Finding related Amazon keywords from the products that rank

Two keywords are related when the same products rank on both. How the overlap method works, how to measure it, and how to turn it into decisions.

Most keyword research starts from words: take a seed term, add modifiers, collect suggestions. That finds phrases, but it doesn't tell you whether a phrase belongs to your market. A more reliable test looks at the results instead of the words. If the same products rank on page one for two keywords, Amazon is treating those searches as the same shopping need, whatever the wording.

This article explains the method behind "reverse" keyword research, the research it rests on, how to measure overlap properly, where it misleads, and how to turn the numbers into decisions.

The core idea: keywords are related when their results are

A keyword is a question; page one is Amazon's answer. Two keywords that get the same answer are, for practical purposes, the same question. Two keywords with completely different answers are different markets, even if they share most of their words.

That gives two directions to work in:

Keyword to keyword. Compare the page-one product sets of two keywords. High overlap means related.
Product to keyword ("reverse lookup"). Start from products that rank on your core keywords and list the other keywords where those same products rank. Each hit is a candidate related keyword.

Both are the same calculation viewed from different sides: a table of which products rank on which keywords.

The research behind it

Grouping queries by shared results is not a seller trick. It is one of the oldest ideas in search research.

Click-based clustering. Beeferman and Berger (KDD 2000) clustered queries from a search engine's click log by the pages users clicked, treating queries and clicked URLs as a graph and merging them step by step. Wen, Nie and Zhang (WWW 2001) stated the principle directly: "If two queries lead to the selection of the same document… then they are similar." They also found that combining click overlap with word overlap worked better than either alone.
Result-based relatedness. A Xerox patent by Natalie Glance (filed 1999) judged queries related by the documents they return, not by their terms. Its example rule: two queries are related if their top results share at least one document.
Purchase-based equivalence at Amazon. Amazon researchers wrote that "two queries should be considered equivalent if they lead to the purchase of the same or similar set of products" (Hiranandani et al., 2019). They also noted that this signal is noisy for rare tail queries, which is why later Amazon work maps tail queries onto popular queries with similar meaning.
Relevance labels. Amazon's public Shopping Queries Dataset (ESCI) contains about 2.6 million human judgements for roughly 130,000 queries, labeling each product as Exact, Substitute, Complement or Irrelevant for the query (Reddy et al., 2022).

Sellers can't see clicks or purchases for other people's products. But page one is the visible result of those behavioral signals, as described in how Amazon organic ranking works. Ranking overlap is the closest public proxy for what the research measures.

How to measure overlap

Jaccard: shared products over all products

The simplest measure is the Jaccard index: the number of products on both pages divided by the number of distinct products across both pages.

Jaccard = shared products ÷ (products on A + products on B − shared products)

It runs from 0 (nothing in common) to 1 (identical sets). It ignores order, so a shared product at the top counts the same as one at the bottom of the page.

Rank-weighted overlap

Order matters on Amazon: the top of the page gets most of the attention, as covered in share of page one. Rank-biased overlap (RBO), from Webber, Moffat and Zobel (ACM TOIS, 2010), was designed for exactly this case: comparing ranked lists that only partly share members, while weighting top ranks more heavily.

RBO compares the two lists at every depth (top 1, top 2, top 3 …), takes the share of overlap at each depth, and sums those shares with weights that shrink geometrically. A parameter p sets how fast they shrink. With p = 0.9, the formula in the paper puts about 86% of the total weight on the first ten ranks. Lower p focuses harder on the very top.

Use Jaccard when you ask "do these keywords show the same products at all?" Use a rank-weighted measure when you ask "do they show the same products where it matters?"

Compare against a set, not one keyword

Overlap with a single head term is fragile. A sturdier signal is how many of your core competitors (products that rank on several of your core keywords) appear on a candidate keyword's page one. If most of them are there, the candidate is part of your market even when its overlap with any single keyword looks modest.

A worked example: "dog crate"

The example below is illustrative. Letters stand for generic products; the pages are shortened for readability.

Keyword	Page-one products	Shared with "dog crate"	Jaccard
dog crate	A B C D E F G H	(reference)	1.00
metal dog crate	A B C D E F J K	A B C D E F	0.60
collapsible dog crate	J K R A D B F H	A B D F H	0.45
dog kennel	A B C E G L M N	A B C E G	0.45
dog crate for car travel	P Q R A S T U V	A	0.07
dog crate cover	W X Y Z … (covers only)	none	0.00

How to read it:

"metal dog crate" shows mostly the same products. It belongs in the set, but it adds little new information; it behaves like a variant of the head term.
"collapsible dog crate" and "dog kennel" have the same Jaccard score, but in this example A and B sit at the top of both "dog kennel" and "dog crate," while the shared products on "collapsible dog crate" are spread lower. A rank-weighted measure would rate "dog kennel" as closer.
"dog crate for car travel" shares one product. Page one is mostly travel carriers: a neighboring market, not the same one.
"dog crate cover" shares nothing. In ESCI terms, its results are complements. High relevance to your customer, zero competition for your spot.

