US Patent 12,561,383 · A9 Search · Technical Analysis & Seller Playbook

How Amazon Patent 12,561,383 Reveals the Architecture Behind Search Bar Query Recommendation

A complete technical analysis, behavioral taxonomy, and seller optimization playbook — built on the A9 patent that documents the dual-pipeline system classifying shopper intent and generating the next recommended query before anyone types a letter.

Filed by A9 · Assigned to Amazon Technologies · Granted Feb 24, 2026

Specification

Progressive narrowing

Generalization

Broadening after friction

Equivalence

Known-item retrieval

Substitution

Competitive switching

Complement

Cross-sell & adjacency

Irrelevant

Session reset event

Intent Types

Pipeline Stages

Walkthroughs

Metadata Layers

Archetypes

Scorecard Max

The Two-Layer Thesis

For five years, most sellers have optimized Amazon listings one way. Keywords in titles, bullets, and backend fields. A9 rewarded relevance through lexical matching, and that model worked. Past tense.

Amazon's search engine appears to operate on two layers: keyword retrieval and intent prediction. Patent 12,561,383 documents that second layer — a system that classifies shopper behavior and predicts what they will search next, built on top of keyword matching infrastructure that still drives product discovery. Keywords remain foundational. But a second system now operates on top of keywords, and most sellers have never encountered documentation of how that system works.

Filed by A9 and assigned to Amazon Technologies, patent 12,561,383 documents a dual-model machine learning system that classifies customer intent from search query sequences and generates predicted next-queries aligned with that intent. A system here doesn't wait for a customer to type. Prediction happens first. Results get precomputed and cached for near-instant delivery.

What does the patent actually cover? Not chatbots. Not conversational AI. It covers a search bar — autocomplete suggestions, "related searches," "customers also searched for," and recommended query refinements appearing when millions of shoppers type into Amazon's search box every day. Foundational infrastructure powering predictive search bar recommendations across Amazon's entire marketplace.

Architecture here operates on six exhaustive intent categories: Specification, Generalization, Equivalence, Substitution, Complement, and Irrelevant. Every query reformulation a customer makes gets classified into one of those six types. I call that taxonomy an Intent Hexagon, where every query transition maps to a vertex and every product competes within one or more vertices.

The Headline

Amazon Search now operates on intent prediction in addition to keyword matching. Sellers who optimize only for keywords are missing a critical second layer — the intent classification system that determines which queries the search bar recommends next.

The Dual-Pipeline Architecture

The patent reveals a fundamental architectural decision: Amazon runs two separate machine learning models in sequence. One classifies why a customer changed their search query. The other generates the next query the customer is likely to type — constrained by that classified intent. These are not the same model.

Stage 01

BERT Encoding

Bidirectional query encoding

BERT converts query text into dense vector representations. Self-attention lets every token attend to every other token simultaneously, capturing full bidirectional context in one pass. "Running shoes" and "jogging sneakers" produce similar embeddings. Keywords still matter — but the system now understands meaning, not just tokens.

Stage 02

Intent Determination

Classify the transition

Each pair of consecutive queries gets classified into one of six intent categories via multi-class cross-entropy loss. A softmax probability distribution across the six categories picks the dominant intent for the transition. "Laptops" → "gaming laptops" = Specification. "iPhone 15" → "smartphones" = Generalization.

Stage 03

Intent Prediction

Forecast the next intent

Masked self-attention runs over historical intent sequences with six parallel attention heads — one per intent type. Six simultaneous interpretations of every customer's search history are maintained until generation. Products visible across multiple intent heads gain broader recommendation exposure.

Stage 04

VAE Query Generation

Generate 3–5 novel queries

A Variational Autoencoder captures the statistical distribution of query transitions and samples from it. Three loss functions (reconstruction, KL divergence, uniformity) tune the system to produce 3–5 diverse, relevant, stable query suggestions — including queries the customer would not have generated independently.

FIG. 1 — System architecture showing queries 104 flowing through Encoding Module 118, Intent Determination Module 106, Intent Prediction Module 110, and Query Prediction Module 114 — **FIG. 1** — US 12,561,383 B1. The two-pipeline architecture drawn in the patent itself. Encoding Module **118** produces feature embeddings **120(1–3)** from consecutive queries. Intent Determination Module **106** classifies the transitions. Intent Prediction Module **110** forecasts the next intent. Query Prediction Module **114** generates the output query **116**. Every module name in the writeup maps to a numbered box in this drawing.

