The usual advice on keyword harvesting is to wait until a search term has two or three orders at a good ACOS, then make it an exact-match keyword. Amazon’s own guides never say that. We tested every order level — one, two, three, four to ten, eleven or more — on 2,846 keywords from a real KDP catalogue, and checked whether the new keywords added sales or just moved them.
The rules, and who actually says them
| Question | Amazon’s own guides | The usual advice |
|---|---|---|
| When to promote a search | No number. Run about two weeks, then add high-performing queries as exact match keywords. | Two or three orders at or under your target ACOS |
| When to add a negative | After at least 20 clicks | Anywhere from 5 to 20 clicks with no order |
| Negate it in the old campaign? | Not mentioned | Usually yes |
So the thresholds people argue about come from blogs and tool makers, not Amazon. Here is what each one looked like in our data.
What it is based on
One publisher’s US Amazon Ads account, 27 books, April 2025 to September 2026. Every keyword in the study was created from a customer search term that had already made at least one order in an automatic, broad or phrase campaign for the same book. Before promotion those searches had 49,602 clicks and 10,170 orders; as keywords they took 43,904 clicks and 7,770 orders. No book or keyword is named. The account’s break-even ACOS is 38%, so “under break-even” below means 38% or less.
To compare like with like, the level-by-level table uses the 466 keywords that got at least 20 clicks — enough for a fair test.
What each order level produced
| Orders before | Keywords | First 20 clicks: conversion | First 20 clicks: under break-even | Lifetime ACOS | Lifetime: under break-even |
|---|---|---|---|---|---|
| 1 | 147 | 15.1% | 14% | 54% | 14% |
| 2 | 86 | 18.2% | 27% | 50% | 20% |
| 3 | 52 | 14.7% | 15% | 53% | 14% |
| 4–10 | 136 | 19.2% | 18% | 48% | 16% |
| 11 or more | 45 | 17.6% | 13% | 36% | 18% |
Keywords with at least 20 clicks. “Under break-even” is the share of keywords at 38% ACOS or less. Lifetime ACOS is all those keywords’ spend divided by all their sales.
Does a second order make a better keyword?
At the start, yes. Over their first 20 clicks, 27% of keywords from two-order searches ran under break-even, against 14% from one-order searches. That gap is unlikely to be chance: our 95% range for it was 2 to 24 points. Two-order keywords also converted a little better early on, 18.2% against 15.1%.
Over the keywords’ whole life, the gap shrank to 20% against 14%, and at that size it could be chance (the range ran from −4 to +17 points). Three-order keywords did no better than one-order keywords at all, which is not what you would expect if each extra order made a search more reliable. We also compared a lot of slices, and the more you compare, the more one of them looks special by luck.
The honest reading: a second order is a mild plus for the first few weeks, not a reason to hold back a search that has sold once. Waiting for it costs something too — only about one in seven one-order keywords ever reached 20 clicks, so most searches are rare, and a rare search can take months to sell twice.
Where it really changes: four orders and more
The clear steps come later. Keywords from searches with four to ten orders ran at 48% ACOS over their lifetime, and those with eleven or more at 36% — the only level under break-even. These are searches people make often: 32% of eleven-plus keywords reached 20 clicks, against 14% of one-order keywords. Volume, not just quality, is what makes them pay.
The search term’s ACOS: ignore it
The rule “only promote at your target ACOS” did not survive. We tried to predict which harvested searches would end up under break-even from everything we knew before promotion — the search’s ACOS, conversion rate, cost per click, the book’s price, order count. Tested on books the model had not seen, it did no better than a coin toss. The search term’s ACOS on its own scored 0.48, where 0.5 is chance. The full ACOS table.
The one signal that helped: recent orders
| Orders in the 30 days before promotion | Under break-even | ACOS |
|---|---|---|
| None | 7% | 62% |
| 1 | 15% | 53% |
| 2–3 | 18% | 51% |
| 4 or more | 30% | 38% |
A search that was selling right before promotion made a better keyword than one whose orders were months old. Keywords whose search last sold within a week ran at 40% ACOS; those whose last order was more than 90 days earlier ran at 72%. Part of this is the season: many “recent” searches were promoted in the run-up to Christmas, when everything sells better. Outside that season the signal was weak. Still, if you have more candidates than budget, promote the ones selling now first.
Does the new keyword steal sales from the old campaign?
This is the worry behind “negate it in the old campaign”: that a new exact keyword just takes orders the automatic or broad campaign would have made anyway. We compared each promoted search with similar searches that were never promoted, over the same dates — same order count, same recent clicks, same time since the last order, and source campaigns just as active.
| Orders before | Searches | Extra orders in 60 days | Lost by the old campaign | Cost of the extra orders |
|---|---|---|---|---|
| 1 | 424 | +2.5 | about none | 53% ACOS |
| 2–3 | 224 | +3.0 | about none | 42% ACOS |
| All | 672 | +2.5 | 0.1 orders | 53% ACOS |
For searches with one to three orders, the new keyword added sales. A search that has sold once almost never sells again on its own: the never-promoted searches averaged 0.02 orders in the next 60 days. Even on the harshest assumption — that every search would have kept selling at its old rate — about two-thirds of the new keyword’s orders were new.
Searches with eleven or more orders were different. There, the keyword mostly took over sales the old campaign was already making — but it made them cheaper, with ACOS falling from 52% to 39%. For a big search, the keyword is about control and cost, not new sales.
