This chapter covers how Amazon decides which listings to show — the A9 ranking model (rank as performance × relevancy, computed keyword by keyword) and the newer selection layer sitting on top of it: Cosmo, Rufus, and Alexa-for-shopping. It then walks the working methodology end to end: pulling a keyword universe by reverse-engineering competitor ASINs in Helium 10, Data Dive or Sellerize; cutting the raw pull down with relevancy, volume, conversion, CPR and market-availability filters; placing the surviving terms into title, bullets, backend and images under hard character and repetition limits; and proving after publication that Amazon actually indexed and ranks the listing. It closes with the Search Query Performance loop that keeps the work honest, and with the off-Amazon AI-visibility disciplines (AEO, GEO, review depth) that now sit beside classic Amazon SEO.
The first thing to internalise is that Amazon A9 Algorithm is not a relevance engine with a commercial side effect — it is a commercial engine that uses relevance as an input. One source puts it flatly: A9 "has one main goal, and that's to make Amazon money." It surfaces listings likely to complete a transaction, not listings that merely match the words a shopper typed. Two inputs drive it: keyword relevancy (is the listing indexed for the term, and does its content match it) and sales performance (does it convert and sell once it appears).
The useful model is that rank, per keyword, is performance multiplied by relevancy. Relevancy is scored per keyword and only maxes out when the keyword appears in its exact written form in the listing, ideally at the start of the title — broad or plural variants that a human reads as identical earn less credit, because the matching compares strings rather than inferring intent (see A9 Title Keyword Weighting (Position & Phrase Match)). Performance is conversion rate, click-through rate and revenue. Amazon can't wait for a distinct click or purchase against every keyword a listing could match, so every click or purchase spreads partial "broad" performance credit across many related terms — and that credit is only unlocked for a given keyword in proportion to that keyword's relevancy score. A listing with strong performance but weak relevancy on a term still won't rank for it; so will a perfectly relevant listing that never converts. Optimising one half of the multiplication leaves the other half's gains on the table.
This also explains a fact that surprises new sellers: the top organic results for a term function as a sales leaderboard. The #1 organic listing for a search is typically the best seller for that search, not the listing A9 judged most topically apt. Rank is a trailing indicator of sales performance on that keyword, which is why converting well on a term compounds into better placement over time — and why a brand-new listing, with no sales history to rank on, needs paid visibility and launch pricing to manufacture the first sales signal. Campaign mechanics live in PPC Campaign Structure & Bidding; launch pricing sits in Sourcing, Budgeting & Fulfillment Logistics.
Only one of the two inputs is fully under a seller's control, which is the whole reason keyword work comes first: "you don't have full control over your sales, but you do have full control over your keywords." That is the split in Amazon Listing Optimization: SEO vs. Conversion Optimization — SEO pleases the algorithm so the listing is found, conversion optimisation pleases the customer so it sells. The same source draws a sharper line inside the SEO half: general SEO is indexing only, making sure the listing appears for the right terms; effective SEO chooses and places those same keywords so the result actually earns the click. Most self-described listing optimisers do the indexing half and skip the click-through half. The conversion half of the pair — images, A+ content, split-testing — is Listing Content & Conversion Design.
Everything downstream follows from this ordering, formalised as 3-Step SEO Framework (Keyword Research → Build → Validate): research keywords by reverse-engineering competitors, build the listing around them, then create and validate in Seller Central that Amazon indexed and ranks it. The claim attached to it is that sellers who fail have skipped step 1 (guessed at keywords instead of pulling competitor data) or step 3 (stopped at publish). One boundary worth setting now, from Algorithm-Trust Threshold for Demographic Targeting: manual demographic targeting is a pre-data workaround. Once a product has roughly 3–5 months on sale or 1,000+ units sold, stop guessing at who the buyer is and let A9 route the listing — above that threshold Amazon's own matching beats manual inference.
