When should an AI agent change a bid?

The hard part of building an agent that manages bids isn't teaching it to change them. It's teaching it not to.

Most signals are noise

A keyword got four clicks yesterday and no orders. Should the bid come down?

Almost certainly not. Four clicks tells you nearly nothing. If that keyword converts at 10%, then no orders in four clicks is the most likely outcome — it happens about two-thirds of the time. A tool that cuts the bid here isn't optimising. It's reacting to a coin flip.

Now multiply that across a few thousand keywords, every day. A system that responds to every fluctuation generates thousands of changes that are uncorrelated with reality, and calls it optimisation because the dashboard is busy.

Changing a bid has a cost

Each change carries costs that rarely appear in anyone's model:

• It resets your evidence. Performance after a change tells you about the new bid, not the old one. Change often enough and you never accumulate enough data about any bid to know whether it was right.
• It moves you in the auction. Position changes conversion rate, which changes the data you're about to learn from. You are altering the measurement instrument mid-measurement.
• It compounds with other changes. Bid, budget, and placement changes interact. Three individually sensible adjustments can produce a nonsense outcome together.

Three questions before any change

1. Is there enough data for this to mean anything?

Not "is there a signal" but "would I see this signal if nothing were wrong?" If the answer is frequently, wait.

2. Is the predicted improvement bigger than the uncertainty?

A predicted 2% gain with a ±15% range is not a 2% gain. It's a coin flip with extra steps. The prediction's confidence matters as much as its value.

3. Would I still make this change tomorrow?

If the answer depends on today's data, it's a reaction, not a decision. Reactions average out to zero at best.

What "wait" looks like

Restraint isn't passivity. While an agent is not changing a bid it should still be accumulating conversion evidence, watching for changes big enough to be real, tracking competitor movement, and noticing when the product's economics shift.

The agent is working. It just isn't touching anything, because the strongest available action is frequently no action — and a system that cannot express that will always find something to do.

The asymmetry that decides it

Acting on noise has an expected value of roughly zero, minus the cost of the change, minus the evidence you destroyed.

Waiting for signal has an expected value of roughly zero, plus better data tomorrow.

Those are not the same bet. This is why "how often does it change bids?" is the wrong question to ask a PPC tool. The right one is: how does it decide whether a change is worth making at all?