Reverse lookup adds candidates you didn't think of. If products A, B and D also rank on "puppy playpen with crate" and "wire dog cage," those keywords become candidates to test with the same overlap check.

Rules of thumb for thresholds

There is no published standard threshold for Amazon keyword overlap. The bands below are working rules of thumb, not research findings. Calibrate them on your own niche.

Jaccard with your core keywords	Typical meaning	Usual action
0.5 and above	Near-duplicate or same intent	Keep one as the main term; track the other at a lower tier
about 0.2 to 0.5	Related, same market	Add to the market with a fitting tier
about 0.05 to 0.2	Adjacent market, partial overlap	Add only if you sell into it; otherwise watch
below 0.05	Different market	Leave out, or build a separate market

Two practical checks keep the bands honest: look at page one yourself before acting on a borderline score, and check how many of your core competitors appear, not only the raw score.

What reverse lookup can and can't see

Reverse lookup sounds like a window into a competitor's keyword strategy. It is narrower than that.

Only keywords someone captured. A reverse lookup can only return keywords that were searched and recorded. If nobody captured "crate for 70 lb dog," no lookup will show that a competitor ranks there.
Ranking, not revenue. A product appearing on page one does not mean that keyword sells for it. Only the brand owner sees its own Search Query Performance data, which lists the top queries that led customers to its products. That report requires Brand Registry.
Volumes are estimates. Amazon's Brand Analytics shows a search frequency rank and "the three top-clicked products for each search term," not raw search volumes. Any volume next to a reverse-lookup result is an estimate from a third party. A product ranking fifth on a big keyword won't appear in Amazon's top-three list at all.
One snapshot, one location. Page one varies by delivery address, time and personalization. A lookup reflects one capture under one set of conditions. See Amazon search results history for why a single snapshot is weak evidence.
Page one only. A product can be indexed for a keyword and sit on page three. Page-one data won't show it, which is usually fine: page three is rarely where the competition is decided.

Pitfalls that distort overlap

Variations. Amazon may show a different child of the same family on different keywords: the large size on one, the medium on another. If you compare child listings, overlap looks lower than it is. Group children into families first; variations and rankings explains why.
Brand queries. A competitor's brand name returns mostly that brand's products. Overlap with your core keywords will be low or misleading, and the keyword measures brand demand, not your category. Keep brand terms out of the analysis.
Broad heads overlap with everything. A head term like "dog crate" shares products with nearly every keyword in the niche, so high overlap with it proves little. Compare candidates against your whole core set, and treat overlap with the head term alone as a weak signal.
Products that rank everywhere. A dominant product that appears on almost every keyword inflates overlap between unrelated terms. Give such products less weight, the way text search discounts very common words.
Seasonality and events. Around Prime Day, holidays or a seasonal peak, page one can shift toward deal products and seasonal variants. Compare pages captured on the same day, and confirm a relationship across several dates before acting. Amazon's own search reports can be pulled by week, month or quarter, which helps separate seasonal keywords from stable ones.
Near-identical spellings. Amazon researchers describe semantic matching that handles synonyms and spelling errors (Nigam et al., KDD 2019). Plurals and misspellings often return nearly the same page. High overlap here is expected and doesn't justify tracking both at full weight.

Turning overlap into decisions

The point of the measure is a clear decision for each candidate keyword:

Add it. Moderate to high overlap with your core set, and your product fits the page. Add it to the market.
Tier it. Weight by demand and fit, not overlap alone. A high-overlap long-tail term can still be a minor keyword. How many keywords to track covers weighting and why near-duplicates add less than they seem.
Merge near-duplicates. Very high overlap plus near-identical wording means one keyword is enough at full weight.
Split into a separate market. Low overlap with your core set but a page full of products you do sell (a travel carrier, a playpen line) means a different market. Track it on its own so it doesn't drag down averages.
Find competitors. Run the logic the other way: products that appear across many of your keywords are your real competitors, often different from the brands you would name. Tracking competitors on Amazon covers what to do next.

Re-run the check every quarter. Markets drift, and a keyword that was related last year may now show a different product type.

How JustRanks supports this

JustRanks records the whole organic first page of Amazon.com search results for more than 200,000 keywords over time, from a fixed US delivery address, overnight US time; ads never count. Keywords saved in any user's market are captured every day, all others every 48h. Variants are grouped as one product family, and related keywords can be found from products that rank on several of your keywords. Missing keywords can be added at no extra cost, and each keyword in a Market gets a tier (Core, Important, Standard or Minor) that sets its weight.