Negative keywords are usually described as "removing terms you do not want."
A more accurate framing: they move budget from positions that will not convert to positions that will. The difference matters, because it sets the test — not "I dislike this term" but "money on this term does not come back."
Three classes safe to negate
One: impressions with no conversion
The order keyword lookup's conversionType sorts terms into four classes, and I is ineffective impressions — visibility, no conversion.
This is the cleanest class to negate, because the judgment is not a guess: the endpoint has already marked it from actual behaviour. How to optimize Amazon keywords covers how to treat all four.
Two: irrelevant terms whose visibility is entirely purchased
Every term from the traffic lookup carries a badges array naming where it places you: naturalSearching, amazonChoice, sponsorBrand, sponsorVideo, ads and others.
A term with ads and no naturalSearching means the marketplace does not consider you naturally relevant to it. Your visibility there is entirely bought.
That is not automatically bad — pushing a term during a launch looks exactly like this. But when a term sits in that state for a long time without converting, the platform's judgment and your conversion data agree: it does not belong in your targeting.
Three: low relevance terms
The keyword miner returns relevancy. Low relevance on a term you are still running usually means spillover from broad match.
This class deserves a look before acting, because relevance is a computed value and occasionally underrates genuinely relevant long-tail phrasing. Low relevance plus no conversion is the safe condition; either one alone can misfire.
Combining the three
When all three signals point at the same terms, negating is safe:
| Signal | Field | Meaning |
|---|---|---|
| Impressions, no conversion | conversionType is I | Money spent, nothing returned |
| Visibility entirely bought | badges has ads, no naturalSearching | The platform sees no natural fit |
| Low relevance | Low relevancy | Semantically it really does not match |
All three hit → negate. Only one → keep watching.
The point of the tiering is that negating is irreversible within a campaign cycle, and killing a term that was about to gain traction costs more than a few extra days of ad spend.
When not to negate yet
This is the boundary worth emphasising most.
Do not negate on thin data. A term with 200 impressions and no order does not establish that it will not convert — if the category's normal conversion rate is around 1%, 200 impressions were never expected to produce even one order.
Negating on an insufficient sample cuts exactly the terms that have not had a chance to prove themselves.
The method is plain: estimate the category's normal conversion level, then check whether the term's impressions were ever enough to produce one conversion. If not, keep watching rather than negating.
The keyword conversion endpoint's clickConvRate gives that estimate a reference point — it is the market's click conversion rate for the term rather than yours, but it is enough to judge the order of magnitude.
Rule out the other problem first
One situation looks like "should be negated" and is not.
A term gets impressions and clicks and no orders. If relevance is the problem, negating is right. But there is another possibility: the term is fine and the listing does not hold people.
The conversion endpoint separates the two:
| Symptom | Where it breaks | What to do |
|---|---|---|
Low clickingRate | Nobody clicks in the results | Main image, price, title — not a negatives problem |
Low clickConvRate | They click and do not buy | A detail page problem; negating will not fix it |
| Both low | Probably relevance | This is the case for negating |
If clickConvRate is low across all your terms, the problem is the listing rather than the term selection. Negating one by one there is using ad budget to paper over a product page problem, and it will not hold.
Exclude at query time too
The excludeKeywords parameter drops unwanted roots while mining and while checking conversion, and filterRootWord collapses by root.
This is a different thing from negating in the ad console — that acts on your targeting, while this acts on your candidate set. Used together, the same irrelevant terms stop reappearing in your selection process.
For selection itself, see Choosing Amazon PPC keywords.
Three things this data cannot do
It cannot give you your own search term report. Which searches actually triggered your ads and what each cost lives in Seller Central as authorized data. What is here is market-side distribution, useful before the fact and not a substitute for the report afterwards.
It cannot tell you which negation type to use. Exact or phrase negation depends on how wide you want the exclusion, and that is a decision in the ad console.
It cannot judge relevance for you. relevancy is a computed value. Whether a term fits your product still needs a human to look.
Questions
Should the negative list be cleaned periodically? Yes. After product changes or category shifts, a term you negated may have become relevant. Revisit the list quarterly rather than only ever adding to it.
Can negating too much hurt traffic? It can, which is exactly why you do not negate on a thin sample. The goal is raising the efficiency of each unit spent, not minimising impressions.
What about irrelevant terms surfaced by automatic campaigns? That is the main source of negatives. Exposing those terms is part of what automatic targeting is for: after a run, negate the ones that do not convert and promote the ones that do into manual campaigns.
Can negatives be chosen from public data alone? For pre-emptive candidate exclusion, yes. For review after the fact you need your own search term report — only it knows what your ads actually triggered.