Amazon PPC benchmarks from 4,088 search terms and 120 days on one working account: how much spend goes nowhere and how concentrated sales really are.
Almost every Amazon PPC statistic you can find is either a vendor’s marketing number or an “industry average” with no stated source. So here is the opposite: every number below comes from one working Amazon account — my own, Rowdy Rooster Woodworks — pulled the day this was published, with the method written out at the bottom so you can argue with it.
The window: 120 days, 29 April to 26 August 2026, US marketplace. 4,088 distinct search terms, 695,942 impressions, 13,850 clicks, 2,550 orders. Account ACoS over the period: 35.4%.
Reported as percentages and ratios throughout. This is one account, not a survey — see the limits section before quoting it as an industry figure.
1. Four out of five search terms never produced an order
3,284 of 4,088 terms — 80.3% — produced zero orders. Only 19.7% produced even one. And 3,281 of those zero-order terms did not merely appear: they took clicks and were paid for.
23.5% of all ad spend in the period went to search terms that never produced a single order. Not overspend on weak terms — spend on terms with nothing to show at all.
That number is why the negative-keyword automation exists, and it is the specific component doing the work here. I run nine negative automations, one per campaign group, and the rule is deliberately blunt: 20 clicks or more with zero orders over a 30-day lookback → add the term as a negative exact at campaign level. One tighter variant fires at 10 clicks over 14 days for a faster-moving product. Right now there are 473 negative exacts live on the account. Every one of them is a term I am no longer paying for. (The free PPC waste calculator runs this same test on your own search-term report.)
2. About 1% of search terms drive well over half of sales
The top 10 terms produced 37.1% of sales. The top 50 — 1.2% of all 4,088 terms — produced 57.1% of sales and 57.8% of orders.
An Amazon account is not a broad portfolio with a long profitable tail. It is a very short head and a very long tail of noise. Which makes finding the head, and then owning it deliberately, the entire job.
The component for that is the harvest automation. I run eight of them. They read qualifying search terms on a schedule, promote the winners into an exact-match campaign at a bid I set, and — if the negate-in-source box is ticked — drop a negative exact back into the source so two of my own campaigns stop competing for the same search. To date they have produced 221 exact-match harvests across 109 distinct terms, plus 30 ASIN harvests into product-targeting campaigns. (The workflow, in full: auto to exact harvesting.)
3. One order proves nothing — and the data says how much nothing
This is the finding I did not expect to be so clean. 609 terms produced exactly one order. Only 132 reached three or more.
If you promote a search term the first time it converts, you are promoting 609 coincidences for every 132 real signals. My harvest rules gate on orders greater than or equal to two or three within a 30-day lookback, and I picked those numbers by instinct long before I ran this query. The instinct turns out to be doing real work: the threshold is the difference between harvesting a pattern and harvesting a fluke.
Set your own threshold to match your volume — a bigger account can afford to be stricter. The principle transfers even when the number doesn’t.
4. The most expensive click is the cheapest customer
Same 120 days, same campaigns, same products — split by where the ad appeared:
| Placement |
Conversion rate |
ACoS |
Relative cost per click |
| Top of Search |
29.7% |
31.3% |
1.71× |
| Rest of Search |
14.0% |
35.5% |
1.00× |
| Product Pages |
12.3% |
45.6% |
1.12× |
Top of search costs 53% more per click than product pages and 71% more than rest of search. It also converts 2.4× better than product pages — and it has the lowest ACoS of the three. The priciest click produced the cheapest customer, and it is not close.
The component here is placement optimization, which runs inside the AI feature rather than as a standalone rule. Five of my AI profiles have it switched on, capped at a 10% step per change and a 200% ceiling on any one modifier, and it is human-review-only by design — it proposes, it never applies. Over the period it put up 18 placement adjustments: I applied 17 and rejected one. Those, plus earlier modifier work, are the 290 placement updates the logs record. The technique it is executing is a low base bid with a high top-of-search modifier on proven targets, which buys the position that converts without raising the bid everywhere. (Written up here: bidding for top of search.) One hard caution: modifiers stack on dynamic bidding, so an 80¢ bid can clear well over $3.
5. Break-even ACoS ranged from 28.9% to 67.1% — inside one catalog
Across nine FBA products in the same account, break-even ACoS ran from 28.9% to 67.1%, averaging 49.6%. That is a 2.3× spread between products sold by the same seller in the same category.
So the “keep ACoS under 25%” rule would have been wrong on every single one of them — strangling nine profitable products to hit a number that describes none of them. My account ran 35.4% ACoS against a 49.6% average break-even: about 14 points of headroom, which reads as reckless if you believe the rule and comfortable if you know your own numbers.
The component is the per-product P&L: it subtracts COGS, inbound shipping, referral fee and FBA fulfillment fee per unit, so break-even is computed per SKU rather than assumed. That is the difference between having a target ACoS and having the right one. (Run yours: break-even ACoS calculator, or read the per-SKU breakdown.)
6. What ran unattended, and what didn’t
Worth stating plainly, because it is the part most tools are vague about. I run 35 automations. Thirty-two of them propose changes and wait for me to approve them — every harvest, every negative, every placement adjustment. Two AI Managers are set to apply on their own, on two specific products I have watched long enough to trust. One dayparting automation reconciles a schedule I set, so there is nothing to approve.
That split is the honest version of “human in the loop.” Approval is the default, and unattended is a thing you switch on deliberately, per product, once you have earned confidence in it. It works that way because autonomous bidding in previous tools damaged this account more than once.
Two supporting numbers from the same logs: of 20,457 recorded actions, 161 failed — 0.8%, mostly Amazon API rejections that were retried. And one-click undo has been used zero times, which I read as the approval queue catching things before they ship rather than as evidence the undo is unnecessary.
Separately, on the review side: 2,414 review requests sent through Amazon’s official Solicitations API over 90 days, with 253 orders correctly skipped as ineligible and zero failures. That one does run unattended, because requesting a review through Amazon’s own endpoint cannot go wrong the way a bid can.
Method, and what this is not
Search-term figures come from Amazon’s search-term report for the US profile, 29 April to 26 August 2026, aggregated per term across the period. A zero-dependency parser that computes these same figures from any Amazon search-term report — wasted spend, sales concentration, harvest and negate candidates — is published as MIT-licensed code, with a synthetic sample report, in our public GitHub repository. The full calculation method is documented there as well. Placement figures come from placement-level performance data over the same window. Break-even uses the current price and Amazon’s own referral and fulfillment fee estimates per ASIN, minus my recorded COGS and inbound shipping; four further products are FBM and carry no FBA fee, so they are excluded rather than counted as zero.
The honest limits. This is one account in one category — wood finishing products — run by someone who manages it daily. It is not a survey and it is not an industry benchmark. A category with different margins, price points or competition will produce different numbers. Ad-attributed sales also run above what the storefront records, because Amazon attributes on a window; ratios inside this dataset are consistent, but do not read the sales figures as total business revenue. And 120 days includes seasonality I have not adjusted for.
What does transfer is the shape: most terms are noise, a handful carry the account, one order is not proof, position matters more than click price, and break-even belongs to the product rather than to a rule of thumb. If your account disagrees with mine, your account is the one that’s right — the point is to go and look.
.png){kind=small}
Partner