Amazon keyword research: from listing vocabulary to validated terms

Keep listing vocabulary separate from search behavior

Amazon keyword research identifies relevant language for a product and tests it against customer search evidence. Competitor listings are one source of ideas. Customer queries, advertising reports, and eligible account analytics answer different questions; none should be substituted for another.

Review language is another useful source of phrasing, but a review is not a search log. A customer can describe a product one way after buying and search another way beforehand. Keep those signals labeled throughout the workflow.

1. Build a relevant, reproducible listing corpus

Select comparable products in one marketplace and language. Record the collection window, child ASINs, parent relationships where known, and fields collected. Mixing titles, reviews, A+ content, and specifications without identifying them makes a term's apparent importance difficult to interpret.

1. Save the title and bullet text for each selected ASIN, retaining the original text for checking.
2. Deduplicate repeated ASINs. Identify repeated family copy so one large family does not silently dominate the corpus.
3. Distinguish a valid empty field from a retrieval failure. Exclude a document missing the fields required by your analysis and report that exclusion.
4. State whether each document is a child listing or a deduplicated family. Use the same unit in every coverage denominator.
5. Keep marketplace languages separate. Translate examples for explanation, not as a substitute for English-market query research.

Output: a corpus manifest: ASIN or document ID, marketplace, field set, language, timestamp, family, and inclusion status. If the source cannot be reproduced, present its figures as a historical published example rather than a newly verified dataset.

2. Normalize text without losing product meaning

Normalization makes comparable forms match. A practical method is to lowercase text, normalize Unicode, split punctuation consistently, and remove a documented list of common function words. Preserve numbers, units, brand terms, and negation for contextual checking even if you exclude them from one scoring pass.

Decide how to handle compounds, plural forms, and spelling variants before comparing listings. In the German example below, the published tokens transliterate umlauts: oel represents Öl, and fassungsvermoegen represents Fassungsvermögen. That is an analysis convention, not recommended customer-facing copy. Stemming, lemmatization, and merging synonyms can change counts; record those choices.

Inspect phrases as well as individual words. “Less fat” and “fat-free” make different claims. “Two dishes at once” implies a capability that not every appliance has. A frequent word is not a safe recommendation until its surrounding phrase and your product's capabilities are understood.

Output: a versioned cleaning recipe and a candidate list linked back to source passages. This is a reproducible workflow to use for new research; the historical sample does not publish every cleaning setting needed to recreate it.

3. Use coverage and TF-IDF for different jobs

Listing coverage answers how widespread a term is in the selected corpus:

coverage(term) = documents containing the term / eligible documents × 100

Count a document once for coverage even if it repeats a word ten times. Report the numerator, denominator, field set, and matching rule whenever raw data is available. A missing document is not automatically a document without the term.

TF-IDF emphasizes terms that occur within a document but are less common across the corpus. One explicit implementation is:

TF(term, document) = term count / total retained tokens
IDF(term) = ln((N + 1) / (document frequency + 1)) + 1
TF-IDF = TF × IDF

Here, N is the number of eligible documents and document frequency is the number containing the term. This smoothed formula is an instructional implementation; the exact historical Scoutvo weighting settings are not supplied with the July example. Do not assume numerical scores are comparable between implementations.

A term can have high coverage and modest IDF because it is common across the category. A rare brand name can have high IDF but be inappropriate for your listing. Coverage identifies shared vocabulary; TF-IDF helps inspect distinctive language. Neither measures search volume, sales, or a ranking factor.

4. Read the July amazon.de example with its limits

The published July 2026 air-fryer sample reports 1,692 unique words across 54 listings. It labels these term percentages as listing coverage, not TF-IDF scores or query volume. The original article supplies rounded percentages but no term-level raw counts, so counts cannot be reconstructed reliably from the display.

These are German-market findings, not recommended keywords for an English-language marketplace. Exact collection times, a corpus export, and field-level availability are not published here. The nominal corpus is 54 listings; the raw data needed to confirm every term's eligible denominator is unavailable. Treat the values as reported historical figures, not a reproducible current benchmark.

The four flagged tokens suggest passages to inspect. They do not prove that the analyzed product was excluded from searches or that using them would increase sales. “Less fat” would need a defined comparator and product-specific substantiation. “At the same time” needs a truthful explanation of what can be cooked simultaneously. Do not copy a competitor's percentage claim.

There is no universal rule that terms above 90% are mandatory or that 40–70% coverage offers a ranking opportunity. Product fit, query evidence, brand restrictions, and clear language determine whether a term belongs.

5. Validate relevance and query evidence

For each candidate, keep a small decision record: source phrase, product capability, supporting specification or test, intended meaning, available query evidence, and proposed placement. Reject competitor brand names and claims you cannot support.

For eligible brands, Amazon Brand Analytics provides query and search-funnel dashboards. Access requires an eligible selling account and brand role. Read the selected report's definitions, marketplace, period, and scope; incomplete query coverage does not establish zero demand.

Use customer queries from your own advertising reports where available, keeping advertising traffic distinct from organic search. Autocomplete can suggest language but is not a precise search-volume measurement. Vendor estimates need the provider, marketplace, date, and an explicit estimate label. If a term has only competitor-text evidence, keep it as a vocabulary hypothesis.

Output: an approved shortlist of relevant terms, with evidence gaps visible. Never translate a high-scoring German term and assume equivalent English demand.

6. Place useful terms in accurate, readable copy

Use the title for the product's identity and important differentiating facts, then explain supported benefits and usage in bullets and descriptions. Avoid making every field repeat the same list. Follow the current title, attribute, and content requirements for your category and marketplace.

Amazon's keyword guide recommends natural placement rather than stuffing. Its search-term guidance addresses the Generic Keyword field used for matching. Check the current field rules in your own marketplace before editing; do not rely on an assumed universal character or byte limit.

Use backend terms for relevant additional vocabulary allowed by the applicable rules. Do not publish misleading synonyms, unrelated products, unsupported performance claims, or competitor brands. Use A+ content to explain verified product information; do not promise that every content surface is indexed in the same way as titles or backend fields.

Output: a proposed edit with the original wording preserved, the changed fields identified, and product or policy approval recorded. See the listing audit for attribute and content checks.

7. Verify the saved edit and measure outcomes

Before changing copy, record the current field values and a relevant baseline period. After saving, confirm the correct child ASIN and marketplace, the stored values, and the rendered listing. A successful save does not establish search indexing.

If you perform manual search checks, log the query, time, delivery location, device, and whether a placement is sponsored. Seeing a result is one observation; failing to find it in a limited search is not definitive proof of non-indexing. Investigate suspected catalog or suppression problems using available Seller Central diagnostics.

For performance, define the denominator before comparing periods. Query impressions, clicks, purchases, units, and product sessions are different measures. Evaluate available funnel data across comparable periods and note price, advertising, inventory, seasonality, and Featured Offer changes. Where an eligible experiment is available, use it to test content rather than attributing every before-and-after change to keywords.

Illustrative calculation: if a defined query report records 200 clicks from 10,000 impressions, click-through rate is 2%. A later 3% rate is an observed difference; without controlling other changes it is not proof that a wording edit caused the increase. These figures are explanatory, not Scoutvo customer results.

Retain accurate wording even when a term fails to show a measurable effect. Revisit candidates when products, customer language, or the competitive set change—not merely to increase a keyword count.

Frequently asked questions