Cosmo (Amazon's Rufus-Powering Algorithm) is the system described as powering Rufus, Amazon's conversational shopping assistant — and it is presented as distinct from A9 rather than a replacement for it. Mechanically, the account runs: each Rufus query is dynamically routed to a different underlying LLM via Amazon Bedrock (Amazon Nova, Anthropic models) depending on whether it's a simple lookup or a research-heavy question; retrieval-augmented generation layers live data on top of the base model; and five data sources feed it — catalog data, reviews, Q&A, external web and publication sources (curated review sites, notably not Reddit or Quora), and customer behaviour weighted toward roughly the last two to three weeks rather than all-time history. The consequence sellers are told to draw is that Cosmo wants a technical attribute-node mapping — a knowledge graph in which a product is tagged as "used with" complementary items, "used in location X" — so it can surface for conversational queries that never contain the product's literal keywords. That reframes keyword density as a secondary signal underneath structured attribute mapping.
Cosmo's 15 Semantic Relations is the operational version of that graph: a taxonomy of 15 relations Cosmo is claimed to check a listing against — what the product is, "used for function," "used for event," "used for audience," "used in location," and so on. The practical split given is that a title should hit roughly 5 of the 15, prioritised via RICE Prioritization Framework (Reach, Impact, Confidence, Effort), while the full listing — title, bullets, description, backend search terms — should cover all 15 somewhere. The diagnostic is a "Cosmo doctor check": scan each field against the 15 relations and flag the gaps, e.g. a beard-oil listing that never says what event it suits, who it's for, or where it's used.
Rufus (Amazon's AI Shopping Assistant) matters to sellers for two concrete behaviours: it scans listing images to extract contextually relevant keywords and attributes, and it mines customer reviews for sentiment to summarise why a product is liked or disliked. That makes image content and review sentiment de facto SEO inputs, not just conversion aids. Essential Candy, in the Helium 10 Scale Stories case, used Rufus in the opposite direction — querying it to sanity-check which benefit claims ("helps with nausea," "ginger for morning sickness") its own reviews actually supported before writing new copy, and prioritising a reshoot specifically because Rufus reads images. The framing offered is explicitly anticipatory: this work doesn't solve a problem that exists today, it positions the listing for when more shoppers route through Rufus.
The same shape appears one layer out in ACO (Alexa Optimization), a framework credited to Andrew Bell (a name given as Andrew Bezos in the NPO and QPO material) and built from Amazon patents, science papers and hands-on work on 4,000+ ASINs. Its central claim is architectural and worth repeating because it prevents a common overreaction: there is no A10. A9 still builds the candidate pool; Alexa for shopping — reaching roughly 100 million shoppers — decides whether a retrieved product gets understood, trusted and selected. A product no query retrieves can never be chosen, however well optimised for the assistant. Its seven moves are: stack high-volume terms into natural noun phrases of around 75 characters; bridge semantically to rooms, occasions, recipients and styles the shopper never typed; map features to outcomes so the listing is inferable into situational queries; keep doing ordinary A9 keyword work; build a per-ASIN mission map; complete every structured attribute field (missing fields cause exclusion before scoring, not a scoring penalty); and treat the whole product page as indexed surface — title, up to 10 bullets for brand-registered sellers, description, native A+ text, lifestyle images with and without text — plus third-party web content, since a review outlet's copy can win the citation over the brand's own listing if it's better written.
QPO (Query Planning Optimization) names the reframe underneath all of this: an AI assistant fans a single prompt like "find something for my kids in our home" into dozens of sub-searches, branching by room, style, budget, material, recipient and use case. "You're no longer trying to rank for a keyword, you're trying to be eligible across the whole family of queries that one shopping mission generates." QPO is operationalised by mapping those query plans from autocomplete and Alexa follow-up prompts, then covering the resulting phrase family with NPO (Noun Phrase Optimization) content.
