📋 Live API Reference: Field names and parameters may change. For MCP calls, refresh and inspect the current ZooData MCP inputSchema and outputSchema; for direct HTTP calls, also consult the OpenAPI spec at https://zoodata.ai/api/v1/openapi-spec. Keyword compatibility field: all eleven current keyword and product-traffic request schemas retain granularity, but the only supported value is week. The bundled CLI always sends granularity=week. Never send day, month, lately_day, or legacy lookbackDays; use the weekly period boundaries returned by the service.

ZooData — Commerce Data Infrastructure for AI Agents

200M+ Amazon products. 36 Amazon and keyword-intelligence endpoints. One API key.

Quick Start

1. Get key: zoodata.ai/api-keys (1,000 free credits)
2. export ZOODATA_API_KEY='hms_live_xxx'
3. Base URL: https://api.zoodata.ai/openapi/v2 — all POST with JSON body
4. Auth: Authorization: Bearer YOUR_API_KEY
5. New keys need 3-5s to activate. If 403, wait and retry.

Capabilities & Data Flow

• Network: only https://api.zoodata.ai (Bearer ZOODATA_API_KEY). Setting ZOODATA_BASE_URL to an untrusted host (anything other than api.zoodata.ai / *.zoodata.ai / localhost) makes the CLI refuse the request and withhold the key — the Bearer token is never sent to an untrusted host.
• Execution: bundled shared ZooData CLI {skill_base_dir}/scripts/zoodata.py (Python 3, stdlib-only). This data-layer reference skill allows the complete literal subcommand surface exposed by the bundled client's current top-level help.
• Local files: none by default; reads the optional credential store ~/.zoodata/config.json; the Local Review Toolkit uses a private temporary working dir (created with mktemp -d, removed when the fallback completes) during the review fallback.
• Sent to the API: keywords, category paths, ASINs, marketplace/date and numeric filter values only. Never sent: budget, experience level, risk tolerance, or any other user-profile text — profile inputs map client-side to numeric filters.
• Credits: every API call consumes account credits. For broad or ambiguous requests, state the estimated credit cost and confirm with the user before running multi-call scans.

Shared CLI contract

Before selecting or invoking a bundled CLI command, read and apply references/cli-contract.md; reapply it after every result. It is the local source of truth for invocation, command identity, execution-environment permission handling, composite reuse, exit-status handling, authoritative transport status, retries, terminal interface failures, and partial results.

For market endpoint schemas and quirks, load references/openapi-reference.md § 2; that file owns their request and response contracts. references/reference.md is a summary index and links to the owner instead of restating the market schemas.

Local Interface Failure Output

For this API-reference skill, a terminal interface failure must produce one concise localized notice stating that the ZooData API lookup could not be completed, followed by the succeeded and failed endpoint identifiers. Do not continue into endpoint guidance, schema interpretation, or another API call. Do not expose control tokens or internal retry logs unless the user requests diagnostics.

⚠️ Critical API Pitfalls (ALL skills must follow)

1. Commerce product search using a broad query → resolve and lock categoryPath before interpreting category-sensitive product, competitor, brand, or price-band results. An explicitly labeled products/search category probe may run without a locked category only to resolve that category. Market discovery uses markets/search filters and returns category IDs; market snapshot, distribution, and history requests use a resolved categoryId. Do not apply this rule to /openapi/v2/keywords/* Keyword Intelligence endpoints: their keyword / query inputs are Amazon search queries and do not require categoryPath.
2. Brand/price-band queries MUST include --category to avoid cross-category contamination
3. Market revenue: use the full-category or selected-sample field identified in references/openapi-reference.md § 2; do not calculate revenue from price × sales.
4. Sales = monthlySalesFloor (lower bound). Fallback: 300,000 / BSR^0.65, tag as 🔍
5. Use API fields directly: sampleOpportunityIndex, sampleTop10BrandSalesRate — never reinvent
6. reviews/analysis needs 50+ reviews. Fallback chain when sample is insufficient:
   1. Lightweight: realtime/product → ratingBreakdown (star distribution only, no themes)
   2. Full 11-dim insights: realtime/reviews (raw text, up to 100) + local Map/Reduce via the Local Review Toolkit below — see "Local Review Toolkit" section
7. Aggregation endpoints (price-band, brand) without categoryPath produce severely distorted data
8. Price-band and brand endpoints only accept keyword (not categoryPath) — cross-validate returned products
9. mode is CLI-local, NOT an API parameter → zoodata.py expands --mode client-side into the filter sets in PRODUCT_MODES ({skill_base_dir}/scripts/zoodata.py, 13 presets) before the request; sending mode raw → 422
10. CLI filter flags ≠ API field names → --sales-min → monthlySalesMin; --ratings-max (review count) → ratingCountMax, not ratingMax (a different valid field — max star rating — that returns wrong results silently, no 422). Pass categoryPath as a JSON array (["Electronics"]), never a string. Unknown fields (salesMin, ratingsMax, …) → 422

