Most advice on Amazon Ads says to wait until a search term has proven itself — two or three orders at a good ACOS — before you make it a keyword. We tested that against seventeen months of data from a real publisher’s US account: 2,846 exact-match keywords, each one created from a search term that had already sold. Two of the most common rules did not survive.
What the study is based on
Everything below comes from one publisher’s Amazon Ads account in the US marketplace, advertising a catalogue of KDP books. No book, keyword or campaign is named; only totals are shown.
| This study | |
|---|---|
| Period | April 2025 to September 2026 |
| Whole account | 254,473 clicks and 32,569 orders across 2,882 campaigns |
| Keywords studied | 2,846 exact-match keywords |
| How each one qualified | The same words had made at least one order as a customer search term, in a different campaign for the same book, before the keyword existed |
| Clicks on those keywords | 43,904 |
| Orders from them | 7,770 |
| Left out | Search terms from exact-match campaigns, and keywords that never received an impression |
We grouped the keywords by how many orders their search term had before it was promoted, then followed what each keyword did afterwards. To keep later bid changes from muddying the picture, we also measured every keyword over its first ten clicks.
Finding 1: one order is as good as three
| Orders before | Keywords | Ever sold | ACOS | Conversion | First 10 clicks |
|---|---|---|---|---|---|
| 1 | 1,030 | 27% | 58% | 16.3% | 14.6% |
| 2 | 746 | 25% | 55% | 17.2% | 16.6% |
| 3 | 337 | 30% | 57% | 15.8% | 14.4% |
| 4 | 201 | 37% | 51% | 19.5% | 17.9% |
| 5 | 118 | 39% | 56% | 16.5% | 18.1% |
| 6–10 | 272 | 46% | 46–52% | 12–22% | 14–20% |
| 11 or more | 142 | 46% | 38% | 19.5% | 16.6% |
ACOS and conversion are each keyword’s lifetime figures. “First 10 clicks” is conversion over each keyword’s first ten clicks, the fairest comparison. Rows from 6 to 10 orders are small, so they are shown as one range.
Keywords promoted after one, two or three orders behaved the same: about 55–58% ACOS and 14–17% conversion. The first real step up came at four orders, and only search terms with eleven or more orders produced keywords that ran at around 38%.
The “ever sold” column rises with order count, but mostly because search terms with more orders are searched more often, so their keywords get more clicks and more chances to sell. Click for click, a one-order keyword sold almost as often as a three-order one.
One nuance: over their first 20 clicks, keywords from two-order searches did start better than one-order ones, but the gap did not last and three orders did no better than one. Every level, in detail: keyword harvesting by order level.
Finding 2: the search term’s ACOS predicts nothing
The second rule — only promote search terms with a good ACOS — fared worse. Here are the keywords that reached ten clicks, grouped by the ACOS their search term had before promotion:
| Search term ACOS | Keywords | Keyword ACOS, first 10 clicks | Ever sold |
|---|---|---|---|
| 0–30% | 266 | 65% | 80% |
| 30–50% | 221 | 62% | 84% |
| 50–70% | 112 | 67% | 74% |
| 70–100% | 44 | 68% | 84% |
| 100–150% | 40 | 58% | 83% |
| Over 150% | 15 | 51% | 80% |
At every order level, the correlation between the search term’s ACOS and the keyword’s ACOS was close to zero (between −0.09 and +0.01). A search term at 20% did not make a 20% keyword: only 8% of fair-tested one-order terms that started under 30% stayed under 30%. And search terms that looked expensive did as well as the cheap ones.
The reason is arithmetic. An ACOS worked out from one to three orders mostly measures luck: one sale after a cheap click looks brilliant, the same sale after an expensive click looks terrible. Neither tells you much about the next hundred clicks.
Finding 3: most keywords that sell, sell within a few clicks
| Keywords that reached | How many | Had sold by then |
|---|---|---|
| 5 clicks | 946 | 51% |
| 10 clicks | 698 | 71% |
| 15 clicks | 556 | 84% |
| 20 clicks | 466 | 90% |
| 30 clicks | 327 | 97% |
Among the keywords that made a sale, the first one typically came by the third click; three in four had sold by the seventh and nine in ten by the thirteenth. These are counted day by day, so the true figures are a little lower still.
The catch is how few keywords got that far. 1,254 of the 2,846 — 44% — never received a single click, and another 894 got between one and nine. Most of the keywords that “didn’t work” were never tested.