One honesty note the material itself insists on, and this chapter keeps: public detail on Cosmo comes largely from vendor-interview content (e.g. ZonGuru) promoting paid optimisation services rather than from Amazon documentation, so specific figures — $10B incremental sales, a 60% purchase-likelihood lift — should be treated as vendor-sourced claims pending independent confirmation. ACO is likewise practitioner reverse-engineering. The pragmatic read: the tactics here (complete your attribute fields, say who and where and what-for, write phrases that read like human language) cost little and are defensible on plain conversion grounds, so act on them; the mechanics are hypotheses, not documented facts.
Start with what not to use. Google vs. Amazon Search Intent (Informational vs. Transactional) rules out Google keyword data outright: "Google searches are mostly informational… Whereas Amazon searches are transactional. People are searching because they're ready to buy." Volume, phrasing and relevance signals don't transfer between the two, so research has to come from Amazon-specific tools. Free keyword tools get the same treatment — "don't even try using the free keyword tools. You get what you pay for. Garbage in equals garbage out."
The core technique is the reverse ASIN lookup, and it is the same idea in all three tool stacks: instead of brainstorming seeds, pull the keywords competitors already rank and convert on. Reverse ASIN Lookup (Helium 10 Cerebro) specifies the sourcing precisely — search your main keyword on Amazon, take the top 3 organic listings (skip anything marked Sponsored; those paid for placement rather than earned it), pull each B0 ASIN from the product URL, and enter them in strict rank order, which deliberately weights results toward the strongest performers. Widen to 5–6 ASINs for lower-volume products. Then run a second pass: search the single top-volume main keyword directly in Helium 10 Cerebro rather than via ASINs, which surfaces terms the ASIN pass missed; Helium 10 auto-dedupes on import, so merging only adds what's new. The same lookup doubles as a niche-difficulty gauge — a dense, high-strength keyword profile on a top competitor signals an entrenched incumbent, a thin one signals room (this overlaps with Product Research & Validation).
Tool choice mostly comes down to how many competitors you can analyse at once. Cerebro is capped at 10 ASINs per run; Data Dive's Divebox, per Reverse ASIN Lookup (Data Dive), handles roughly 20 — and that gap is cited as the specific reason some longtime Helium 10 users moved their keyword research to Data Dive while keeping Helium 10's Blackbox for product research. Sellerize covers the same ground differently: Sellerize Keyword Re-ranker pulls every keyword a given competitor listing ranks for (run the top 3 related listings, filter, dedupe, merge), while Sellerize Keyword Hunter Pro expands a seed into a "semantic core" filtered to keywords with demonstrated organic sales history rather than every term that has ever been searched. On the Helium 10 side the seed-expansion counterpart is Helium 10 Magnet (Seed Keyword Expansion), which grows outward from a term you already know describes the product and fills adjacent gaps competitors' listings don't surface.
Keyword List Seed-and-Steal Assembly (Dedupe & Manual Verification) is the assembly order: seed research, steal from winners, merge and dedupe, then manually verify — search each keyword on Amazon in All Departments and confirm it actually returns products like yours rather than trusting tool output blind. Misspellings are the step most sellers skip. Helium 10 Misspellinator runs the master list and splits typos into autocorrected (Amazon silently fixes these, so targeting adds little) and non-autocorrected (Amazon does not fix them, so a shopper who mistypes sees a different results page — worth indexing for directly); discard the first group, keep the second. ChatGPT-Assisted Misspelling Generation is the non-Helium-10 route: prompt with "Give me a downloadable CSV file containing all the common Amazon search only high-intent typos for the following keyword phrases," paste the list, then strip the original-keyword column and header row before reimporting.