The real problem: what the new keyword pays per click
If keywords from one-order searches add sales and convert as well as the rest, why do they run above break-even? Price. We compared what each keyword paid per click with what its search had paid before promotion:
| Promoted in | Keyword cost per click vs the search’s | Under break-even | ACOS |
|---|---|---|---|
| Oct–Dec 2025 | 1.05× | 26% | 41% |
| Jan–Mar 2026 | 1.39× | 4% | 70% |
| Apr–Jun 2026 | 1.52× | 3% | 62% |
| Jul–Sep 2026 | 1.51× | 1% | 72% |
Keywords with at least 20 clicks, by the quarter they were promoted.
Conversion held steady at 15–21% throughout. What changed was the bid: keywords promoted in 2026 started well above what their search had been costing, and paid about half as much again per click for the same orders. Clicks across the whole account also got about a third dearer over the same period, so not all of it was the keyword bids. But at the late-2025 price, the 2026 keywords would have run near 50% ACOS instead of over 70%.
Two cautions. A new exact keyword may win more top-of-search placements than the automatic campaign did, and those clicks cost more but can be worth more. And the late-2025 group had the Christmas season behind it. Neither changes the practical point: start a new keyword near what the search already costs, and raise it only if it earns it.
Free tool · no sign-inSearch term harvesterDrop in your search term report: it lists the searches to promote with a starting bid at what each one already costs per click.How to harvest, level by level
| Search term had | What to do | Expect (lifetime ACOS) |
|---|---|---|
| 1 order | Promote. Start the bid at the search’s own cost per click. Judge it after about 20 clicks. | about 54% |
| 2–3 orders | Promote, the same way. Expect a slightly better start. | 50–53% |
| 4–10 orders | Promote these first if budget is tight. | about 48% |
| 11 or more | Promote now. Expect it to take over the search from the old campaign, at a lower cost. | about 36% |
- Rank by recent orders, then order count. Not by the search term’s ACOS.
- Start the bid at what the search already pays per click, capped at your break-even bid.
- Give each keyword a fair test — about 3 ÷ conversion rate clicks — before you judge it. When to pause.
- Negate a search after about 20 clicks with no order, as Amazon suggests. Five clicks is too few: a search that converts at 15% shows no order in its first five clicks more than 40% of the time.
- If the keywords sell above break-even, look at the bid before the keyword. The price per click, not the search, was what pushed ACOS up in our data.
The common rules, scored
| Rule | Who says it | What our data says |
|---|---|---|
| Wait for 2–3 orders | Blogs and tool makers | Partly: two orders gave a better start, not a reliably better keyword. Waiting delays every winner. |
| Only promote at your target ACOS | Blogs and tool makers | Not supported: the search’s ACOS predicted nothing. |
| Negate after 5–10 clicks | Blogs and tool makers | Too early for books converting at 15–20%. |
| Wait 20 clicks before a negative | Amazon | Supported: about 3 ÷ conversion rate. |
| Harvesting just moves sales | A common worry | Not for 1–3 orders: those sales were mostly new. True at 11+, where it made them cheaper. |
The limits of this study
- One account. One publisher, one marketplace, books only. Your numbers will differ; test the patterns.
- Small groups. 86 two-order and 45 eleven-plus keywords reached 20 clicks. Treat single percentages as rough.
- Matched, not randomised. Promoted searches may have been searched more often than their matches, which would flatter the extra-sales figure. The worst-case figure (about two-thirds new) allows for that.
- Seasonality. Much of the late-2025 advantage came with Christmas.
- Page reads and organic sales are not counted.
Common questions
How many orders should a search term have before I harvest it?
One is enough. In our data, keywords from two-order searches started a little better, but over their lifetime one, two and three orders performed about the same (50–54% ACOS). The clear improvement came at four orders and especially eleven or more (36%).
Do search terms with two orders perform better than those with one?
Early on, yes: 27% of two-order keywords ran under break-even in their first 20 clicks, against 14% of one-order keywords. Over the keywords’ whole life the gap shrank to 20% against 14% and could be chance, and three-order keywords did no better than one-order ones.
Does harvesting a search term just steal sales from the auto campaign?
Not for searches with one to three orders: compared with similar searches that were never promoted, each new keyword added about 2.5 orders in 60 days while the old campaign lost almost none. For searches with eleven or more orders, the keyword mostly took over existing sales, but at a lower ACOS.
What is keyword harvesting in Amazon Ads?
Taking customer search terms that made sales in automatic, broad or phrase campaigns and adding them as exact-match keywords, so you can bid on each one directly.
What does Amazon recommend for harvesting search terms?
Amazon’s guides give no order or ACOS threshold. They suggest letting a campaign run about two weeks, then adding high-performing queries as exact-match keywords, and waiting for at least 20 clicks before adding a search as a negative.
What bid should a harvested keyword start at?
About what the search was already costing per click, capped at your break-even bid. In our data, keywords that started about 1.5 times above their search’s cost per click ran at 62–72% ACOS; those that started near it ran at 41%.
- Promote a search after its first order; a second order helps the start, not the lifetime.
- The clear steps are at 4–10 orders (48% ACOS) and 11+ (36%).
- Ignore the search term’s ACOS; recent orders are the one signal that helped.
- For 1–3-order searches, the new keyword adds sales instead of moving them.
- Start bids at the search’s own cost per click: overpaying, not bad keywords, drove ACOS up.