On Missing Key (no credentials configured)

BEFORE calling any endpoint, verify credentials are configured. The reliable check is python {skill_base_dir}/scripts/zoodata.py check — credentials-only by default, no endpoint calls and no credit usage; exits non-zero if no key is found in env vars OR config files. A [ -z "$ZOODATA_API_KEY" ] test alone is NOT sufficient — a user may have only ~/.zoodata/config.json set.

When no key is found through any mechanism:

1. STOP. Do not run the workflow. Do not call zoodata.py (you'll just get the same credential error and burn tokens).
2. Do NOT fall back to a "partial analysis from training data" / "industry common-sense headlines" / "for reference only" preview. Your training data is stale, has no per-ASIN granularity, and presenting it as analysis — even disclaimed — misrepresents what this skill produces. The deliverable is data-backed; without data, there is no deliverable.
3. Tell the user, in their language, all three of:
   • "ZOODATA_API_KEY is not set — I need this to run the analysis."
   • Get a free key (1,000 credits, no credit card): https://zoodata.ai/en/api-keys
   • Configure via one of:
     • export ZOODATA_API_KEY='hms_live_xxx' (session only)
     • mkdir -p ~/.zoodata && chmod 700 ~/.zoodata && (umask 077; echo '{"api_key":"hms_live_xxx"}' > ~/.zoodata/config.json) (persistent; keep the file private — 0600)
4. Optionally state in one sentence what the workflow will produce once the key is configured (deliverable shape only — no numbers, no market color, no "common sense" preview).

On 401 Invalid Key

When zoodata.py returns a structured error with _transport.status=401, apply this route regardless of whether the preserved server error object contains status or uses the CLI fallback message:

1. STOP further endpoint calls immediately. Do not retry — a rejected key won't be accepted on a second try; every subsequent call will return 401 too.
2. Keep the selected credential authoritative. Do not inspect, compare, export, or switch to a lower-priority legacy credential after rejection. A legacy credential may be selected only when neither new source is configured; trying another endpoint or asking to continue does not change this precedence.
3. Report to the user:
   • The selected ZooData credential was rejected (likely invalid, revoked, or expired)
   • If any partial findings were collected before the failure, show them and mark as partial
   • Fix at https://zoodata.ai/en/api-keys (verify the key, regenerate if needed)
4. Do not fabricate or guess the data the failed calls would have returned. This includes "training-data fallback" / "industry common-sense" headlines disguised as preview — those are fabrications.

On 402 Credit Exhausted

When zoodata.py returns a structured error with _transport.status=402, apply this route regardless of whether the preserved server error object contains status or uses the CLI fallback message:

1. STOP further endpoint calls immediately. Do not retry. Do not switch endpoints as a workaround — 402 is account-level (key/subscription), not endpoint-level.
2. Report to the user with all four of:
   • Which step in the workflow was reached (e.g. "Completed step 3/5: brand analysis")
   • Partial findings already collected (show the actual data, not just a list of completed steps)
   • Returned credit metadata when available; if it is absent, say it was not returned rather than estimating it
   • Top-up link: https://zoodata.ai/en/pricing
3. Do not fabricate or guess the missing data to "complete" the report. Mark partial findings explicitly as partial. No "training-data fallback" / "industry common-sense" filler — substituting public-knowledge prose for missing endpoint data is still fabrication.

