What is Manual bidding?

Manual bidding is the arrangement where the advertiser sets a figure and the platform uses it. A keyword carries a maximum cost per click; an ad group or placement carries one; every auction that entity enters is bid at that number, adjusted only by whatever multipliers the account has configured.

It is the original model, and for a long time it was the only one. Its defining property is that nothing is inferred: the bid on Tuesday afternoon is the bid on Saturday morning, and it is the number somebody typed.

That property is both the case for it and the case against it.

Predictability. Nothing changes without somebody changing it. When an account is behaving strangely, a manual campaign is one fewer moving part, and diagnosing anything is easier with a fixed baseline.

Working on thin data. A model with a handful of conversions to learn from produces predictions barely distinguishable from noise, and a campaign bidding on those predictions delivers erratically. A manual bid does not degrade in that situation; it simply stays where it was put.

Deliberate data-gathering. A campaign whose job is to accumulate conversions so that a later strategy has something to learn from is better served by a stable bid than by a model chasing a target it cannot yet meet.

Encoding a decision. Sometimes the right bid is a policy rather than an optimisation — a defensive bid on the brand name, a ceiling somebody has agreed, a deliberate underbid on a segment the business does not want. A manual figure states it plainly.

The limitation is granularity, and it is severe.

A single figure applies to every auction the entity enters. The same number is bid for a person who has visited the site three times this week and a person who has never heard of the business, on a phone at eleven at night and a desktop at eleven in the morning, against a query that matches tightly and one that matches loosely, in a city where the product is available and one where delivery takes a fortnight.

Some of those differences matter enormously to whether the click converts. None of them is visible to a manual bid.

An automated strategy computes a separate number for each of those auctions, using signals about the person, the context and the moment that the advertiser cannot see and could not act on at that speed even if they could. Where the account has enough conversion history for those predictions to be any good, the difference is not marginal.

Bid adjustments are the partial answer: multipliers applied by device, location, audience membership, schedule and a few other dimensions, layered on top of the base bid.

They work, within their design. A business that cannot deliver to a region can bid it down; a business whose phone leads close far better can bid mobile up.

Three cautions. They compound, so several adjustments landing on one auction can produce a bid nobody intended and few accounts audit. They are set from aggregate reports, which at fine granularity are frequently too thin to support a conclusion. And they are static, so an adjustment set on last quarter's pattern persists through whatever changes since.

The rule that keeps them useful: set an adjustment to encode a business fact somebody can state in a sentence, not to chase a pattern in a report.

The practical question about moving off manual bidding is not philosophical. It is whether the account has enough of the event to support a model.

Each platform publishes its own threshold, in events accumulated within a window rather than in days; the learning phase is where those figures live. An account comfortably above its platform's threshold is leaving performance on the table by bidding manually. An account well below it will find that an automated strategy delivers unpredictably and produces a cost figure that swings without explanation.

Between those, a common and sensible pattern is to optimise toward a shallower, more frequent event — a lead rather than a sale, a registration rather than a purchase — until the deeper one has volume, rather than to stay manual or to jump to a target the data cannot support.

Fix measurement first. A model optimising toward a wrong conversion is worse than a manual bid, because it will pursue the error efficiently.

Set the first target near what the campaign has actually achieved recently, not at what somebody hopes for. A target below anything the account has reached suppresses delivery rather than lowering cost.

Change one thing. Moving the bid strategy, the budget, the creative and the structure in the same week produces a result nobody can attribute.

Wait out the learning period. The campaign will deliver unevenly while the model builds its predictions, and reading the numbers during that stretch and reverting is how accounts end up back on manual bidding convinced that automation does not work.

And keep a manual campaign somewhere if the account is large enough to afford it. A stable, predictable comparison point is worth its cost the first time something unexplained happens.

Even in accounts that run automated strategies almost everywhere, a few places keep a manual bid on purpose.

Brand defence is the clearest. A campaign on the business's own name has a job that is not really an optimisation: appear, cheaply, above anyone bidding on the name. Conversion rates there are high enough that an automated strategy will happily bid far more than necessary, because it correctly predicts the click will convert. A modest manual bid buys the same position for less.

Campaigns under legal, contractual or partner constraints are another. Where a maximum price has been agreed with somebody, a number that cannot drift is the point.

New campaigns in unfamiliar territory are a third. A manual bid for the first fortnight produces conversions at a known price and builds the history a strategy will later need, which is a more orderly start than pointing a model at a target and watching it search.

The pattern in all three is the same: manual bidding is the right tool when the correct bid is a decision rather than a prediction. Where the bid genuinely depends on how likely this particular impression is to convert, one number cannot express it, and no amount of adjustment layering will make it able to.