Keywords

Concept of the skill

Keyword work is demand translation. Raw queries are evidence of language, intent, and platform constraints; the skill turns them into targetable clusters, page/listing assignments, and measurement baselines without pretending that keywords alone create rankings. In AI search, the unit of demand is shifting from the exact keyword string toward the underlying question, entity, and topic — but the translation discipline is the same: rank evidence by provenance, map demand to the surface best able to satisfy it, then measure.

Coverage

Keyword work starts before writing or page building. This skill covers:

• Seed expansion: turning product, audience, competitor, and problem language into candidate queries.
• Intent mapping: classifying terms as informational, navigational, transactional, or commercial investigation.
• Evidence-quality triage: ranking keyword sources by reliability before trusting volume, competition, or suggested phrases, so observed demand is never confused with a model's guess.
• Query normalization: preserving raw queries while grouping variants, misspellings, synonyms, locales, and branded/non-branded language into useful decision records.
• Question and entity research: surfacing the who/what/when/where/why/how questions and the named entities behind a topic, so pages and listings can be found by AI answer engines as well as classic search.
• SERP and marketplace-result validation: checking the current result set, page/listing type, and platform-native signals before assigning intent or owner surface.
• Clustering: grouping related terms by meaning, shared results, and target surface so one page or listing owns one primary intent.
• Platform field translation: mapping query language into Etsy tags/titles, Amazon titles/search terms, Shopify page fields, and SaaS/content page targets.
• Marketplace semantic-intent checks: capturing implied use cases, buyer questions, compatibility, and product claims for intent-based marketplace discovery (e.g. Amazon COSMO/Rufus) without pretending exact-match repetition is enough.
• Long-tail strategy: finding specific marketplace and product phrases where relevance is stronger than raw volume.
• Cannibalization detection: finding cases where two owned surfaces compete for the same intent.
• Query value re-weighting: prioritizing clusters by demand, click-or-citation potential, conversion fit, owner fit, and evidence confidence instead of raw volume alone.
• Tracking methodology: choosing a baseline, cadence, and platform-specific measurement surface — including AI-citation/visibility signals — after keyword changes.

Philosophy of the skill

Keywords are not ranking spells. They are evidence of how people describe a need, what stage of the buying or learning journey they are in, and which surface can answer that need best. A good keyword pass reduces ambiguity: one query cluster, one intent, one owner surface, one measurement plan.

Agents often make three mistakes without this discipline. First, they stuff repeated terms into titles or tags instead of choosing precise phrases. Second, they copy web SEO rules into marketplace fields where character, byte, and matching rules differ. Third, they let multiple pages chase the same query, splitting relevance and making tracking unreadable.

A fourth mistake is now common in 2026: treating a keyword's raw search volume as its value when most searches end with zero clicks. A fifth, sharpened by AI research tools, is laundering an LLM's generated term list into "demand" it never measured. The right move is still translation — query language becomes a target map, not a paragraph full of repeated words — but the map now weighs question/answer intent, entity coverage, click-or-citation potential, and evidence provenance alongside volume.

Boundary Routing

1. Keyword Research Process

Seed, Expand, Cluster, Map

Evidence Quality Ladder

Not all "keyword data" is equal, and the biggest failure mode for an AI agent doing this work is treating its own generated term list as if it were measured demand. Rank evidence by provenance and never silently blend tiers:

Two preservation rules:

• Preserve the raw query. Keep the exact searcher phrasing alongside any normalized/clustered form. Normalization for clustering must never overwrite the original wording — the raw query is the evidence; the cluster label is your interpretation.
• Record the tier. Each candidate carries its provenance tier so a downstream reviewer can see whether a target rests on measured demand or on an unvalidated guess. A Tier-4 term presented as if it were Tier-1 demand is the core hallucination risk this ladder exists to stop. If a term cannot be traced to a source, keep it in a hypothesis bucket instead of assigning it to a page/listing.

Intent Validation Rules

Intent is not decided from wording alone. Before assigning a cluster:

1. Inspect the current result set or marketplace results for the query in the target locale.
2. Note the dominant surface type: guide, product page, collection, category, marketplace listing, comparison page, support page, video, local pack, forum, or AI answer.
3. Treat a query as mixed when the result set is split across incompatible surface types.
4. Assign the owner only after the target surface can satisfy the dominant intent better than the alternatives.
5. Re-check intent after major platform changes, seasonal shifts, or when Search Console query groups change.