A raw Cerebro pull can return 2,600+ keywords, so the cut is mechanical. Cerebro Filtering Protocol (Organic / Competitor Rank / Number of Competitors) applies three filters in sequence — Match Type = Organic, Competitor Rank ≤ 60, Number of Competitors ≥ 2 — which typically lands around 200–300 keywords. Rank >60 excludes terms too obscure to be reliable; a single ranking competitor excludes the idiosyncratic. A manual pass then strips irrelevant terms and competitor brand names, and the reason for the second is financial, not tidiness: brand terms left in the list end up spending your PPC budget advertising someone else's brand. Cerebro's output also carries an Amazon Recommended Rank — Amazon's own signal of how relevant a keyword is to that ASIN. Run your own ASIN through before launch and drop anything that returns a blank score: Amazon doesn't consider the term relevant to your listing, whatever its search volume.
Data Dive does the cut by sorting instead of filtering. Data Dive Four-Bucket Keyword Sorting System splits every pulled keyword into four buckets: trash (discarded), master list (≥30% relevancy and ≥450 monthly searches), outliers (relevancy <30% but volume >2,000/mo), and residue (low relevancy, low volume, or beyond the top 500 terms). Relevancy here has a specific definition — the percentage of the analysed competitor set ranking in the top 45 organic positions ("page one") for that term — and the thresholds are adjustable in settings. The overflow tiers are the point, not waste: outliers are often high-volume terms most competitors systematically miss, and manually scanning residue sorted by relevancy surfaces variant-specific terms ("waterproof" for a nylon bag) that deserve promotion into the master list. A large outlier bucket concentrated on one ASIN — the example given is 66 keywords carrying 1M+ search volume — flags a competitor whose dominance is worth investigating. Residue skewing heavily to brand names is itself a signal that generic terms in the niche are hard to rank for.
Two more cleanup tools. Data Dive Roots Feature (Bulk Keyword-Group Exclusion & Prioritization) groups keywords by shared root word so you can act on a whole family at once — bulk-excluding a competitor brand root in phrase form (removes every term containing the word) rather than exact form (removes only the literal term, leaving long-tail variants sitting in the list) — and, in the other direction, shows which roots carry the most cumulative relevancy and volume, i.e. which words most deserve title and bullet space. Relevance-Sort Keyword Cleanup is the blunt version: sort by relevance score, treat anything around 10 or below as unrelated, and strip wrong-variant terms in the same pass. What survives becomes the Master Keyword List & Listing Scorecard — search volume, traffic potential, buyer intent, difficulty and opportunity score per keyword — which then drives listing copy, PPC targeting and rank tracking as a single shared artefact rather than three separate lists. One benchmark worth pulling while you're in the data: SV on Page One (Cumulative Top-45 Search Volume), the cumulative monthly search volume a product ranks for inside the top 45 positions. It tracks closely with sales performance, so comparing your own total against top competitors' totals says how far short of competitive coverage the listing currently is.
A filtered list of 200–300 relevant keywords is still not a target list. The selection logic that separates them is the most valuable idea in this chapter, and it inverts the beginner instinct. High-Conversion Keyword Selection Strategy starts from the A9 mechanics in the first section: rank is driven by sales generated through a keyword, not by how often that keyword is searched. So a low-volume, high-converting term out-ranks a high-volume, low-converting one over time — and the same logic carries into paid, where a lower-volume high-conversion term produces a lower effective ACoS than a high-volume term that clicks and doesn't buy. The cited example is stark: for the same product, a high-search-volume keyword converted at ~6% while a lower-volume buyer-intent keyword converted at ~52%.
The strategy replaces search volume with three axes, and the screening thresholds are concrete. Drop any keyword converting below ~20%. Keep only keywords where sellers outside the top 3 capture at least ~40% of sales — that is Market Availability Metric, the percentage of a keyword's sales not absorbed by the top 3 listings. Read it as "how much of this keyword's demand is up for grabs": roughly 70%+ is a green light, since a new listing isn't landing in the top 3 at launch anyway, while ~10% is a red flag — the incumbents take nearly everything, so page one wouldn't translate into revenue. The third axis is CPR (Units to Rank) Metric, the number of units a listing must sell within an 8-day window to reach page one for that term; lower CPR means the keyword is cheaper and faster to win. Applying the two-metric filter first shrinks hundreds of keywords to a workable set before relevance cleanup and competitor merging.