Architectural Proof

Intent classification operates on top of retrieval. Keywords drive retrieval. Intent classification drives which queries the search bar suggests next. Sellers optimizing for keywords are doing necessary work — on one layer.

The Intent Hexagon

Every time a customer modifies a search query, the system classifies that modification into one of six exhaustive, mutually exclusive intent categories. Not a spectrum. A taxonomy. Classification operates on the difference between consecutive queries, never on individual queries in isolation. The same keyword carries entirely different intent depending on what preceded it.

Illustration by Andrew Bell · The Intent Hexagon

The Intent Hexagon — my illustration. Not a figure from the patent. This is my own visual model of the six-type taxonomy the patent defines: every reformulation the search bar sees resolves into exactly one of these vertices. The set is exhaustive. The set is mutually exclusive. Classification runs on the difference between consecutive queries, not on the query in isolation. — Andrew Bell

01 · Intent Type

Specification

Progressive narrowing

Shoppers add keywords, introduce brand names, or specify colors and materials. They have converged. Purchase intent runs highest during Specification.

"laptops" → "gaming laptops RTX 4090 16 inch"

Play: Titles balance a core category term with primary and secondary modifiers. Backend keywords in tiered separation — primary (core), secondary (use case), tertiary (attributes). Bullets progressively reveal specification depth.

02 · Intent Type

Generalization

Broadening after friction

Keyword removal signals dissatisfaction. Shoppers who broaden search terms are voting against current results — expanding scope because what they found feels too narrow or irrelevant.

"electric standing desk 60 inch" → "standing desk"

Play: Rank for base category terms. Allocate ~25% of backend keywords to broad category terms without restrictive modifiers. Core category placement in title positions 1–3 matters most.

03 · Intent Type

Equivalence

Known-item retrieval

Highest conversion intent on Amazon. Exact model numbers, brand-specific product names, SKU-level terms. No discovery happening. A retrieval mission.

"Canon EOS R6 Mark II"

Play: Product name, model number, generation, and variant must match exactly what shoppers search. Model codes and product line identifiers function as search anchors. Exact-name traffic routes to the listing that spells it right.

04 · Intent Type

Substitution

Competitive brand switching

Replacing one brand name or characteristic with another. When shoppers enter Substitution mode, the search bar proactively recommends additional alternatives not yet considered.

"Nike marathon running shoes" → "Asics marathon shoes"

Play: Backend keywords benefit from including competitor brand names used descriptively. A+ Content includes comparison modules with non-disparaging feature, price, and quality positioning.

05 · Intent Type

Complement

Cross-sell & adjacency

Phone cases after phones. Chargers after laptops. Keywords related to, but functionally different from, previous results. Products in the Complement window receive highest algorithmic lift immediately after a purchase event on a related primary product.

"Canon EOS R6 Mark II" → "Canon RF 50mm lens"

Play: Backend keywords target complement patterns ("with," "for," "alongside"). Bundling and Sponsored Brands campaigns targeting complement keywords capture the add-to-cart moment.

06 · Intent Type

Irrelevant

Session reset event

A nuclear option. Keywords completely unrelated to prior queries trigger Irrelevant — and the patent language runs precise: the system "may exclude" prior queries from the prediction pipeline entirely. One Irrelevant classification resets the customer's intent context.

"dog food topper probiotics" → "birthday cards funny"

Play: Title and bullet 1 must immediately communicate what the product serves and what boundaries exist. A+ Content explicitly sets product boundaries to self-filter irrelevant traffic and reduce bounce.

Classifications here don't run binary. The system computes a softmax probability vector across all six intent types for every reformulation. A transition might register 60% Specification, 25% Complement, and 15% Generalization simultaneously. Products can match multiple intents at once, and optimizing for a spread of plausible intents hedges visibility across a probability distribution.

Context functions as the variable. Keywords stay constant. "Nike shoes" after "Adidas shoes" registers as Substitution. "Nike shoes" after "shoes" registers as Specification. Identical keyword. Completely different intent.