Use wording as the first clue; use result-set fit as the decision.

Output Format: The Query Map

The deliverable of a keyword pass is not a flat list of terms — it is a structured query map, one row per candidate, so a downstream agent or human can act on it without re-deriving the research. A flat list hides provenance, intent, ownership, and measurement; the map makes each of them an explicit, checkable column.

Example row:

A row missing owner surface, baseline metric, or re-check date is a research note, not a target — promote it only when those are filled.

Search Intent Types

Do not assign a target until intent is clear. A single phrase can change owner depending on results: "custom mug ideas" likely wants inspiration; "custom mug bulk order" likely wants a transactional page.

Conversational / answer-seeking phrasing is not a separate intent type. A conversational query like "what are the best waterproof hiking boots for wide feet?" still expresses one of the four intents above (here, commercial investigation). Use query fan-out to discover these conversational and AEO-triggering variants, then map each back to its underlying intent rather than creating a fifth category.

Zero-Click Value Adjustment

A keyword's strategic value is no longer its search volume alone. With AI Overviews and assistant answers now resolving a large share of queries on the results page, definitional and FAQ-style terms with high volume often drive few clicks, while product, local, and transactional terms with moderate volume convert better than their volume suggests. When prioritizing, weight each cluster by demand × click-or-citation potential, not demand alone:

• High volume, low click-through (definitional, "what is X") — worth covering for AI-answer visibility and topical authority, but do not let raw volume promote it above commercial intent.
• Moderate volume, high click-through (product, comparison, "buy/near me", branded) — usually the higher-value target even at lower volume.
• Treat exact-volume figures from any keyword tool as estimates; the relative ordering and the intent matter more than the headline number.

Score priority clusters qualitatively against these factors (there is no per-query "zero-click probability" number any platform exposes — do not invent a formula):

The practical rule is qualitative: a broad definitional term may deserve a support article or FAQ, but it should not outrank a moderate-volume query with clearer commercial intent and measurable owner fit.

2. Platform Field Translation

Platform fields are not interchangeable. Translate the same keyword cluster differently depending on where the buyer searches.

Etsy Rules

• Write human-first titles. Etsy's search uses natural-language understanding, so robotic exact-match stuffing ("Blue Mug, Ceramic Mug, Gift Mug") is not rewarded and wastes the title.
• Lead the 140-character title with the primary keyword for shopper readability, not for a ranking boost: Etsy's current guidance states that where a term appears in the title does not affect listing ranking, but the leading words are the slice shown on mobile search and Google previews — so put the most decision-relevant terms first.
• Use tags as phrases, not isolated fragments, when the phrase reflects how a shopper searches.
• Etsy allows up to 13 tags per listing — use all 13 slots when the terms are accurate and distinct (most sellers leave slots empty or repeat title phrases, wasting matching paths). Use them for synonyms, recipients, materials, and occasions rather than repeating title words.
• Each Etsy tag has a 20-character limit (spaces count toward it), so choose compact buyer language over long keyword strings.
• Fill the listing's attributes completely. They act as structured data and power category and gift filters, often matching more reliably than free-text tags.
• Do not repeat the same concept across title, tags, and attributes unless repetition adds a real matching path.
• Include audience, occasion, material, style, use case, and product type when those details are true.
• Treat Etsy's AI title-recommendation tools as a first-party suggestion surface, not authority: before accepting suggestions, preserve useful existing search phrases in tags/description/attributes, apply changes to a small sample when risk is high, and re-check stats after enough time. When the work is live Etsy field application or Search Visibility diagnostics, hand off to the etsy skill — this skill supplies the demand map and evidence.