Around those, a few coarser floors and gauges. Minimum Keyword Search-Volume Threshold (~300/Month) sets a rule of thumb of roughly 300 monthly searches to be worth targeting at all; around 100/month is simply not a good keyword. Title Density (Helium 10 Blackbox Metric) — from Helium 10's Blackbox Keywords tab — counts how many page-one listings already carry the exact phrase, in exact word order, in their title; it's a competitive signal distinct from review count, and setting a low maximum (e.g. 5) filters for keywords where you aren't fighting entrenched title optimisation.
"Right to Win" Keyword Selection Framework converts all of that into one question: not "does this term have volume?" but "can this listing plausibly rank for it, given the budget and the competitors already dominating it?" The worked case is an essential-oil hard-candy brand whose auto campaigns kept surfacing "hard candy" as a top-matching term — a results page owned in practice by Hershey and Jolly Rancher. "It's going to be really tough for you to win against Hershey's budget"; a seller could spend an entire monthly budget on that one keyword and still lose the auction. The fix was to redirect toward long-tail, benefit-specific terms where differentiation gives a real shot — "nausea," "ginger candy" — identified from Cerebro output filtered for semantic relevance. The brand's own branded keyword was also ruled out, judged non-incremental.
Two principles govern how wide to spread. Keyword Concentration Principle (Top 3–5 Keywords Drive 80–90% of Sales) claims that for most products the top 3–5 keywords account for roughly 80–90% of sales intent, making concentration of copy and PPC on a handful of terms a resourcing rule that sits upstream of any long-tail prioritisation — most of the tail isn't worth prioritising at all. Midtail Keyword Prioritization says where the non-obvious insight lives when you do look wider: longtail terms carry too little volume to act on, head terms carry huge volume with diluted relevance, so the actionable problems surface in the midtail — specific enough to represent a real intent, big enough that fixing them moves sales.
Intent vs. Research Keywords (Sequencing Framework) adds the time dimension, and it is the rule most likely to save money. Intent keywords are for shoppers who already know what they want ("retinol under eye patches") — high conversion, short path. Research keywords are for browsers ("birthday gifts for women," ~340k searches/month, of whom only ~0.8% buy anything); low conversion there isn't a defect, it's a longer journey, and the objective shifts to being remembered. The sequencing rule is strict: rank high-intent product-specific terms first, build reviews and conversion history, and only then expand into research terms. Doing it in the other order is called "an expensive way to light your money on fire," reserved for brands already doing five-figure units per month. Once the precondition is met, the payoff can be fast — a documented case moved from organic rank 71 to rank 2 in five days on "birthday gifts for women" with a short ad burst selling 150 units, and held rank 2 after the campaign stopped.
Sometimes the right answer is that a keyword doesn't belong in this listing at all. Keyword Segmentation via Product Variant Listings says that when a valuable term spans a different use case, launch it as a separate variant rather than broadening the existing listing — stretching copy to cover an adjacent use case dilutes both the copy and the buyer match, costing conversion on the terms that matter most. And Hero Variation & Variation-Level Keyword Demand warns against launching a variant on its search volume alone: best-selling multi-variant listings typically concentrate ads and ranking on one cheap, generic hero variation that absorbs traffic for the whole page, then convert shoppers to pricier or less-common variants once they've landed. The hero's job is to be the cheapest entry point that gets clicks — not the highest-margin SKU.
