What is Automated bidding?
Automated bidding hands the per-auction bid to the platform, which predicts the value of each impression and bids accordingly. The advertiser still sets the objective and the target, and those two inputs decide almost everything about the result.
What is being automated
Automated bidding moves one decision to the platform: what to bid in each individual auction. Everything above that decision stays with the advertiser.
The advertiser still chooses which outcome matters, sets the target, sets the budget, writes the ads, defines the audience, decides what counts as a conversion and decides what a conversion is worth. The platform contributes a prediction — how likely this particular impression is to produce the chosen outcome — and converts that prediction into a bid.
Stated that narrowly, the case for it is straightforward. The prediction depends on signals the advertiser cannot see and could not act on at auction speed if they could: the person's recent behaviour, the device and context, the time, the query's exact wording within a match, patterns across a volume of data no account holds on its own. A single manual number cannot express any of that.
The case against it is not that the prediction is bad. It is that the objective and the target are still the advertiser's, and a good model pointed at a poorly chosen objective pursues it very efficiently.
The three inputs that decide the outcome
The event. A bidder optimises toward whatever event the account nominated. Nominate a click and it finds people who click. Nominate an add-to-cart and it finds people who add to carts. The single largest improvement available in most accounts is moving the optimisation event closer to the thing the business actually sells, and the constraint on doing so is volume rather than ambition.
The value. Where the strategy is value-based, the number sent back with each conversion is what the model maximises. A static value on every conversion is not a value signal. Revenue is adequate where margins are even and misleading where they are not. Profit is the version that matches what the business wants and costs some engineering to supply.
The target. A target derived from contribution and close rate bounds the campaign at what it can afford. A target taken from a benchmark or from the account's best week bounds it at a number the market may never supply, and the symptom is not expensive delivery — it is no delivery.
The learning period is real and is not a formality
Every material change puts a campaign into a state where the model is rebuilding its estimates. Delivery is uneven, cost per outcome swings, and neither is representative of what the campaign will do once it settles.
The changes that trigger it are the significant ones: the target, a large budget move, a new creative set, a structural edit that changes what the campaign contains. Small adjustments are absorbed.
The failure mode is a loop. A campaign is changed, it delivers badly while learning, the account reads that as the change failing and reverses it, which triggers another learning period, and so on. Accounts stuck in this loop conclude that automated bidding does not work for them, and what is not working is the cadence of changes.
The discipline is unglamorous: change one thing, wait at least a full conversion window and preferably longer, then read the result. Where several campaigns need changing, stagger them.
What it cannot fix
Measurement. A model optimising toward a duplicated tag will find the duplication. A model optimising toward a conversion that does not predict revenue will find conversions that do not predict revenue. Automated bidding amplifies whatever the measurement says, which makes measurement quality more important under automation than it was under manual bidding, not less.
Creative. Auctions rank on expected value, so an ad nobody engages with is expensive to serve at any bid. No bidding strategy compensates for a weak hook or an offer nobody wants; it just pays more for the same indifference.
The page after the click. Everything between the click and the conversion is outside the auction entirely.
An impossible target. If no auction clears at the target, the bidder declines them all. The campaign underspends, reports a cost near the target on very little volume, and looks broken.
Where it genuinely underperforms
Three situations, all about evidence.
Thin conversion volume. Below the platforms' own threshold the predictions rest on almost nothing, and the variance in delivery can be worse than a manual bid's predictability.
A campaign whose job is not optimisation. Brand defence wants to appear cheaply above anyone bidding on the name. A value-predicting model correctly sees that those clicks convert and bids accordingly, which buys the same position for more money.
Very recent structural change. A campaign rebuilt last week has a model trained on a campaign that no longer exists.
Reading a campaign that is running one
The useful diagnostics change under automation.
Bid-level reports stop meaning much, because there is no bid to read. Impression share and its two lost columns matter more, because they separate a budget constraint from a competitiveness one. The spread of costs around the average matters, because a target is an average and a wide spread is normal. And the gap between the target and what was achieved, tracked over weeks rather than days, is the honest read on whether the target is one the market can supply.
Above all, the question to ask of a struggling automated campaign is which of the three inputs is wrong — the event, the value or the target — before asking anything about the algorithm.