Amazon Rules

• Keep the product title specific to the product and within current category policy.
• Treat the 200-character title ceiling (most categories, including spaces) as a maximum, not a writing target.
• Follow Amazon's Jan 21, 2025 title-quality rules: most categories cap titles at 200 characters including spaces, restrict certain special characters (unless part of the brand name), and disallow repeating the same word more than twice (prepositions, articles, and conjunctions excepted). Verify the live category template — caps and allowed characters vary by category, and some enforce far shorter limits (e.g. 80–120 chars). Front-load the brand + main keyword + key differentiator for mobile visibility.
• Map intent, not just literal phrases: Amazon retrieval has two stages — candidate generation/matching still rewards the right literal terms being present in indexable fields, while the ranking layer (Amazon's publicly described "COSMO" common-sense knowledge system, surfaced through the Rufus shopping assistant) increasingly infers shopper intent and use-case fit. So the title/bullets must carry the precise literal terms AND describe the real use case and audience; do not assume exact-phrase stuffing alone wins ranking. (Treat the algorithm names as Amazon's described behavior, not a guaranteed internal spec; verify against current Seller Central guidance.)
• Capture unspoken intent as evidence. A query like "pregnant dress" implies an unmet need ("expandable waist"); a "toddler party dress" search implies "machine-washable" and "soft fabric". Record implied use cases, audiences, compatibility, occasions, constraints, and buyer questions (mined from reviews and Q&A) in the query map. This is intent evidence — translate it into accurate title/bullet language and product claims; place a term in backend search terms only when it is generic, accurate, and not already captured elsewhere. Capturing an implied need does not license stuffing it or inventing a claim.
• Count bytes, not characters: verify the live Seller Central backend "generic keywords" field (community/staff guidance reports ~249 bytes in the US). Overflowing the byte ceiling can prevent the backend terms from being indexed.
• Separate search-term words with single spaces. Amazon does not count spaces or punctuation toward the limit, and you do not need commas or other separators — extra punctuation earns no additional matches, so omit it.
• Do not repeat title words, brand terms, or other redundant information already captured in indexed fields in the search terms.

Shopify and Google Organic Rules

• Add keywords in readable phrases and sentences. Random keyword strings are bad UX and may be ignored by search engines and AI models.
• Use collection pages for broader product-category demand when the collection can satisfy the query — category terms typically carry several times the volume of single-product terms, and an empty collection page above the grid cannot rank for them; add unique descriptive copy.
• Use product pages for specific attributes, models, variants, themes, or audience/occasion searches.
• Make <title> text descriptive and concise (Shopify recommends ≤60 chars within its ~70-char field); avoid stuffing and boilerplate. Stuffed, boilerplate, inaccurate, or unclear titles may be rewritten by Google in search results.
• Make meta descriptions page-specific and useful (~160 chars practical target), leading with the most important keyword and a reason to click. They are snippet candidates, not ranking guarantees.
• For AI Overviews / generative search (GEO): target the conversational long-tails and the sub-questions an AI answer fans out to, but remember Google says foundational SEO still applies and no special AI-only markup, chunking, or rewriting is required. Surface the entity and question demand here; leave on-page answer structuring to seo-strategy and writing-humanizer.

3. Long-Tail Strategy

Long-tail terms are useful when specificity increases match quality. They are especially valuable for marketplaces, product catalogs, and niche content because they encode product type, audience, style, occasion, material, or problem — and increasingly for conversational and AI-answer queries.

Pattern:

[modifier] + [theme/problem] + [product or page type] + [audience/occasion]

Examples:

Long-tail does not mean every modifier belongs in the title. Use the full phrase to choose the target, then distribute true details across the appropriate fields.

Modifier Facet Matrix

The failure mode when an agent expands a head term into long-tail is inventing modifiers that sound plausible but describe nothing the product or page actually offers. The guardrail is to expand only along named facets, and to attach a modifier only when it is true of the target. Each facet is a row; fill a cell only with an attribute that genuinely applies.

Two rules:

• Expand by facet, validate by truth. A modifier earns a place only if the product/page actually has that audience, occasion, material, etc. A facet cell with no true value stays empty — do not fill it to look complete (this is where Tier-4 hallucination creeps back in).
• Cross-check against the Evidence Quality Ladder. A facet-generated phrase is a Tier-4 hypothesis until validated against autocomplete, first-party search data, or a keyword tool; do not promote it to a target unvalidated.