Finally, classification at scale. AI-Assisted Intent Clustering (ChatGPT Two-Prompt Method) offloads buyer-intent tagging to ChatGPT in two prompts: first ask it to identify the top keyword families by buyer intent present in the exported keyword-plus-volume CSV, then ask for a deduplicated CSV assigning every keyword to exactly one intent category, using a stated priority order from most-specific to most-generic to resolve ambiguous terms. That priority instruction is what makes the classification deterministic; without it, borderline keywords get assigned inconsistently or duplicated across categories. The resulting intent column is then filtered one group at a time into campaign build-outs in PPC Campaign Structure & Bidding.
Before any copy gets written, split the list. Helium 10 Four-List Keyword System (Main, Related, Misspelled, Master) keeps four lists rather than one pile — Main (primary, highest-relevance terms), Related (adjacent secondary terms), Misspelled (the non-autocorrected typos), and Master (the consolidated list that feeds copy). The reason is downstream, not aesthetic: each list becomes its own campaign, so match-type and relevance data stay clean per segment and misspelled-keyword traffic can be read separately from main-keyword traffic. The analogy offered is a food truck that logs timing, best-sellers and promo effects versus one that just serves whoever walks up; dumping every keyword a tool spits out into one campaign and hoping Amazon sorts it out is the "backwards" approach that makes it impossible to see what's working. Practically, you pass through the master list several times, tagging a subset into each saved Helium 10 Folders (Keyword List Storage) folder so the lists can be pulled back into the builder, index checker and tracker later without re-running research.
Helium 10 Listing Builder is where lists become copy. As you type it marks each keyword used or unused, so nothing researched gets silently dropped; when the title, bullets and description are drafted, clicking "hide used" isolates the leftovers, which go straight into the backend field rather than being forced into visible copy. It has two shortcuts for people who find copywriting hard: an AI path (upload a product photo, click analyse, generate a draft, then refine against the master list — and it now pulls Search Query Performance data for the ASIN automatically), and Competitor ASIN as Listing Builder Template, which pre-fills the builder with a top competitor's listing and auto-crosses off whichever of your keywords already appear in it. The discipline with the second one is explicit: use it as a reference point, then rewrite every section so the final listing is original and tailored to your own keyword list, not a copy of the competitor's.
A9 Title Keyword Weighting (Position & Phrase Match) establishes the hierarchy: title keywords carry the most weight of any field; earlier positions carry more weight than later ones; and full phrases matched intact are weighted more heavily than the same words scattered across the copy. "The keywords you include in the title carry the most weight… And the keywords closest to the beginning of the title have more weight." A keyword earns top relevancy only in its exact written form, so write "toiletry bag for men" verbatim rather than trusting "men's toiletry bag" to pick up the credit. The first five words are claimed to carry outsized weight for both A9 and the human shopper, who mostly doesn't read past them.
Amazon Listing Title Structure Formula gives the layout: lead with the shortest, most-searched relevant keyword phrase, then say what the product is, what it does, what's included, and what differentiates it. Use | rather than commas as separators. Brand placement depends on recognition — a new or unknown brand goes at the end, because an unrecognised name up front earns no click-through and reads as noise ahead of the words that say what the thing is; a brand with real search volume goes at the front, where recognition converts. Check the brand's own search volume before deciding. RICE Prioritization Framework (Reach, Impact, Confidence, Effort) decides what earns title space: Reach, Impact, Confidence and Effort — with the distinctive twist that Effort here means character cost against a hard field limit, not labour. High reach, high impact, low character cost goes in the title; long-tail, character-expensive terms get pushed to bullets or backend.
There isn't one title formula, though. Metric-Diagnostic Title Strategy Selection picks the pattern from whichever top-line metric is weak: a sessions/CTR structure (brand + product type + the 1–2 highest-traffic keywords) only when 1–2 keywords drive 70–80% of the ASIN's sales; a CTR/CVR structure leading with the single strongest differentiator when conversion is already strong but clicks aren't; and a keyword-coverage structure for long-tail categories or brand-new ASINs with no dominant term yet.