The Hexagon in Action

Ten walkthroughs of real shopping journeys. Each traces a multi-query session, showing which intent the system assigns at every transition — and what that classification means for sellers competing at each stage.

Walk 01

The Marathon Runner

A broad fitness search progressively narrows, then discovers complementary gear.

running shoes→marathon running shoes

Specification

marathon running shoes→Nike marathon running shoes

Specification

Nike marathon running shoes→Asics marathon shoes

Substitution

Asics marathon shoes→running socks moisture wicking

Complement

Seller takeaway: A running-shoe seller must rank for the broad term, the use-case modifier, and survive brand substitution by appearing when competitors are searched. An accessories seller should target complement queries that follow shoe purchases.

Walk 02

The Home Office Upgrade

A professional narrows into specs, hits a dead end, generalizes, then discovers accessories.

standing desk→standing desk electric 60 inch

Specification

standing desk electric 60 inch→electric standing desk

Generalization

electric standing desk→Uplift desk

Equivalence

Uplift desk→monitor arm for standing desk

Complement

Seller takeaway: Standing-desk sellers need keyword coverage at both specific and general levels. Accessory sellers should optimize for complement queries tied to desk purchases.

Walk 03

The Skincare Journey — Saturation Cliff

Four consecutive Specification queries without conversion trigger the predicted Substitution.

face cream→face cream anti aging

Specification

face cream anti aging→face cream anti aging retinol

Specification

face cream anti aging retinol→retinol cream for women over 50

Specification

retinol cream for women over 50→CeraVe retinol cream

Substitution

Seller takeaway: This is the Saturation Cliff in action. After sustained Specification without conversion, the system's predictions shift toward Substitution and the recommended queries begin surfacing competitors the shopper had not yet considered.

Walk 04

The Camera Enthusiast — Post-Purchase Inversion

Equivalence lookup triggers a three-step Complement chain after purchase.

[browsing]→Canon EOS R6 Mark II

Equivalence

Canon EOS R6 Mark II→Canon RF 50mm f/1.8 lens

Complement

Canon RF 50mm f/1.8 lens→camera bag mirrorless

Complement

camera bag mirrorless→SD card 128GB UHS-II

Complement

Seller takeaway: This is the Post-Purchase Inversion Effect. Once the primary product is purchased, the system flips to Complement predictions. Accessory sellers who optimize for camera-model-specific queries capture the entire chain.

Walk 05

The Budget Kitchen Shopper

Price sensitivity triggers generalization-to-substitution, then ecosystem Complement.

KitchenAid stand mixer→stand mixer

Generalization

stand mixer→Hamilton Beach stand mixer

Substitution

Hamilton Beach stand mixer→stand mixer dough hook attachment

Complement

stand mixer dough hook attachment→kitchen scale digital

Complement

Seller takeaway: When a shopper generalizes away from a premium brand, the system detects price sensitivity. Mid-tier brands win these generalization-to-substitution transitions. Accessory sellers should target the full baking ecosystem.

Walk 06

The Pet Owner — Session Reset

A strong Equivalence session is erased by one Irrelevant classification.

[browsing]→Purina Pro Plan sensitive stomach dog food

Equivalence

Purina Pro Plan sensitive stomach dog food→dog food topper probiotics

Complement

dog food topper probiotics→birthday cards funny

Irrelevant

Seller takeaway: The Irrelevant classification erases all prior session context. The dog-food affinity, the probiotic interest, the health-conscious signal — all zeroed out. Every product must earn its own relevance from scratch after a session reset.

Walk 07

The Fitness Tech Comparison Shopper

A sustained Substitution loop resolves, then opens an accessory Complement chain.

Fitbit fitness tracker→Garmin fitness tracker

Substitution

Garmin fitness tracker→Apple Watch SE fitness

Substitution

Apple Watch SE fitness→Apple Watch SE band replacement

Complement

Apple Watch SE band replacement→screen protector Apple Watch SE

Complement

Seller takeaway: Sustained Substitution signals active comparison. The moment Substitution shifts to Complement, a purchase decision has been made. Accessory sellers targeting the winning brand capture the full post-decision chain.

