How to Use The Helium 10 MCP and Claude for Amazon Keyword Research Andrew Bell , Amazon Lead & AI Expert 17 minute read Published: November 18, 2026 Share: URL copied Your 2027 Sellers Season Pass—Now 35% Off Start strong with 35% off Helium 10 Diamond for 6 months, plus a $50 MCP credit, exclusive MCP skills, and early access to seller checklists. Offer ends December 6. Get Your Offer Table of Contents Turn a large keyword bank into product meaning, shopper missions, and a content plan.Start with the product you actually sellConnect the data and choose the comparisonsGive Claude a structure that can grow with the bankUse all 15 COSMO relations as a reference vocabularyLet one keyword belong to several useful viewsBuild the mission around the requirementsAsk Claude for views the team can act onWork two rows into one useful image briefGive each content field a distinct jobA practical prompt for your next keyword bankMake the bank useful enough to write fromSources and case note Achieve More Results in Less Time With Helium 10 Sign Up For Free Your 2027 Sellers Season Pass—Now 35% Off Start strong with 35% off Helium 10 Diamond for 6 months, plus a $50 MCP credit, exclusive MCP skills, and early access to seller checklists. Offer ends December 6. Get Your Offer Turn a large keyword bank into product meaning, shopper missions, and a content plan. Open a wall-decor keyword bank and look at the requirements inside the phrases. Within a few rows, the research can move from product identity to placement, construction, and environmental suitability. “Metal flower wall art” raises a question about identity and construction. “Hallway wall decor” introduces a placement. “Outdoor metal wall decor” adds an environmental requirement. Sorting those phrases by search volume gives you one useful view. Understanding the requirements inside them gives you the next decision. Helium 10 MCP and Claude give you a practical way to investigate those differences. Helium 10 supplies keyword evidence through the connection. Claude can help organize the returned rows into categories, requirements, possible uses, and questions worth answering. With a disciplined workflow, you can move between those views while keeping every phrase attached to its original data. The same approach becomes more valuable as the bank grows. A collection of thousands of rows may contain several ways to describe the same product, multiple shopping missions, and repeated questions about the same missing fact. Give that collection a useful structure, and the research starts explaining what the product page needs to do. Work in that order: establish product truth, retrieve relevant demand, interpret relationships, test product fit, and give the supported opportunities a content job. Start with the product you actually sell Before collecting more keywords, establish what the exact product can support. In the September wall-decor session behind this article, the retrieved listing described a whitewashed flower decoration combining MDF and metal. That detail affects how we evaluate a phrase such as “metal wall decor.” The phrase may be commercially relevant to the category. The finished copy still needs to describe the complete construction. Pack quantity creates another useful example. “Metal flowers wall decor” appeared in the retrieved bank. A plural search phrase does not establish that this particular item includes several pieces. The product record decides the quantity. Ask Claude to prepare a compact product truth card: exact ASIN and marketplace, product type, materials, dimensions, design, included items, installation, care, and supported environments. Keep the source beside each fact. Give unresolved details their own field so they remain visible throughout the research. Images can help identify visible form, finish, and components. A photograph of a product outdoors still leaves its weather suitability to be established through appropriate product evidence. The truth card becomes the reference for every later classification. A phrase can then be relevant, conditional, or unsuitable for a reason you can inspect. Connect the data and choose the comparisons MCP means Model Context Protocol. Helium 10’s connection lets supported AI assistants request data from available Helium 10 tools. Its current setup guide lists Claude Web and Desktop as supported, with connection through Customize, Connectors, and the Helium 10 connector. Follow the official guide for current account requirements and authorization.[1] Once connected, ask Claude which keyword-research functions your session can use. The current Helium 10 documentation includes ASIN lookup, seed expansion, competitor comparison, and multi-source keyword discovery.