Smart Bidding Still Needs a Babysitter. Here's the Job

Smart Bidding automates one decision — the bid, per auction, toward your target — and does it with signals no human can process. What it does not automate is everything around that decision: whether the target is realistic, whether the data feeding it is intact, whether budget or learning states are distorting it, and whether tomorrow resembles yesterday. That perimeter is the babysitting job. It is judgment work, it is still yours, and an AI reading the account makes it a few visible minutes a week instead of a vague monthly worry. The job has five parts, and each has a prompt.

What Smart Bidding actually optimises — and what it ignores

Smart Bidding predicts conversion probability auction by auction and bids accordingly, using signals unavailable to manual bidding (Google's overview lists the mechanics). Given a sane goal and clean conversion data, it out-bids any human.

The blind spots are structural, not bugs. It optimises toward recorded conversions — break tracking and it cheerfully optimises toward a lie. It treats the target as truth — hand it an impossible CPA and it obeys by buying almost nothing. It has no idea a conversion's value differs from its price — margins, lead quality, lifetime value live outside the account. And it extrapolates yesterday — a sale, a season, a launch arrives as a surprise unless you warn it. Every item on that list is a judgment input. The list is not a criticism; it is the job description for the human half. The machine bids; you govern what it bids toward.

Before you blame the algorithm

Every "Smart Bidding stopped working" session should open with five questions, in order: Is conversion tracking recording what it recorded last month? Is the budget capping delivery? Is the target within sight of what the account has actually achieved recently? What changed — targets, structure, conversion actions — inside the learning window? And has demand itself moved? Most complaints end at one of the five, and none of the five is the bidder's fault. The weekly prompt below asks all of them at once, which is faster than the meeting where everyone has a theory.

The five signals a bid strategy is struggling

Five patterns, all visible in data the account already holds — and all routinely blamed on the algorithm when they are actually constraints:

1. Volume collapse after a target change. Conversions fall hard right after a tightened target — the strategy is refusing auctions it no longer believes in. The tightening did the damage, not the market.
2. "Limited by target" status. The strategy's own confession that the efficiency bar is capping volume. Budget sits unspent while the target does the strangling.
3. Budget constraint on a value strategy. "Limited by budget" under target ROAS means the system cannot buy the auctions it wants — performance degrades in a way no target tweak will fix.
4. Actual drifting from target for weeks. A gap that persists beyond noise means the target and the market disagree; one of them has to move, and it will not be the market.
5. Perpetual learning. A strategy re-entering learning every week is being churned by repeated edits — usually a hand that cannot stop adjusting.

The weekly prompt that surfaces all five:

For every bid strategy in account 123-456-7890: strategy type,
status, days since last status change, target versus actual CPA or
ROAS over the last 30 days, and any target changes in change history
this quarter. Flag the five struggle signals: post-change volume
drops, limited-by-target, limited-by-budget, sustained target-actual
gaps, and strategies that keep re-entering learning.

Minutes to read, and each flag arrives with its evidence attached — which matters, because four of the five signals have fixes that point in different directions.

One structural note: portfolio strategies — one strategy governing several campaigns — pool their learning and their status, so a signal at portfolio level needs reading against every member campaign. Tell the prompt which campaigns share a strategy, because a "struggling" portfolio is sometimes one bad member dragging four healthy ones.

When to change targets — and how much at once

Target changes are the sharpest tool on this perimeter, and the discipline is boring on purpose:

• Move gradually. Small steps, singly — think a modest fraction of the current target, not a halving. Each large jump both shocks volume and partially restarts learning, so aggressive corrections cost twice.
• One change per learning cycle. Change, then wait out the window before judging or changing again. Serial adjusters generate signal number five and then blame the machine.
• Change targets for economics, budgets for scale. Wanting cheaper conversions is a target conversation; wanting more at current efficiency is a budget one. Crossing the two — raising budget to fix a target problem, tightening targets to fix overspend — is the most common self-inflicted wound in automated bidding. When in doubt, write the intended outcome down first — cheaper, or more — and let that sentence choose the lever.

Have the AI log the discipline: every target change, its date, its size, and a review date past the learning window — a one-line ledger that ends relitigating from memory. The ledger pays for itself the first time someone asks why volume fell in March and the answer is a dated line instead of a debate.

Learning phase resets: what genuinely triggers them

Not every edit resets learning, and knowing the difference removes a lot of superstition. Genuine triggers: switching strategy type, large target moves, major conversion-action changes — new counting rules, new primary actions — and structural surgery on what the strategy governs. Mostly harmless: adding a keyword, refreshing an ad, modest budget moves within normal range.

The status column is the arbiter — "Learning" with a reason beats any folk theory. During learning, performance wobbles by design; the babysitter's job is to hold the review date and stop anyone (including you) from "fixing" a strategy that is mid-calibration. If a client or a boss needs the wobble explained, the status column is again the friend: "the strategy is learning, and here is the date we judge it" is a sentence that buys patience, with the advantage of being true. Where the line sits between governing the machine and doing its job for it is the subject of the wider automation layers guide.

Seasonality adjustments before sales, not during

Smart Bidding extrapolates recent history, so a predictable conversion-rate spike — sale weekend, launch day — arrives as a pleasant surprise it underbids into, and the return to normal as a nasty one it overbids through. A seasonality adjustment is the warning mechanism: you declare the expected conversion-rate shift for a short window, the system pre-adjusts, and it unwinds cleanly at the window's end.

Used right, it is for brief, sharp, predictable events — days, not seasons — set just before the event with an honest estimate from last year's data. Used wrong — long windows, guessed magnitudes, set mid-event — it distorts bidding worse than saying nothing. The AI's contribution is the estimate: "Compare conversion rate during last year's sale window to the surrounding weeks and tell me what adjustment that history supports." If last year's three-day sale converted at half again the surrounding fortnight's rate, that history — not optimism — is the adjustment you declare, for those three days only, removed on schedule. Declare it before the event; during is too late to be a warning.

The babysitting job, in total: one weekly status prompt, target changes made gradually and logged, learning windows respected, seasonality declared ahead of the calendar. A machine that bids better than you, governed by judgment it cannot have — that is the division of labour working as designed. Handled this way, Smart Bidding stops being a thing that happens to the account and becomes a thing the account does on purpose, which is the entire difference between automation you supervise and automation you suffer.