Walk 08

The Baby Registry Builder

Specification, generalization recovery, then a long Complement chain.

baby crib→baby crib convertible white

Specification

baby crib convertible white→convertible crib

Generalization

convertible crib→crib mattress waterproof

Complement

crib mattress waterproof→crib sheets organic cotton

Complement

Seller takeaway: Baby products exemplify long complement chains. A crib triggers mattress, sheets, bumper, mobile, and monitor in sequence. Sellers who optimize for the full nursery ecosystem capture traffic at every complement transition.

Walk 09

The Material-Conscious Shopper

Substitution on material (not brand) followed by specification and generalization.

stainless steel water bottle→glass water bottle

Substitution

glass water bottle→glass water bottle with silicone sleeve

Specification

glass water bottle with silicone sleeve→BPA free water bottle

Generalization

BPA free water bottle→water bottle brush cleaner

Complement

Seller takeaway: Substitution is not always brand-based. Material substitution is a distinct pattern the system tracks. Sellers of eco-friendly or material-specific products should optimize for material comparison queries.

Walk 10

The Gaming Setup — Full Lifecycle

Deep Specification, Saturation Cliff, Substitution, and a multi-product Complement chain.

laptop→gaming laptop

Specification

gaming laptop→gaming laptop RTX 4090 16 inch

Specification

gaming laptop RTX 4090 16 inch→gaming laptop RTX 4090 16 inch under 2000

Specification

gaming laptop RTX 4090 16in under 2000→Alienware gaming laptop

Substitution

Alienware gaming laptop→laptop cooling pad

Complement

laptop cooling pad→gaming mouse wireless

Complement

Seller takeaway: The full lifecycle of the Intent Hexagon in a single session. The Saturation Cliff is illustrated in the patent itself. The complement chain following purchase can extend across three, four, or more related products. Sellers who understand the sequence can position at the transition points where the system's predictions shift toward their product type.

Behavioral Signals & System Dynamics

Five named dynamics that govern how the system translates shopper behavior into predicted intent and generated queries. Each is documented or strongly implied by patent text.

Dynamic 01

The Post-Purchase Inversion Effect

When a customer purchases or adds to cart, the system significantly increases the predicted probability of Complement intent for the next search. The customer who just bought a phone won't search for more phones — the system begins generating phone-case and accessory queries before the customer types a single character. If you sell accessories, your maximum algorithmic visibility window opens the instant a customer converts on a related primary product. The window is architecturally engineered.

Dynamic 02

The Saturation Cliff

Patent 12,561,383 illustrates a pattern with an example of four Specification queries. After approximately four consecutive Specifications without conversion, probability increases that the customer will switch to Substitution — actively seeking alternatives to everything narrowed toward. Recommended queries begin surfacing competitors. Brands dominating a Specification sequence but failing to convert inside that window face increasing probability that the system predicts Substitution intent next.

Dynamic 03

The Session Reset Event

When the system classifies a reformulation as Irrelevant, all queries before that classification may get excluded from subsequent intent counts. Prior session context may no longer inform predictions. Sellers who lost a customer to an Irrelevant shift cannot rely on session continuity for recovery because the prediction pipeline may operate without benefit of prior behavioral signals.

Dynamic 04

The Behavioral Signal Hierarchy

Patent text explicitly enumerates three tiers of interaction: navigation to a product page, addition to cart or list, and initiation of purchase. No collapsing into a single "engagement" metric. Each tier carries different weight in the intent prediction model. A page view nudges. A cart-add shifts. A purchase inverts.

Dynamic 05

The Personalization Stack

An identical search query produces different intent predictions for different customers. Patent text enumerates five metadata layers enriching every query before encoding: human language indication, geographic characteristics, demographic data, hardware and software components of the device, and network identifiers indicative of geolocation. No single "search result page" exists. Millions exist, personalized at the encoding layer.

Additional Patent Details

Six technical details enumerated in the patent that shape how the system behaves in production.

Multimodal Input

Voice, image, and video are valid inputs

Voice queries convert via speech-to-text. Image searches use recognition to extract text. All modalities reduce to text embeddings before intent analysis. Voice-search optimization is keyword optimization under the hood.

Precomputed Cache

Results are cached before the click

Search results for high-probability predicted queries get precomputed and cached. Products in precomputed result sets benefit from reduced latency — a structural advantage distinct from listing quality.