On 422 Validation Error

For every parsed HTTP response from zoodata.py, treat _transport.status as the authoritative outer status; response-body and nested status-like fields do not override it. When the CLI returns HTTP 422 / VALIDATION_ERROR, read the preserved structured server error on stdout, including its message/details and _query.params. Do not retry the unchanged request. Correct the named fields first; the CLI exits non-zero while preserving the server error fields for the calling agent. Current keyword requests retain granularity for compatibility but accept only week; legacy lookbackDays remains unsupported.

36 Amazon and Keyword Endpoints

Known Quirks

• Market request quirks, including the live sampleType values and retired parameters, are owned by references/openapi-reference.md § 2.
• listingAge is a string enum (30d, 90d, 180d, 1y, 2y).
• Many search/list endpoints return .data as an array — use .data[0] for the first record. But some commands may return non-array payloads inside data, so inspect the actual response shape before indexing.
• ratingCount not reviewCount everywhere
• bsr (int) in products vs bestsellersRank (array) in realtime
• buyboxWinner.price — NOT top-level price in realtime
• realtime/product does NOT return: monthlySalesFloor, fbaFee, sellerCount
• realtime/product cold-start: first call for an uncached ASIN may return success: true with an EMPTY data (asin: "") while the live fetch warms up — retry once after a few seconds before concluding "no data" (still billed 1 credit per call)
• reviewCountMin/Max filters currently broken (API-56)
• reviews/analysis may 500 for certain ASINs (API-58) — retry different ASIN
• Rate limits: follow the limit exposed by the service for the active request; a client pacing default is not the service quota. The shared references/cli-contract.md owns automatic rate-limit recovery for bundled CLI calls.
• categories uses categoryKeyword (not keyword) and parentCategoryPath (not parentCategoryName)
• reviews/analysis: mode required ("asin"/"category"), use asins (plural array) not asin
• realtime/reviews: returns 10 reviews/page fixed (no pageSize param); 1 credit/page; cursor-paginated; hard cap = 100 reviews (10 pages); supports marketplace US/UK only
• keywords/detail accepts exactly one of keyword / keywords[] (max 20), resolves date to the nearest available weekly snapshot, and returns input-ordered data.items[]; an unmatched item has status=empty, not top-level data: null
• keywords/market-profile accepts one of keyword / keywords[] (max 20), requires date, and supports weekly granularity only. The published compact MCP contract differs from the legacy HTTP shape; load references/openapi-reference.md § 12b before imposing score or availability fields. A subject-specific calculation failure can return HTTP 500 for the whole batch.
• keywords/trend-profile accepts one of keyword / keywords[] (max 20), requires date and 1–4 unique windowPeriods selected from 4/8/12/26, and supports weekly granularity only.
• keywords/extends requires query (not keyword), uses the latest available weekly snapshot, supports queryType = phrase or fuzzy, and may legitimately return empty data.rows[]; the MCP schema has no date parameter, and the CLI has no --date option for this command
• All eleven current keyword and product-traffic request schemas retain granularity for compatibility and accept only week; all currently support only marketplace US. The bundled CLI sends both values explicitly. Never send another granularity or legacy lookbackDays. Use returned weekly period boundaries instead of inferring a rolling window.
• Batch keyword fields on detail, market-profile, trend, trend-profile, and product-traffic-terms-trend must already equal LOWER(TRIM(value)); the bundled CLI normalizes them. Single-keyword fields accept surrounding whitespace and letter case where the endpoint schema says so.
• Keyword endpoints are keyword-query workflows; for inputs named keyword or query, use the Amazon search query / keyword phrase being analyzed
• For keyword endpoints that require date or dateTo, prefer T-1 or earlier and avoid the current date unless the user explicitly asks for today's lookup
• keywords/search-results requires date + keyword; exploreTypes values are ORG, SP, SB, SBV, SPR
• Competitor traffic-term lookup is consolidated into keywords/product-traffic-terms; route retired-interface behavior and the exact request contract to references/openapi-reference.md § 16.
• keywords/product-traffic-structure-profile compares the resolved current week with the previous week; it is not a multi-week trend. See references/openapi-reference.md § 17 for its exact contract.
• keywords/product-traffic-trend is ASIN-level across all keywords and has no keyword dimension; use product-traffic-terms-trend for an ASIN × keyword series.
• keywords/product-traffic-trend-profile provides the server-calculated ASIN-wide trend profile; route exact window, detail, and billing questions to references/openapi-reference.md § 20.
• keywords/search-results is the default source for explaining what products currently appear on a keyword SERP because it already returns listing-level product fields
• products/search is a broader ZooData product-database query and must not be presented as Amazon live keyword SERP ordering