[1] Start with the exact product and add comparison products that serve a similar need. For wall decor, inspect motif, construction, size, setting, and pack quantity. A floral wall sculpture may provide useful category language even when its finish differs. A three-piece set may introduce phrases that need closer quantity review. Those differences belong beside the proposed competitor ASINs. The saved session followed a clear sequence: 195 competitor candidates were reported Five competitors were selected and approved The expanded keyword lookup reported 255 phrases Thirty phrases were displayed in the saved conversation The session ended with a proposed allocation of 21 keyword placements The initial target-product lookup returned an empty result. Expanding through the selected competitors supplied the reported bank. That is a practical example of using the conversation to make the next research decision. Keep the reported count and the inspectable rows visible throughout the work. We can inspect the 30 displayed phrases. The remaining reported phrases are outside this worked example. The COSMO-inspired annotations and Claude workflow below are new teaching analysis built around that recorded case. Give Claude a structure that can grow with the bank A large bank becomes easier to work with when the exact keyword and its annotations live in connected records. Keep an untouched source table. Give every source row a stable ID. The worked example uses WD01 through WD30, matched to the original displayed row order; choose a stable convention for each new bank. Preserve the exact phrase, original metric names and values, source label, marketplace, reporting period when available, and retrieval settings. Keep a separate normalized field for matching. Trimming spaces or standardizing case should leave the original wording intact. That separation does two useful things. You can revise a classification without changing the evidence, and you can check every output against the original row. Use the following division of work when preparing a larger analysis: Deterministic preparation: count rows, inspect columns, identify blanks, normalize a matching key, flag duplicate observations, and validate numeric fields with code or spreadsheet operations Semantic interpretation: identify product types, attributes, candidate relationships, ambiguous wording, and possible shopper missions Evidence review: connect each requirement to the product truth card and identify the facts still needed Output checks: reconcile IDs, review changed values, check coverage, and inspect classifications before building the content plan Figure 1. A proposed Claude workflow follows three recovered row IDs through relation candidates, attribute dimensions and product-evidence checks; the classifications are teaching examples, not an executed Claude run. Use these steps as the operating brief for Claude. Their execution depends on the tools and file access enabled in your session. Claude supports CSV uploads; its guidance identifies additional requirements for XLSX and notes that content limits still apply.[2] A file that uploads successfully still needs a row-count and coverage check. For a very large bank, ask Claude to create a batch manifest before classification. This is a proposed organizational method using the tools available in your Claude session, rather than an automatic Helium 10 connector feature. Each batch should carry the same product truth card, relation definitions, decision rules, and output fields. Record the expected IDs, processed IDs, and any failed or deferred rows. Choose batch size after checking the actual file and available processing capacity. Consistency matters across batches. If “conditional” means one thing in the first group and another in the last, the merged output will hide different standards inside the same column. Freeze the definitions for the run. When you change a rule, identify the earlier rows affected and review them again. You can now ask increasingly precise questions of the same evidence. Show the floral identity terms. Show indoor placements. Show phrases with unresolved material requirements. Show the questions that need an image. Show every candidate waiting on the same missing specification. Use all 15 COSMO relations as a reference vocabulary Amazon’s COSMO paper describes commonsense knowledge generation from shopping behavior. Table 2 publishes 15 relation types. The formal head in its knowledge representation is a shopping behavior; our product-centered keyword map below adapts that vocabulary for seller research.