Discovery Engineering

Queries shoppers wouldn't have typed

Recommended queries may contain terms a customer was unaware of and unlikely to type manually. Amazon actively expands customer vocabulary. Distinctive features, novel materials, and unique attributes can surface even when no one explicitly searches for them.

Delta Analysis

The system reads what changed

Intent is inferred by comparing consecutive queries — what keywords were added, removed, or substituted. The system measures the manner of change as its primary input, not the content of any individual query.

Dual Placement

Search bar and results page

Predicted queries appear in two contexts: inside the search interface (autocomplete) and alongside results pages. Two visibility channels, same prediction model, different placement windows in the shopper's journey.

Session Boundaries

Intent prediction is session-scoped

The sequence of queries is bounded by a session — a temporal grouping of activity. Cross-session context is not guaranteed. A customer returning hours later may start with a clean slate. Session boundaries are prediction boundaries.

Normalization

Counts are both raw and normalized. Whether a customer searches 3 or 30 times, the proportional distribution of intents matters more than raw numbers. Proportionality, not volume.

Post-Transition

The model explicitly learns what typically follows each intent. "After Generalization, Specification follows within 2 queries" is the kind of pattern the model captures. I call this Sequential Intent Positioning — products matching the next move in the choreography get recommended.

Six Heads

Six parallel attention heads run simultaneously, one per intent type. The system doesn't commit to a single interpretation of the journey until the final prediction. Products visible across multiple heads gain broader exposure.

Seller Optimization Playbook

The Intent-Optimized Listing Framework. Four listing surfaces, each addressing one or more intent types. Every edit traces back to a classification the system performs internally.

Title

Balance broad and specific

Leads with core category term (Generalization capture). Follows with primary modifier (Specification capture). Brand name placed where Equivalence traffic rewards recognition. The first 80 characters do most of the ranking work.

Women's Professional Tennis Shoes — Lightweight, Breathable, Court-Optimized

Covers: Generalization ("shoes"), Specification ("women's tennis shoes"), Irrelevant filter ("professional," "court-optimized").

Bullet Points

Structured in intent priority

Bullet 1 filters Irrelevant traffic and confirms fit inside three seconds. Bullets 2–3 carry Specification depth — features backed by specific evidence. Bullet 4 runs Substitution defense — explicit reasons to choose over alternatives. Bullet 5 opens Complement hooks — "pairs well with" ecosystem references.

Backend Keywords

Tiered allocation

Recommended starting mix: ~60% Specification (long-tail modifiers), ~15% Generalization (core category), ~15% Equivalence / Substitution (alternatives and competitor terms used descriptively), ~10% Complement (related-product language). Tune from Search Query Performance data.

A+ Content

Four structured modules

(1) Quick relevance confirmation. (2) Specification details and features. (3) Equivalence / Substitution positioning vs. alternatives — comparison modules. (4) Complement ecosystem showing what pairs. Every module does one job the hexagon names.

Intent Mapping — the Method

A9's classifier lives deep inside Amazon infrastructure. Sellers never see the labels — only downstream effects. Intent Mapping of the Intent Hexagon converts A9's machine-learning taxonomy into a seller-facing lens: six transition types become six coverage dimensions a seller can score, diagnose, and rewrite toward.

Seven Steps, Six Subagents, One Brief

Gather Demand Inputs

Pull a Helium 10 Cerebro or Magnet export. Filter by search-volume threshold. Separate terms into five input categories: core category, phrase variants, modifiers, use cases, and adjacent needs.

Classify Inputs by Intent Type

A Classifier subagent loads reference material on the six types plus worked examples. Every keyword receives a dominant intent label. Ambiguous keywords also receive a secondary label. Output: a classification table ready for journey modeling.

Model Likely Journey Paths

A Journey Modeler produces 3–5 realistic flows specific to the product. Each flow names a starting query, a sequence of transitions, and a likely ending query. Flows describe how a shopper probably moves through a session.

Diagnose Vulnerabilities

A Vulnerability Diagnostician walks the current listing through every modeled journey. Output lists where the listing loses relevance, why, and which page element causes the loss.

Align Page Elements to Journey Stages

Every diagnosed vulnerability is mapped to a specific listing surface — title, bullets, hero image, secondary images, A+ content, backend terms, brand store. No fix floats without an anchor.