Keyword Intelligence Endpoints

These eleven endpoints fill the gap between raw catalog data and search-demand/search-visibility intelligence.

Keyword value boundary:

• Keyword endpoints provide estimated search, visibility, rank, traffic-share, and impression-point signals
• They do not provide a seller's first-party ABA Search Query Performance funnel by themselves
• Treat keyword value, profitability, and conversion potential as directional unless the user supplies ABA-SQP impressions, clicks, cart adds, purchases, click share, purchase share, and conversion rate
• Seller-artifact acquisition, stage selection, field interpretation, and user-facing output policy belong to the amazon-keyword-traffic-analysis skill. This API reference does not prescribe a blanket caveat or one seller view for every subject.

/openapi/v2/keywords/detail

• Input: exactly one of keyword / keywords[] (1–20), required date, optional marketplace; compatibility-retained granularity supports only week
• Data window: resolves the requested date to the nearest available weekly snapshot at or before that date
• Date rule: prefer T-1 or earlier for date; avoid current-date lookup unless explicitly requested
• Response shape: data.context + data.items[], preserving request order
• Item fields: identity, status=ok|empty, snapshotData, emptyReason, nullable errorCode, nullable errorMessage
• snapshotData fields include estimateSearchCount, abaRank, Top3 click/conversion shares, marketCharacteristics, totalSkuCount, SKU/brand/title coverage, organic/ad counts, and Top48 benchmarks
• Do not read legacy estimateSearchCountWeekly, totalSkuCnt, or top-level data:null

/openapi/v2/keywords/market-profile (metric layer)

• Availability: standard production endpoint under the documented base URL
• Input: exactly one of keyword or keywords[] (1–20), required date, optional marketplace; compatibility-retained granularity supports only week
• Response shape: data.context + data.items[], preserving request order
• Read references/openapi-reference.md § 12b for the compact MCP and legacy HTTP response contracts, field identity, availability, and billing. Compact profiles return business evidence without normalized scores; do not require legacy scoringSpec, item status, or support flags. Legacy score-bearing results remain interpreted under their own returned scoring context.
• A subject-specific calculation failure can currently produce HTTP 500 for the whole batch. Treat it as a service failure, not an item-level empty result; do not automatically fan out all subjects into single calls.
• Three-layer boundary: use data-layer keywords/detail for source snapshot fields, metric-layer keywords/market-profile for stable deterministic profile objects, and the Agent + skill layer for evidence composition, confidence, explanations, limitations, and actions
• Metric-first access: call the matching metric before its source data endpoint. Descend only when the Agent needs an indicator or evidence grain omitted by the metric contract, the metric endpoint is unavailable and transparent data-based calculation is valid, no metric exists, or raw evidence is explicitly requested. Incomplete metric calculation coverage is a conclusion limit—not by itself a reason to call same-source data.
• Batch-first execution: after selecting the endpoint, collect all subjects with identical non-subject context and prefer its batch contract over repeated single calls. Deduplicate case-insensitively, preserve order, chunk compatible sets at the endpoint limit (20 for current keyword batches), and merge results back into global input order. Batch support never justifies an extra cross-layer call.

