How to set an Amazon bid

A useful Amazon bid starts with what a click can afford to cost, then accounts for the evidence, the current bid and the limits on the next change. Bid calculations cannot guarantee profitable sales: product economics, conversion quality, placement and stock still matter. This guide explains the reference arithmetic and the checks around it.

The one piece of arithmetic

ACOS is spend divided by sales. At the level of a single keyword, that is the same as cost per click divided by revenue per click:

ACOS = CPC ÷ RPC

Revenue per click is just total sales from that keyword divided by total clicks on it. Rearranged, you get the only formula you actually need:

Target CPC = RPC × Target ACOS

A keyword with $100 of sales across 10 clicks earns $10 per click. At a 30% target, it has a $3.00 reference target CPC. Paying $2.00 may leave room for a cautious increase if the sample, inventory and other checks support it. Paying $5.00 calls for a review, but the reference does not guarantee what the next auction will cost.

Notice what this does not do: it does not increase or decrease by an arbitrary percentage. It computes the destination. Whether you arrive in one step or several is a separate decision — see Clamp below.

Calculate from CPC, never from bid

This is the single most costly error in Amazon bidding and it is completely invisible while you are making it.

Your bid is a status. It is whatever it happens to be set to right now, possibly by you, possibly by an automation, possibly last Tuesday. Your CPC is data. It is what the market actually charged you over the period you are looking at.

Percentage rules applied to the bid compound this error. Imagine bids were cut hard yesterday. Your 30-day ACOS still reflects last month's higher CPCs, because that is when the spend and sales happened. A rule that says "ACOS is still high, cut another 15%" is now cutting from an already-cut bid based on evidence that predates the first cut. Run that weekly and the account collapses while every individual decision looks defensible.

The fix is to always compute the destination CPC from performance, then compare it to where you are.

The four situations

Every target in your account is in one of four states. Two call for less money, two for more.

Above target, with sales. The formula applies directly. The keyword converts, you know its revenue per click, you know what it can afford. Move toward it.

Spending with no sales. There is no RPC, because there is no R. But you are not without information — you know your average order value and roughly how many clicks it usually takes you to convert. So project forward:

Target CPC = Target ACOS × ( AOV ÷ ( clicks so far + average clicks to convert ) )

This is a rule-based projection, not a forecast of the next sale. Additional non-converting clicks lower the reference by growing the denominator. The assumptions, evidence thresholds, floors and size-of-change limits still matter; the formula cannot recover spend already incurred or guarantee a target cost per acquisition.

It also replaces the arbitrary threshold. "Archive anything over $20 with no sales" is far too patient for a $15 product and far too aggressive for a $200 one. Target CPA — AOV × target ACOS — adapts by itself.

Below target, converting well. Room to grow, but do not use the RPC formula here. A keyword with one click and one sale on a $40 product has an RPC of $40 and implies a $12 bid at a 30% target. That is not a signal, it is a coincidence. Step up modestly instead — 5–25% depending on how badly you want the volume — and let the ceiling from the formula act as a hard stop you approach rather than a destination you jump to.

Barely any traffic. Not enough clicks to calculate anything. Step the bid up until it generates enough data to be judged properly. This is the one case where a percentage increase is genuinely the right tool, because there is nothing to compute from.

Why steps, not jumps

Even when the arithmetic says a bid should halve, moving it there in one action is usually wrong.

A large bid change moves you into a different competitive tier. On broad, phrase and auto targeting, that changes which search terms you match at all — so next week's data describes a different set of queries than last week's. You have not measured your change; you have replaced the thing being measured.

Capping how far any bid can move in one run (a Clamp) keeps each change attributable. It is slower and it is the only way to learn anything.

The grace band

If your target is 30%, do not optimise a keyword sitting at 31%. Leave a band — commonly about 10% either side — where you do nothing.

Without it you generate constant small changes on targets that are already fine, each one resetting the clock on measurement and adding noise. The band is not laziness; it is what makes the changes you do make interpretable.

Bids and placements are one system

For a single placement adjustment, before other applicable bidding adjustments:

Placement-adjusted bid = base bid × (1 + placement modifier)
Realized CPC = actual spend ÷ actual clicks

The adjusted bid is not the amount necessarily paid for a click. Dynamic bidding and other applicable controls can also affect the auction bid. See Amazon's bidding guidance.

Read bids and placements together when investigating: "I lowered my bids and my ACOS went up."

Here is the mechanism. ACOS is CPC ÷ RPC. If CPC fell and ACOS rose, RPC must have fallen further. That can reflect conversion rate, order value, placement mix, attribution lag or other changes. Compare the same scope and dates. Top of search is not automatically the best-converting placement in every account.

If you cut a base bid, inspect realized placement performance before planning the next change. Separating bid and placement changes helps interpretation, though it does not eliminate seasonality or other confounding factors.

Floors, ceilings and sanity

Three guardrails worth having regardless of method:

A bid ceiling derived from the formula — never pay more per click than the keyword's revenue can support at your target. A floor below which a bid is not worth maintaining. And a cap on how many changes any one campaign receives at once: if your data says a campaign needs forty simultaneous adjustments, the likelier explanation is that something is wrong with the data.

What AutoPPC does with this

AutoPPC applies the relevant saved rules to the target's situation, limits the size of the move and retains the evidence behind the proposal. Formula-based cuts and evidence-gated raises follow different cases. Inventory-dependent raises require adequate stock evidence.

Bid and placement modifier never move on the same campaign in the same run. That is not a setting — it is an Invariant, because a change you cannot attribute is a change you cannot learn from.