Rewrite and Rebuild

A rewrite plan executes edits traceable to identified vulnerabilities. Intent Mapping produces strategy. Execution happens downstream inside whatever copy pipeline the seller runs. Scope boundary stays clean.

Score Journey Coverage

A Scorer applies a rubric across six coverage dimensions. Scale runs 1–10 per dimension, 60 total. Grading bands translate the score into Poor, Developing, Strong, or Optimized.

Six Journey Archetypes

Real shopper journeys repeat in predictable patterns. Six archetypes drawn from my analysis of Helium 10 keyword graphs alongside A9's reformulation patterns.

Gen → Spec → Spec → Equ

Refinement Path

Broad to narrow to known-item. "laptop" → "gaming laptop" → "gaming laptop RTX 4090" → "Alienware m18 R2." A listing prepared for retrieval but not for narrowing loses visibility at step two.

Spec → Gen / Sub

Failed Refinement Path

Specification narrows too far, reverses course, either back to Generalization or sideways to Substitution. Failed refinement kills conversion if the listing cannot reappear at the broader query.

Equ → Comp → Comp

Known Item to Ecosystem

Shopper searches for a known item, commits mentally, begins building around the purchase. "AirPods Pro 2" → "AirPods Pro 2 case" → "Apple USB-C adapter." Brand store architecture drives ecosystem capture.

Spec → Sub → Sub → Comp

Comparison Path

Narrow, reject candidate, compare alternatives, then add a complementary item after selection. A+ content earns decisive weight during the comparison phase.

Spec → Gen → different Spec

Recovery Path

First narrowing path stalls. Shopper backs up and narrows along a different axis. Broad title coverage plus multiple feature axes in bullets let a listing ride recovery paths without disappearing.

Any → Irrelevant

Abandonment Path

Session ends or pivots to an unrelated category. Strong bullet 1 and a clean hero image filter wrong shoppers early. Abandonment drops conversion metrics when a listing fails to filter.

Six Failure Modes

Patent 12,561,383 classifies transitions. Intent Mapping adds a diagnostic layer on top: what happens when a listing cannot survive a given transition. Six failure modes, one per intent type.

F01 · Specification

Failure to Hold Specification

Listing retrieves on broad terms but vanishes as shoppers add modifiers. Title carries core category and no meaningful attributes. Bullets lack progressive feature depth. Images fail to prove attribute claims.

F02 · Generalization

Failure to Capture Generalization

Listing ranks only on long-tail modifiers. Strip a qualifier and the listing disappears. Hero image fails to confirm broad category fit. First half of title over-specializes before establishing category.

F03 · Equivalence

Failure to Win Equivalence

Product sold under a name shoppers do not actually type. Model numbers missing from title. Variant language inconsistent across bullets, images, and A+. Exact-name traffic routes elsewhere.

F04 · Substitution

Failure to Survive Substitution

Comparison shoppers arrive, scan for differentiation, find none, select a competitor. A+ content lacks comparison modules. Bullet 4 repeats features instead of arguing choice rationale.

F05 · Complement

Failure to Benefit from Complement

Adjacent search surfaces exist but the listing does not participate. No ecosystem visibility. Brand store fails to route Complement traffic. A+ comparison modules do not link related SKUs.

F06 · Irrelevant

Failure to Filter Irrelevant Traffic

Listing attracts wrong shoppers. Title ambiguity produces impressions outside category fit. CTR and conversion erode. Organic rank degrades downstream.

A Worked Audit

Theory without a worked example leaves a seller reading instead of operating. Here is Intent Mapping run end-to-end against a real product category — a mid-tier 32 oz insulated water bottle.

FIG. 3 — Worked examples showing a shoes query sequence classifying as Substitution and a cell phone query sequence classifying as Complementary — **FIG. 3** — US 12,561,383 B1. Amazon's own worked examples. **Example A**: a five-query shoe session (*shoes → tennis shoes → athletic shoes → Brand A athletic shoes → blue Brand A athletic shoes*) plus a "Navigation to interface" user interaction produces a **Substitution** prediction — output query *blue Brand B athletic shoes*. **Example B**: three cell-phone queries ending on Brand A Model B, plus a *Purchase Brand A Model B* interaction, produces a **Complementary** prediction — output query *Brand A Model B phone case*. Purchase events flip the predicted intent from Specification toward Complement.