/openapi/v2/keywords/trend

• Input: exactly one of keyword / keywords[] (1–20), required dateFrom / dateTo, optional marketplace; compatibility-retained granularity supports only week; maximum 93-day range
• Data window: weekly-granularity points across the requested date range
• Date rule: prefer T-1 or earlier for dateTo; avoid current-date lookup unless explicitly requested
• Response shape: data.context + data.items[].series[], preserving request order
• Item fields: identity, status=ok|empty, series[], emptyReason, nullable errorCode, nullable errorMessage
• Series fields: periodStartDate, periodEndDate, estimateSearchCount, abaRank, abaTop3ClickShareRate, abaTop3ConversionShareRate

/openapi/v2/keywords/trend-profile (metric layer)

• Input: exactly one of keyword / keywords[] (1–20), required date, required unique windowPeriods[] selected from 4/8/12/26, optional marketplace; compatibility-retained granularity supports only week
• Response: data.context + data.items[].rows[]; every requested window returns one row with rowContext, status=ok|empty, emptyReason, and trendProfile
• Available profiles contain independently guarded searchDemand and abaRank dimensions with trend and {value,direction} entries under trendEvidence. trendPattern is not required; search demand can also return window-specific volatilityLevel and lastPeriodWindowPosition.
• Evidence includes first/last/change values, normalized slope, direction consistency, aligned/eligible period counts, plus demand volatility/window position or ABA best/worst rank
• Use this metric endpoint before raw keywords/trend for trend-shape and volatility judgments. Descend only for required weekly points or fields omitted from the profile.
• Preserve null empty reasons rather than inventing one. Billing is per keyword with at least one status=ok window; use returned credit metadata.

/openapi/v2/keywords/extends

• Input: required query; optional marketplace, page, pageSize, queryType, sortBy, sortOrder; compatibility-retained granularity supports only week; no date is required
• Important quirk: seed field is query, not keyword; queryType supports phrase and fuzzy
• Data window: latest available weekly snapshot; the MCP schema has no date parameter, and the CLI has no --date option for this command
• Response shape: data.context + data.query + data.queryType + data.rows[]
• Row fields: matchData.{query,keyword,site,relevanceScore} and keywordSnapshot, whose dataWindow.currentPeriod and snapshot metrics use the same current field families as keywords/detail
• Do not flatten rows to legacy term, seedKeyword, or estimateSearchCountWeekly; empty rows[] is normal

/openapi/v2/keywords/search-results

• Input: required keyword / date; optional marketplace, page, pageSize, exploreTypes, sortBy, sortOrder
• Compatibility-retained granularity supports only week; do not send legacy lookbackDays
• Data window: latest available weekly period at or before the requested date; use the returned period boundaries
• Date rule: prefer T-1 or earlier for date; avoid current-date lookup unless explicitly requested
• Response shape: data.context + data.identity + data.rows[]
• Row fields include latestObservedAt, exploreType, absolutePosition, pageIndex, pagePosition, asin, title, brand, price, currency, link, imageLink, rating, ratingCount, recentSales, hasVideo, estimateImpressionPoint, keywordTotalEstimateImpressionPoint
• Interpretation rule: use this endpoint first for "what is on page 1 / what products dominate this keyword / what does the SERP look like"
• Do not substitute products/search when the question is about observed keyword SERP composition or ordering

/openapi/v2/keywords/product-traffic-terms

• Provides traffic-driving keyword rows for any target ASIN, including competitor research; the former competitor-specific lookup is consolidated into this route.
• Read references/openapi-reference.md § 16 for the request, response, field, date, filtering, pagination, retirement, and billing contract.
• Apply references/cli-contract.md to every result.

