the platform never sees.
Run the comparison honestly
When testing one strategy against another, use the platform’s own experiment tooling rather than running them in separate campaigns. Separate campaigns compete with each other, share the same query space and produce results that mostly reflect which one got the better inventory. A proper split holds the auction constant and splits users. It takes longer, and it is the only version worth acting on.
Handing the auction to the machine is the right default in 2026. But it is a delegation, not an abdication. The bidder will pursue whatever you define as success, at scale, without hesitation and without telling you when the definition was wrong. Getting the definition right is now most of the job.
Keep reading: AI Advertising Optimization