[3] The full set is used_for_func, used_for_eve, used_for_aud, capable_of, used_to, used_as, is_a, used_on, used_in_loc, used_in_body, used_with, used_by, xIntersted_in, xIs_a, and xWant. The unusual spelling of xIntersted_in follows the published table. Its three x-prefixed relations concern the shopper: interest, identity, and desired activity. Keep those shopper-oriented hypotheses distinct from claims about the product. Figure 2. The 15 relation labels and tail domains in COSMO Table 2, arranged into five editorial reading groups with new teaching examples. The atlas shows the individual relations, their tail domains, and teaching examples. Read it as a set of questions you can bring to a keyword bank. Which product concept appears? Which function, setting, activity, audience, or complementary item might be relevant? What would support that interpretation? You will use different parts of the vocabulary for different products. Wall decor gives us useful identity and location examples. A kitchen utensil may offer clearer functional examples. Clothing may introduce body placement, seasons, and activities. A complete reference helps you recognize those opportunities without requiring every product to fill every relation. Keep the canonical relation, the proposed tail, and the evidence in separate fields. The relation names the connection. The tail records the concept at the other end. Confidence and review status explain how far the analysis has progressed. For our flower decoration, a product-centered teaching interpretation might be: Flower wall decoration → used_for_func → add dimensional floral detail to a wall That purpose follows plausibly from the described form. “Improves your mood” would require a different level of evidence. A useful relation makes the product easier to explain while keeping the claim proportionate. Material, color, dimensions, and motif belong in separate attribute fields. The 15-relation vocabulary helps organize contextual meaning; the attribute checks preserve the physical requirements inside the phrase. Let one keyword belong to several useful views Consider “small bathroom wall decor.” The phrase includes a product concept, a location, and an ambiguous size modifier. “Small” could describe the item or the bathroom. The exact phrase alone leaves that unresolved. Claude can record “wall decor” as the category and “bathroom” as the candidate location, while separately flagging the size ambiguity and the need for environment-suitability evidence. The row can then appear in a category view, a bathroom view, a size-review view, and an evidence-request view. Each view answers a different research question. The source row remains one observation. That distinction matters when you summarize the bank. If a phrase appears in three groups, its search-volume observation should still be counted once in a deduplicated overall total. Group totals can overlap. Label them accordingly. Keep different sources, dates, and metric definitions separate until you have a defensible comparison rule. The saved keyword metrics also need that care. The transcript includes returned fields named searchVol, cps, and priority. Preserve their labels and values. An unexplained score should remain unexplained until its definition can be verified; assigning it a conversion meaning would add information the source did not establish. Here is how several displayed phrases change the research decision: “Metal flower wall art” offers a close identity candidate. Confirm the floral form and disclose the MDF-plus-metal construction “Hallway wall decor” suggests a placement question. Dimensions, projection, mounting, and available wall space determine fit “Small bathroom wall decor” adds two checks: the meaning of “small” and the product’s suitability for that environment “Outdoor metal wall decor” creates a specific evidence request. Keep outdoor-use claims pending until the exact product’s suitability is established “Wrought iron wall decor” introduces a material requirement. Evidence of metal components alone does not establish wrought iron Figure 3. Nine selected rows from the 30 displayed historical keywords show overlapping literal phrase dimensions and candidate location tails. SV preserves the historical provider-reported search-volume estimates; 255 phrases were reported in the original pool. At the end of this pass, every decision should have an inspectable reason. Retain a phrase as a research candidate, hold it while a fact is unresolved, or exclude it from the approved shortlist when the product fails a requirement. Preserve all three groups so a later reviewer can follow the choices. Retaining a phrase for research does not approve every claim it might imply. Keep phrase status and claim approval as separate fields. Build the mission around the requirements A location label gives us a place. A mission adds the task and the conditions that determine whether the product can serve it. Take “hallway wall decor.” A possible mission is to add a floral accent to an indoor wall. The next questions become practical: How much space is available? How far does the piece project? What mounting method does it require? How does the finish look beside the existing room colors? Those questions lead to a content plan with substance. A dimension image helps judge scale. A side view shows projection. A mounting explanation addresses installation. A construction close-up clarifies the mixed materials. These are generated planning questions. The keyword bank gives us the language to investigate; the proposed mission remains a hypothesis until supporting customer evidence or a shopper’s own request establishes it. You can also examine how requirements change across a possible