/openapi/v2/keywords/product-traffic-structure-profile

• Production supports one ASIN or a batch of up to 20 ASINs, retains granularity with week as its only supported value, and compares the resolved current week with the previous week.
• Read references/openapi-reference.md § 17 for the request, response, status, field, date, batching, and billing contract.
• Apply references/cli-contract.md to every result, including a server-provided endpoint migration response.

/openapi/v2/keywords/product-traffic-terms-trend

• Provides weekly product-side traffic, placement, keyword-context, and product-observation history for one ASIN and its named keyword subjects.
• Read references/openapi-reference.md § 18 for the request, response, status, field, date, batching, range, and billing contract.
• Apply references/cli-contract.md to every result.

/openapi/v2/keywords/product-traffic-trend

• Provides ASIN-wide raw weekly traffic and term-coverage history across all observed keywords; it is the weekly-detail companion to the four-week metric profile.
• Read references/openapi-reference.md § 19 for the request, response, status, field, date, batching, and billing contract.
• Apply references/cli-contract.md to every result.

/openapi/v2/keywords/product-traffic-trend-profile

• Provides the server-calculated ASIN-wide four-week trend profile; use the raw trend endpoint when weekly points are required.
• Read references/openapi-reference.md § 20 for the request, response, status, field, date, batching, and billing contract.
• Apply references/cli-contract.md to every result.

Local Review Toolkit

When /reviews/analysis lacks aggregation (ASIN has <50 reviews or no daily snapshot), fall back to live raw reviews + your own LLM. The toolkit does NOT call any external LLM — you (the calling skill's LLM) perform the Map/Reduce steps.

Workflow:

# 1. Fetch raw reviews (up to 100, cursor-paginated, ~60s, 10 credits at full)
zoodata.py reviews-raw --asin B0XXXXXXXX [--marketplace US] [--max-pages 10]

# 2. For EACH review, render the per-review Map prompt
zoodata.py review-tag-prompt --review '<single review JSON>' \
    [--product-title "..."] [--product-category "..."]
# → Your LLM produces a JSON object with sentiment + 11 dimension arrays
#   (mentioned_scenarios, mentioned_issues, mentioned_positives, mentioned_improvements,
#    mentioned_buying_factors, mentioned_pain_points, user_profiles, mentioned_usage_times,
#    mentioned_usage_locations, mentioned_behaviors, keywords)
# Suggested map parallelism: ~20 concurrent if your LLM supports it

# 3. Collect candidate phrases per dimension. For EACH dimension render the Reduce prompt
zoodata.py review-reduce-prompt --label-type positives \
    --candidates '["comfortable","comfy","very comfortable",...]'
# → Your LLM produces {clusters: [{canonical, members}, ...]}
# Suggested chunk size for `keywords` dim when >150 candidates: 150 per call

# 4. Aggregate into reviews/analysis-compatible consumerInsights
zoodata.py review-aggregate --reviews raw.json --tagged tags.json --clusters clusters.json
# → Output shape matches /reviews/analysis: reviewCount, avgRating,
#   sentimentDistribution, consumerInsights[], topKeywords[]

When to use the toolkit instead of reviews/analysis:

• ASIN has fewer than 50 reviews
• reviews/analysis returns sparse consumerInsights (missing dimensions)
• Need the freshest possible data (Spider scrape vs. T+1 BigQuery snapshot)
• Need to analyze a brand-new product that has no daily snapshot yet

Field Differences Across Endpoints

Load references/reference.md § Field Differences Across Endpoints for the cross-endpoint field map.

Confidence Labels (all skills)

• 📊 Data-backed — direct API data
• 🔍 Inferred — logical reasoning from data
• 💡 Directional — suggestions, predictions

Strategy recommendations and subjective conclusions are NEVER 📊. Extreme growth (>200%) = 💡 only.

Data Notes

• Sales (monthlySalesFloor) = lower-bound estimate
• Realtime = live; products/competitors = ~T+1 delay
• Marketplace coverage varies by endpoint; follow each endpoint schema
• Each call consumes credits; check meta.creditsConsumed

Links

• zoodata.ai · API Docs · GitHub · support@zoodata.ai