4. Clustering and Topical Authority

Pillar-Cluster Model

Pillar: "Print on Demand Guide"
  - Cluster: "POD pricing strategy"
  - Cluster: "How to design for POD"
  - Cluster: "Best POD providers compared"
  - Cluster: "POD profit margins explained"
  - Cluster: "Start a POD business on Etsy"

Rules:

1. One cluster owns one primary intent.
2. The pillar targets broad understanding; cluster pages target specific sub-intents.
3. Every cluster links back to the pillar, and the pillar links to all clusters.
4. Two cluster pages must not target the same primary keyword unless they intentionally serve different intent.
5. Marketplace listings can cluster too: one design family can split by product type, audience, or occasion when each listing is genuinely distinct.

A well-organized hub-and-spoke (pillar-cluster) structure is also how AI answer engines and featured snippets read topical authority: covering a topic's questions and entities completely, with consistent facts, is what makes a site a candidate to be summarized or cited. Cluster for authority, not just for individual rankings.

Clustering Tests

A cluster is valid when the terms share both meaning and a plausible owner surface. AI-generated or tool-generated clusters are drafts — validate them against raw queries, current result sets, and owner-surface fit before publishing a keyword map.

Cluster Types

Entity Consistency

AI answer engines lose confidence when a brand's facts are inconsistent across surfaces (different hours, names, prices, or descriptions on different sites) and may omit the brand rather than risk a wrong answer. When keyword research surfaces the entities a topic depends on — brand, product, person, place, attribute — flag that those entities must be described consistently wherever they appear. Naming the inconsistency is in scope; fixing the on-page schema/markup is seo-strategy's job.

5. Cannibalization Detection

Keyword cannibalization happens when two or more owned surfaces compete for the same query and same intent. It is not a problem when different surfaces satisfy different intents for the same head term. With zero-click AI answers now common, the subtler version is collapsing several distinct conversational intents into one overly broad page.

Detection Protocol

1. Export query data from Search Console, marketplace analytics, ads reports, or rank tracking.
2. Group by query and primary intent.
3. Flag queries where two or more URLs/listings get similar impressions or ranks for the same intent.
4. Inspect the current results: if the same result set appears for both targets, the surfaces may be competing.
5. Decide whether the surfaces should merge, differentiate, canonicalize, redirect, noindex, or remain separate.

Resolution Options

AI-Era Keyword Research

Current AI tools (deep-research agents, web-search tools, marketplace and Search Console AI grouping) improve discovery and synthesis, but they do not replace keyword research because they do not own platform demand data, marketplace policies, or first-party performance baselines.

Use AI agents and research tools for:

• Expanding seed concepts into candidate facets, synonyms, and conversational variants (query fan-out).
• Finding official platform docs and current policy changes.
• Summarizing competitor language from public pages/listings.
• Drafting clusters that a human or agent then validates against source data.
• Explaining why terms should be merged, split, discarded, or tracked.

Do not use AI agents or research tools for:

• Inventing search volume, conversion, rank, or competition figures.
• Treating generated phrases as evidence that buyers actually search them.
• Replacing Seller Central, Etsy/Shopify analytics, Search Console, Google Trends, ads query reports, or marketplace analytics.
• Creating pages for every query fan-out or AI-suggested variant.
• Promising AI Overview / AI Mode inclusion from keyword edits alone.

Google's current generative-search guidance says foundational SEO still applies, that query fan-out may issue related searches across subtopics, and that there is no need for special AI-only files, chunking, or rewriting. For keyword work the practical change is discovery breadth: capture conversational questions, follow-up phrasing, entities, and comparison dimensions, then cluster them by user need rather than making one page per generated variant.

6. Tracking Methodology

Ranking claims need a baseline and a re-check window. "Improved keywords" is not evidence unless visibility, impressions, clicks, rank, AI-citation share, or marketplace listing quality changed.

Track only terms tied to a decision. If a term does not affect a page/listing assignment, title/tag/search-term field, or content plan, it is research noise. For AI-search visibility, treat presence/citation in answers and downstream conversions as the signal — classic rank position alone undercounts value when most searches are zero-click. Do not promote an "AI citation rate" as a standard platform metric; no major platform exposes one — use the available reports plus citation spot-checks and referrer data.

Search Console Segmentation

When using Google Search Console data, segment before drawing conclusions:

If Search Console's AI-powered configuration or query groups are used, inspect the generated filters/grouping before acting. AI-assisted analysis reduces setup time; it does not remove the need to verify the query set and owner-surface decision.