search journey. “Metal wall decor” to “metal flower wall art” illustrates Specification by adding floral design. Reversing the sequence illustrates Generalization toward a broader category. Both phrases appeared in the saved bank; their sequence here is modeled. Now add “outdoor metal wall decor.” The new setting creates an eligibility check. A relevant phrase can reveal an opportunity for another product in the assortment, or a specification the current product still needs to establish. The research has made progress even when that phrase stays out of the approved copy. Carry the known requirements through the journey. Broadening a query does not tell us that a shopper has abandoned their available wall space, installation limits, or environment. Ask Claude for views the team can act on Once the rows are structured, ask Claude for outputs that a writer, designer, or product manager can use: Identity shortlist: the strongest relevant category and design phrases, with the exact product facts supporting them Requirement map: the materials, dimensions, quantities, environments, and other conditions expressed across the bank Mission map: proposed shopping tasks linked to retrieved phrases, with inferred requirements clearly labeled Evidence queue: missing facts grouped by the decision they block, so one verified specification can resolve several rows Content opportunity map: questions the product can answer, the existing page coverage, and the most useful place to add the answer Look for shared requirements across phrases. Several location terms may lead to the same dimension image. Several material terms may lead to one precise construction explanation. That is an opportunity to make a single content element useful across multiple searches. Prioritize after the relevance and evidence review. Compare compatible demand metrics within the eligible candidate set. Where available, examine actual business contribution separately from broader market demand. A broad phrase with substantial volume and a precise phrase with purchase evidence may deserve different roles. Avoid forcing the entire decision into one opaque score. Show the phrase, its demand evidence, the product fit, the unresolved requirement, and the proposed action. A reviewer should be able to understand why it matters. Figure 4. A proposed inspectable record for WD09 separates the bathroom relation candidate from the ambiguous small modifier, then routes both to product-evidence review. Work two rows into one useful image brief WD09, “small bathroom wall decor,” and WD20, “hallway wall decor,” share a practical information need: scale. The saved phrases carry searchVol values of 690 and 1,852 respectively. Those historical values help identify the source rows; the combined total would not tell us how many shoppers need a dimension image. In the companion teaching workbook, WD09 is Hold because bathroom suitability and the meaning of “small” remain unresolved. WD20 is Retain as an indoor-placement research candidate, with actual fit still dependent on dimensions and installation details. Create one evidence-queue item linked to both IDs: verify overall width, height, projection, and mounting requirements for the exact variant. When those facts arrive, update the shared product record and both row annotations. The hallway content brief can now use the verified measurements. The bathroom row still needs environment guidance and clarification of the size requirement. That gives the designer a specific brief: show the full product straight on with verified width and height, add a side profile for projection, and identify the documented mounting method. Use the correct pack quantity. Keep the setting neutral while environmental eligibility remains open. Insert actual verified values at production; none are supplied by this teaching example. One image can answer part of two search families. The remaining bathroom question stays in the evidence queue. That is the advantage of joining the bank to a shared product record: useful facts travel across relevant rows, and unresolved claims remain visible. Give each content field a distinct job The September session ended with 21 proposed keyword placements across a title, five bullets, and a description. That allocation was a planning result. It did not establish a universal placement formula or a measured ranking improvement. For a new plan, begin with the information the shopper needs. Product identity belongs where the item must be understood quickly. Construction belongs with a clear explanation of the materials. Scale belongs with verified measurements and a useful visual. Installation deserves instructions specific enough to guide a purchase decision. Selected phrases can travel with those explanations when they fit naturally. This gives a writer a reason for using a term and gives the reviewer a way to inspect the claim. Before drafting, ask four questions of each proposed field: What should the shopper understand here? Which retrieved phrase