Most writing on Amazon keyword optimization stops at "put the words in your title, bullets and backend search terms." Nothing there is wrong, but it skips the hard part: you usually hold several hundred to a few thousand terms, and the question is which one to touch first.
It also assumes you are starting from nothing. In practice, as soon as a listing is live it is already getting impressions for a set of terms. Optimizing is not picking words from a blank page — it is classifying the set that is already running.
Separate three questions first
Three different questions come up during optimization, and they map to three different datasets:
| Question | Dataset | Endpoint |
|---|---|---|
| Which terms is my listing showing up for | Traffic keyword lookup | /v1/amazon/traffic/keyword |
| Which of those actually produce orders | Order keyword lookup | /v1/amazon/keyword-order |
| Is this one term worth continuing to fund | Keyword conversion | /v1/amazon/keyword/conversion |
Merging all three into one sheet is the main reason keyword lists go unused. The first tells you where the impressions are, the second tells you where the money comes from, and the third is what you use to decide on a single term.
If you do not have a keyword library yet, Building your own keyword library covers the step before this one: merging ABA, reverse lookup and expansion data into a table you can maintain. This post assumes you already have it.
The separation that matters most: traffic terms are not order terms
Traffic keyword lookup returns the terms this ASIN reached the first three pages of search results for. Order keyword lookup returns the terms that produced purchases.
The two sets overlap and are nowhere near equal. And most failed "keyword optimization" comes from spending the effort on the part of the first set that is not in the second: terms with high search volume where you do get impressions and never get an order.
Those terms are easy to identify, because the conversionType field on the order lookup labels them directly.
Four conversion classes, four different actions
conversionType sorts terms into four classes. They do not call for the same treatment, and that is the point of this post.
| Class | Meaning | What to do |
|---|---|---|
E | Strong converter | Defend it. Confirm page-one organic rank, give it title priority |
S | Steady converter | Maintain. Do not spend more trying to "optimize" it |
L | Losing conversion | Investigate. It used to convert; something changed externally |
I | Ineffective impressions | Impressions, no conversion. Judge relevance, then stop or fix the listing |
L and I deserve separate treatment, because they point at completely different problems.
L, losing conversion, is the class to handle first. The term used to produce orders, so relevance is fine and your listing could once carry it. That it stopped means the change is almost certainly external: a competitor cut price, a new entrant arrived, you went out of stock, your rating dropped. This class does not need keyword changes — it needs someone to find out what happened in that window.
I, ineffective impressions, is a different thing. Impressions without conversion usually means weak relevance: the person searching wants something other than what you sell. In advertising, these terms are pure cost. But before pausing them, settle one question: is relevance genuinely weak, or is it relevant and your presentation fails to hold people? The next section is how to tell those apart.
Before the click and after the click are two problems
The keyword conversion endpoint returns clickingRate (top-three click rate) and clickConvRate (click conversion rate). These two measure different stages, and they are not fixed the same way.
| Symptom | Where it breaks | What to change |
|---|---|---|
Low clickingRate | The search results page | Main image, price, first half of the title, rating count |
Low clickConvRate | The detail page | Review content, image detail, description, price comparison |
| Both low | Probably relevance | This term may not have been worth having |
Row one means people saw you and did not click. Reordering the keywords in your title helps with that; rewriting the detail page description does not.
Row two is the reverse: people clicked and did not buy. No amount of keyword adjustment helps here, because the keyword already did its job.
Separating these two is the line between "optimizing keywords" and "optimizing the listing." Often the data points at the second, and that is not a keyword problem.
Read organic rank and ad rank together
Every term from the traffic lookup carries both rankPosition (organic) and adPosition (advertising), each broken into page, index and position — which page, which slot on it, and which position overall.
There is also a badges array naming where the term places you: naturalSearching, amazonChoice, sponsorBrand, sponsorVideo, ads and others.
Line those two groups up and two kinds of waste show up directly:
One: already ranking organically and still paying for the term. When rankPosition.position is strong and ads are running on it too, that spend is buying a position you already hold. It does not have to stop immediately, but it should be a decision with a reason behind it rather than the default state.
Two: badges contains ads and not naturalSearching. That term's visibility is entirely purchased — it goes to zero when the campaign pauses. It can be deliberate, such as pushing a term during a launch, but you want to know you are doing it rather than believing the term is "won."
Where the terms go
Once the classification is done, the placement rules are short:
- The title carries the few highest-volume terms from class
E, in an order that reads as a sentence rather than a pile - Bullets and description carry the mid and long tail in complete natural sentences, one use case per sentence
- Backend search terms take spelling variants, synonyms and whatever had no home above, without repeating what is already in the title
- A+ content and image copy do not participate directly in on-site search matching; write them for conversion, not for keywords
Character and byte limits are deliberately left out here. The platform's limits change and differ by marketplace, so use whatever your own dashboard states at the time — a number hardcoded into a blog post is worse than no number once it goes stale.
How to verify the optimization worked
You need a re-check mechanism afterwards, or you cannot tell "the optimization worked" from "the category was up this month."
The order lookup returns searchRank, searchRankGv (monthly change) and searchRankGr (monthly rate of change), and supports weekly (reverseType of W) and monthly (M) queries. A workable routine:
- Save a baseline before you change anything, including each term's
conversionTypeandrankPosition.position - Re-check monthly, and look less at how far rank moved than at whether terms changed class
I→EorL→Eis the signal that the work actually landed- The traffic lookup's
monthparameter takes a past calendar month, so you can look back at the state at the time
Step 2 is the important one. Rank numbers drift with the category, while a change in conversionType is a relative judgment and far less exposed to that drift.
Three things this data cannot do
It cannot give you your own backend search term report. That is authorized seller account data. Everything here comes from the public side of the marketplace, and the two do not substitute for each other.
It cannot give you a competitor's backend terms. Same reason — only their account can see them.
It cannot judge relevance for you. Class I tells you a term does not convert, but whether that term should exist at all needs you to look at the words. The data does not know what your product actually is.
If you want to validate a lookup dataset's semantics before committing to it, How to choose an Amazon reverse ASIN tool gives an acceptance checklist you can run yourself.
Questions
How long before optimization shows results? Rank changes and conversion changes do not move together. Re-check monthly rather than watching daily; at a daily resolution the noise is far larger than the signal.
Are more keywords better?
No. Volume dilutes, and a large tail of class I terms is pure cost in advertising. Of the four classes, only E and part of L genuinely deserve investment.
Does keyword order in the title matter?
For clicks, yes — clickingRate measures exactly what people do after seeing you in the results, and the front of the title is what they read first. An order that reads naturally beats a pile.
Is it normal for one term to behave differently across marketplaces?
Yes. Search volume, competitive intensity and phrasing are all independent per marketplace. Both the lookup and conversion endpoints take a marketplace parameter; query each one separately rather than merging marketplaces into one sheet.