7. Anti-Patterns

Public Grounding

Verify high-stakes platform work against current docs before applying. As of the 2026-06-11 review:

• Etsy Help / Seller Handbook: tags support up to 13 tags per listing and 20 characters per tag (spaces count; multi-word phrases allowed); listing titles can be up to 140 characters. Etsy's current guidance favors clear human-first titles and holistic listing understanding (title, tags, attributes, description, photo, reviews); title-term placement does not affect ranking, but leading terms are what shoppers see on mobile and in Google previews. Treat AI title-recommendation tools as suggestions to validate, not authority.
• Amazon Seller Central: the Jan 21, 2025 title policy says most categories may not exceed 200 characters including spaces, restricts some special characters, and bars repeating any word more than twice (except prepositions/articles/conjunctions). Backend "generic keywords" are limited by bytes (community/staff guidance reports ~249 bytes in the US field; verify the live field) — overflow can leave the terms unindexed. Amazon does not count spaces or punctuation toward the limit, so no commas/separators are needed. Capture COSMO/Rufus-style implied intent as evidence, not as license to stuff backend terms.
• Google Search Central: title links and snippets are generated automatically; use descriptive, concise titles, unique descriptions, and avoid keyword stuffing. AI features (AI Overviews) reuse the same indexed, well-structured content and need no special AI-only markup — so the question/entity coverage this skill identifies feeds AI-answer eligibility, while the on-page structuring of those answers is seo-strategy/writing-humanizer work. The generative-AI performance report breaks out by page/country/device/date but not by individual query.
• Shopify Help: add keywords in readable, natural phrases; Shopify lets merchants edit title tags (≤~70 chars, recommend ≤60), meta descriptions (~160 chars), URLs, and alt text; collection pages need unique above-the-grid copy to rank for higher-volume category terms.
• AI-search context (2026): a large and growing share of searches resolve with zero clicks via AI Overviews and assistants, so weight keyword value by intent and click/citation potential, not raw volume, and track answer-engine visibility alongside classic rank.

These are grounding points, not permanent invariants. Re-check them when platform policy matters.

Verification

• Output is a structured query map (raw query, normalized query, modifier facets, source, evidence tier, locale/surface, intent, owner surface, baseline metric, re-check date), not a flat term list.
• Every keyword cluster has one primary intent and one owner page/listing.
• Every candidate carries its evidence tier; no model-generated (Tier-4) term was promoted to a target without validation against measured or platform-native data, and raw searcher phrasing was preserved alongside cluster labels.
• SERP or marketplace-result fit was checked for priority clusters before owner assignment.
• Long-tail modifiers were expanded by named facet and each is true of the target (no plausible-but-false modifiers).
• Supporting terms are distinct synonyms or modifiers, not repeated filler.
• Priority clusters were value-weighted by demand, click/citation potential, conversion fit, owner fit, and evidence confidence — not search volume alone.
• Question and entity coverage was considered for topics that surface in AI answers, and entity facts were flagged for cross-surface consistency.
• Etsy uses all relevant tag slots as accurate phrases within 20 chars each, fills attributes, and leads a clear human-first 140-char title with the primary keyword.
• Amazon search terms stay within the current byte limit (counted as bytes), use spaces (no commas) to separate, do not repeat title/brand/other captured fields; the title is within the 200-char ceiling, repeats no word more than twice, and carries the use-case/audience language the ranking layer infers intent from.
• Amazon priority clusters captured implied use cases, audiences, compatibility, constraints, and buyer questions where COSMO/Rufus-style semantic intent may matter.
• Shopify and Google web guidance uses readable phrases, unique titles/descriptions, collection pages have unique copy, and no keyword stuffing or AI-only markup was prescribed.
• Branded and non-branded query data were separated (in the classic Performance report) when Search Console data is available.
• Cannibalization check was run before creating a new page/listing for a synonym cluster.
• Rank/impression/listing baseline (and AI-visibility spot-check where relevant) was captured before edits and a re-check date was set.
• Final prose, page construction, AEO/GEO content structuring, and IA decisions were handed to writing-humanizer, seo-strategy, or information-architecture as appropriate.

Do NOT Use When