The setup. Mid-tier brand. Stainless steel, double-wall vacuum, leak-proof lid, eight colors. Seller pulls a Cerebro export off a leading competitor ASIN. 1,847 keywords with volume data. After filtering under 300 monthly searches, 412 working terms remain, separated into the five input categories: 38 core, 174 phrase variants, 86 modifiers, 77 use cases, 37 adjacent needs.

Classification. "Insulated water bottle" → Specification (from a Generalization starting point). "Stainless steel water bottle" → Specification. "Hydro Flask 32 oz" → Equivalence. "Yeti Rambler" (after Hydro Flask) → Substitution. "Water bottle brush" → Complement. "Gym water bottle for women" → Specification with Use Case tag.

Journey modeling. Four realistic flows emerge — a Refinement Path, a Comparison Path across Hydro Flask / Yeti / Stanley, a Known-Item-to-Ecosystem path, and a Recovery Path where "insulated bottle leak proof for kids" returns thin results and the shopper broadens.

Diagnosis. The current listing holds Specification depth but over-commits to brand before establishing category — blocking Generalization recapture mid-session. No comparison modules in A+ content, so Substitution shoppers have no anchor. No ecosystem cross-linking in brand store. Bullet 1 buries kid-safe language behind feature lists.

Rewrite map. Title leads with category term before brand. A+ content adds a comparison module built around Hydro Flask / Yeti / Stanley attributes. Brand store adds an "everything for your bottle" landing page with lids, brushes, sleeves. Bullet 1 moves kid-safe language forward.

The Scorecard

Before rewrite vs. post-rewrite across the six coverage dimensions. A separate example — a dog harness listing, mid-tier brand, moderate rank. Audit reveals the following findings.

30/60

Developing

Pre-rewrite total

0–29

Poor

30–44

Developing — most listings land here on first pass

45–54

Strong

55–60

Optimized

Entry Coverage (Gen)

7 / 10

Refinement (Spec)

6 / 10

Retrieval Precision (Equ)

4 / 10

Comparison Defense (Sub)

3 / 10

Adjacency Readiness (Comp)

4 / 10

Traffic Filtering (Irrelevant)

6 / 10

Rewrite targets. Bring Retrieval Precision and Comparison Defense into the 7 range. Add Complement ecosystem through a brand-store rebuild. Move the listing from 30 (Developing) to 45 (Strong) over two rewrite cycles. Water-bottle case moved 25 → 47; post-rewrite conversion climbed 19% inside 60 days across a five-SKU portfolio.

The rubric does not reward perfection. A commodity product cannot earn a 10 on Comparison Defense because the product has limited differentiation. Optimized band sits at 55+ because most listings plateau before 55 without either deep product innovation or sustained content investment.

Common Classification Errors

Classification looks easy in a reference guide and harder under live keywords. Five confusion patterns show up across most audits.

Specification mistaken for Equivalence

"Gaming laptop RTX 4090 16 inch" feels specific enough to read as Equivalence. Test fails. Equivalence requires a named known item, not a narrow description. A shopper typing "Alienware m18 R2" has committed. Describing narrows a candidate set. Naming retrieves one. Specification gets served feature-led content. Equivalence gets served exact-name retrieval content. Misclassifying collapses the two rails.

Substitution mistaken for Specification

When a shopper types a different brand, sellers sometimes log that as narrowing. "Nike shoes" → "Adidas shoes" adds no new attribute. Original candidate gets replaced. Same need, new candidate. Substitution. Treating the transition as Specification misses a comparison moment and underbuilds A+ content that should carry competitive positioning.

Complement mistaken for Substitution

A harness becomes a leash. A camera becomes a tripod. Related products feel like substitutions because of surface similarity. Test: does the primary product stay wanted? Complement keeps the primary wanted and adds a companion. Substitution replaces the primary entirely. Misclassifying collapses cross-sell opportunity.

Generalization mistaken for Irrelevant

"Gaming laptop RTX 4090" → "laptop" looks like a wide swing. Underlying need stays alive — shopper still wants a laptop. Generalization, not Irrelevant. "Gaming laptop RTX 4090" → "hiking boots" breaks need entirely. That is Irrelevant. Distinction drives whether a listing stays in consideration or gets dropped.