Finding 4: pauses ended more tests than bad keywords did
Among keywords promoted after a single order, 470 never spent anything. 369 of those sat in campaigns that were paused after a median of just 14 days. Of the 255 that got one to nine clicks without a sale, 184 had been paused. A campaign full of new keywords, paused after two weeks, settles nothing.
Finding 5: cutting the bid rarely fixes a high ACOS
The account’s target was 38%. 197 one-order keywords ran above it and had their bids cut, the first cut at least three weeks before we measured:
| After the first cut | Keywords |
|---|---|
| Got under the 38% target | 2 |
| Kept selling, still above target (ACOS 64% → 60%) | 16 |
| Clicks but no sale | 3 |
| Live, no clicks in the last 30 days | 7 |
| Paused (keyword, campaign or both) | 169 |
The sixteen that kept selling improved by about four points. On these keywords, a lower bid bought fewer clicks rather than cheaper sales, because the conversion rate belongs to the search, not to the bid.
Finding 6: keywords with no clicks are rarely under-bid
Of 93 live keywords that had never been clicked, 82 were already bidding more than the original search term paid per click when it converted. They were not losing auctions; hardly anyone was searching those exact words.
AcosControl · Start freeEvery search term that sold, ready to become a keywordAcosControl lists the search terms that have made orders and are not a keyword yet. One click turns one into an exact-match campaign.What we recommend
- Promote a search term after its first order. Waiting for a second or third order bought nothing in this data. It only delayed the keyword.
- Ignore the search term’s ACOS when you decide. If you have more candidates than budget, rank them by order count, then by clicks per order.
- Set your expectations by order count, and judge each keyword against your own break-even ACOS, not someone else’s target:
| Search term had | Expected ACOS | Conversion |
|---|---|---|
| 1–3 orders | 57% | 16.4% |
| 4–10 orders | 51% | 18.5% |
| 11 or more | 38% | 19.5% |
- Give every new keyword a fair test: 10–15 clicks, or one royalty’s worth of spend, before you judge it. How to decide when to pause.
- Don’t pause a campaign of untested keywords. If you need to save money, pause the keywords that have had their test, not the ones that haven’t.
- If a tested keyword runs above break-even, decide rather than trim. Lowering its bid will mostly cut its sales. Keep it near break-even if you want the sales, or pause it.
- Don’t keep raising bids on keywords nobody clicks. They cost nothing while they wait, and a higher bid rarely creates searches that aren’t there.
The limits of this study
- One account. One publisher, one marketplace, books only. Your catalogue and prices will shift the numbers; the patterns are what to test.
- Added, not moved — mostly. A follow-up compared promoted searches with similar ones that were never promoted: for searches with one to three orders, the keywords added orders rather than taking them from the original campaign. The comparison.
- Organic lift is not measured. Ad sales count as ordinary sales in KDP, so they may help a book’s rank and organic sales. None of that is counted here.
- Survivorship. Keywords that sold tended to be kept running, so the later rows of the clicks table flatter long-running keywords.
- Attribution. Book ads credit sales for 14 days after a click; recent days were still filling in when we measured.
Common questions
How many orders should a search term have before I make it a keyword?
In this study, one. Keywords promoted after one, two or three orders performed the same: about 55–58% ACOS and 14–17% conversion. The first clear improvement came at four orders.
Should I promote a search term with a high ACOS?
Its ACOS is not a reason to skip it. Across 2,846 keywords, the search term’s ACOS before promotion had almost no relationship (correlation between −0.09 and +0.01) with the keyword’s ACOS afterwards.
How many clicks should a new keyword get before I judge it?
About 10–15. Of keywords that reached 10 clicks, 71% had sold by then; of those that reached 15, 84% had. Keywords that sold usually did so by their third click.
Does lowering the bid bring down a high ACOS?
Rarely, for keywords harvested from search terms. Of 197 that ran above a 38% target and had their bids cut, two got under target; the sixteen that kept selling improved from 64% to 60%.
What ACOS should I expect from a keyword made from a search term?
In this data: about 57% when the search term had one to three orders, 51% with four to ten, and 38% with eleven or more. Compare those with your own break-even ACOS.
- One order is enough to promote a search term: keywords from 1, 2 and 3-order terms performed the same.
- The search term’s ACOS before promotion predicted nothing about the keyword’s ACOS after.
- Most keywords that sell do it fast: 71% of those reaching 10 clicks had sold by then.
- 44% of the keywords never got a click; paused campaigns ended more tests than bad keywords did.
- Bid cuts rarely brought a high-ACOS keyword under target; decide keep-or-pause instead.