naturally fits that explanation? Which product evidence supports it? What new information does this field contribute to the page? The resulting brief connects demand, meaning, evidence, and execution. You can trace a title phrase to its source row, a bullet to the requirement it answers, and an image to the decision it helps the shopper make. A practical prompt for your next keyword bank Use this after the product truth card and source data are available: “Organize this Helium 10 keyword bank for [ASIN] in [MARKETPLACE]. Begin by checking the files, columns, row counts, and available processing tools. Tell me what you can inspect and what remains unavailable. Preserve an untouched source table. Assign stable source-row IDs. Keep exact phrases, original metrics and labels, source, marketplace, date or period when supplied, and query settings. Create separate normalization fields. Flag duplicate observations and retain their provenance. Use deterministic operations for counting, normalization, joins, and validation where tools support them. Create manageable semantic-analysis batches with a manifest of expected and completed IDs. Keep the same product truth card, definitions, and output schema across batches. For each phrase, identify category and attribute requirements. Apply relevant COSMO-inspired relations from the supplied 15-relation reference, with candidate tails and short reasons. Allow multiple relations when useful. Keep unsupported, ambiguous, and inapplicable mappings visible rather than filling every relation. Record product evidence, unresolved facts, interpretation confidence, and Retain, Hold, or Exclude status. Keep claim approval separate from research relevance. Label inferred missions and generated buyer questions. Preserve original source wording alongside all analysis. Build an identity shortlist, requirement map, mission map, evidence queue, and content opportunity map. Cross-reference them by source-row ID. Use compatible metrics, disclose overlapping groups, and prevent duplicate counting within each reported total. Keep generated query ideas outside retrieved demand totals. Validate source-to-output coverage, unchanged original values, unique IDs, allowed relation labels, duplicate handling, and the evidence behind claims. Review ambiguous rows and a sample from every batch. Report unresolved coverage and classification issues. Finish with a proposed content plan showing the shopper question, supporting product fact, chosen phrase, and information job for each field or image. Return the plan for review before drafting or publishing.” Make the bank useful enough to write from A useful keyword bank should help you make the next product-page decision. Which phrase describes the item accurately? Which setting deserves an example? Which claim needs evidence? Which requirement points toward a different product in the assortment? Once the rows are connected, several queries may lead to one useful content improvement. Each recommendation remains traceable to the phrase, the interpretation, and the supporting fact. Helium 10 brings the keyword research into the conversation. Claude gives you a way to work across that research with a consistent structure. COSMO’s relation vocabulary adds questions about meaning, while the product record determines what the answers can support. Start with one product. Build the bank. Preserve the evidence. Then ask what each relevant phrase requires the product page to explain. Sources and case note [1] Helium 10: Getting Started with Helium 10 MCP [2] Claude Help Center: Upload files to Claude [3] Amazon Science: COSMO, Table 2 and Section 3 Historical counts and displayed keyword phrases come from the saved September 9, 2026 wall-decor research conversation. The companion workbook annotates the 30 displayed phrases. The relations, missions, crosscuts, and large-bank workflow in this article are newly authored teaching applications. No new Claude analysis of the reported 255-phrase bank was performed. These applications are not Amazon-issued classifications, native Helium 10 COSMO scores, or statements of Amazon ranking weights. Current setup references checked October 7, 2026. Become a Top E-Commerce Brand Sign up now to access powerful, easy-to-use solutions to help with every part of selling on Amazon, TikTok, and Walmart. Sign Up Today Andrew Bell , Amazon Lead & AI Expert Andrew Bell is an AI expert, Forbes-featured strategist, co-author of Rufus: The Blueprint, and creator of the most popular Custom GPTs for Amazon sellers. Published in: AmazonAmazon Seller TipsBlogKeyword Research Share: URL copied Share: Published in: AmazonAmazon Seller TipsBlogKeyword Research Thought Leadership, Tips, and Tricks Never miss insights into the Amazon selling space by signing up for our email list! Subscribe Achieve More Results in Less Time Accelerate the Growth of Your Business, Brand or Agency Maximize your results and drive success faster with Helium 10’s full suite of Amazon and Walmart solutions. Get Started