Irrelevant mistaken for Noise

Sellers sometimes discard Irrelevant keywords as junk. The patent treats Irrelevant as a session reset. Prior context dies. The seller needs a filter strategy. Bullet 1 and hero image carry decisive weight because a cold-start shopper landing on a listing has no accumulated intent context. The listing must earn relevance from zero inside three seconds.

Two Layers, One System

Most Amazon sellers optimize for only one layer of a two-layer system. They write keyword-dense titles, manage backend search terms, bid on exact-match phrases. That work remains necessary because keywords still drive retrieval. But patent 12,561,383 documents a second layer operating on top of keyword matching.

Patent 12,561,383 documents a system that classifies intent on top of keyword retrieval. Each query reformulation gets classified into one of six Intent Hexagon types. Sessions accumulate behavioral signal stacks. Purchase events significantly increase the probability of Complement predictions. Sustained Specification sequences without conversion increase the probability of Substitution predictions. An algorithm doesn't guess. Classification runs with trained precision, and training has drawn on Amazon's entire behavioral corpus.

Sellers who understand the architecture hold a structural advantage. They can engineer visibility at the precise moment a customer's intent shifts toward their product type. They can position complementary products in an inversion window. They can build semantic richness that survives BERT encoding. They can optimize across a probability distribution rather than for a single keyword.

The optimization target has shifted. A system classifies intent, predicts transitions, generates queries, and precomputes results — all before a customer types a next search. Sellers who understand the named dynamics documented here will capture customers the system already steers toward them.

Closing

Keywords help a product get retrieved. Intent Mapping helps a product survive whatever move a shopper makes next. Sellers who understand their position within the Intent Hexagon can engineer visibility at the exact moment a customer's intent shifts toward their product type.

Handoff Packet — from Strategy to Copy

Artifact 01

Product Truth Card. Identity, IS attributes, IS-NOT attributes, audience, disqualification signals.

Artifact 02

Classification Table. Every working keyword with dominant and secondary intent labels.

Artifact 03

Journey Worksheet. Modeled flows with risk points flagged.

Artifact 04

Rewrite Plan. Failure-mode-tagged edits mapped to each listing surface.

The writer consumes the packet and executes title, bullets, A+ copy, backend terms. Intent labels propagate through copy so the title surfaces Generalization and Specification terms, bullet 4 carries Substitution defense language, and bullet 5 surfaces Complement hooks pulled from the adjacency tier. Seller gets strategy, then copy, then a scorecard that tracks movement across rewrite cycles.

Read the actual patent

The full 17-page grant. Flip through it below — page by page, spread by spread. Filed by A9.com LLC, assigned to Amazon Technologies, issued under the B1 designation (granted without prior publication, so the printed cover date is the grant date).

Use ← → to turn pages · click the dots to jump

Claude Skill · Ready to Install

Run this on your own listings

The full Intent Mapping method — Truth Card builder, 6-type classifier, journey modeler, vulnerability diagnostician, scorer, and synthesizer — packaged as an installable Claude skill. Drop the zip into Claude Code or Cowork and it triggers automatically whenever you paste keywords or ask for a listing audit.

7-step method6 intent types5 subagents60-point scorecard

How to install

Click the download button to save intent-mapping.zip
Unzip the archive — you'll get an intent-mapping/ folder
Drop it into your skills directory (Claude Code: ~/.claude/skills/, Cowork: upload via Skills settings)
Restart Claude and paste keywords — the skill triggers automatically

↓ Download Intent Mapping skill

intent-mapping.zip · 57 KB · SHA-256 942acc42…

ChatGPT · Intent Mapping GPT

Prefer ChatGPT? Run it there instead.

The same seven-step Intent Mapping method packaged as a custom GPT. Paste your keyword export and product description — the GPT runs the full classify → journey → diagnose → score → rewrite loop inside ChatGPT. No install, no local setup.

Same 7-step methodSame 6 intent typesZero installWeb-based

How to use

Click the button to open the Intent Mapping GPT in ChatGPT
Sign in (a free ChatGPT account works)
Paste your Helium 10 Cerebro or Magnet export plus a product description
The GPT returns the classified set, journeys, diagnostic, score, and rewrite plan

↗ Open Intent Mapping GPT

Opens chatgpt.com · Free account required