Keyword Repetition De-Indexing Penalty is the constraint most likely to be actively hurting an existing listing. The claim: repeat an exact search term more than 3 times across title, bullets, backend and description and Amazon can de-index the listing for that term entirely — "the reality is Amazon punishes you when you use it more than three times. So, you lose your ability to rank on it." One cited case had a term repeated roughly 20 times and lost indexing on it purely from copy hygiene, not from weak relevance or bids. Inside the title the cap is stricter, around twice. And the flip side matters just as much: a keyword only needs to appear once anywhere in the listing to be indexed, so repetition buys nothing and risks everything. Audit any term appearing more than 3 times and trim.
Bullet Point 1,000-Character Front-Loading Rule adds a second ceiling: Amazon reportedly only pays attention to the first 1,000 characters of bullet content, so anything past that adds no indexing value — the instinct to fill every available character is counterproductive. Front-load the strongest keywords and selling points into the earliest bullets and the earliest sentences of each. Bullet length is also a conversion lever in its own right: "when bullet points are too long, it becomes a huge turn-off and shoppers are not going to spend the time to read it."
Overflow goes to Amazon Generic Keywords (Backend Search Terms Field) — Amazon's hidden Generic Keywords field, 2,500 characters total with a 500-character-per-line cap, which indexes terms that never appear in visible copy. Terms already used verbatim in the listing don't need repeating here. Two categories belong here specifically. First, the mismatched tier from Two-Tier Keyword List Strategy: after manually searching each master-list keyword on Amazon in All Departments, terms whose results don't match your product get indexed in backend only, never in visible copy or PPC, since their dominant shopper intent points elsewhere and paying for those clicks wastes spend. Second, low-volume residue terms can go in as single words rather than full phrases — because Amazon infers connections between a root word and its long-tail variants, dropping "fantasy" in backend can do the work of stuffing "fantasy booknook" into the title.
That inference is generalised by NPO (Noun Phrase Optimization), the language layer of the ACO framework: rather than stuffing loose keywords, build titles, bullets and A+ content from a head noun plus stacked modifiers — lamp → floor lamp → metal floor lamp → dimmable metal floor lamp for reading nook, with "UL listed / ETL certified" as a proof layer. Its ground rules are worth keeping: only phrases that accurately fit the product, never force irrelevant terms, keep parent phrases (what it fundamentally is) separate from refinement phrases, use stacks only where they read naturally, and source phrasing from autocomplete, ads, reviews, Q&A and catalog data rather than inventing it. Its concrete moves include putting a constraint or proof phrase in the first two bullets, since those act as hard filters agents apply before a product reaches comparison, and adding an audience or occasion line ("gift for horse lovers") that most listings simply omit. NPO also documents a useful mechanism: a listing ranking and converting well for a root term ("book nook kit") can surface for variants like "booknook kit" that appear nowhere in the copy.
Two further surfaces are now machine-read. Hidden SEO Image Text (Invisible Keyword Indexing) embeds keyword-optimised text into the listing image — text not on the physical package — purely so Amazon's image scanning and Rufus index it; Essential Candy was advised to drop its secondary keyword list into blank space on the main image, capturing indexing without cluttering the package or the bullets. And AI Shopping-Assistant Q&A Extraction → Objection-to-Content Workflow turns the assistant itself into a content-gap audit: query Rufus about your own listing to see what it tells shoppers (including the negative-review summaries it surfaces transparently), harvest a larger sample of its auto-generated follow-up questions, extract the Q&A pairs to a spreadsheet, and convert recurring questions and objections into listing content — a "real questions from real shoppers" FAQ image, objection rebuttals, care-instruction imagery that preempts misuse-driven negative reviews. Re-run it monthly; the question set doesn't shift fast. Before publishing, run the draft through the Listing Scorecard side of Master Keyword List & Listing Scorecard — keyword coverage, relevance, exact-match versus broad-match gaps, and competitor blind spots — to confirm you're covering what competitors are missing. The visual and persuasive craft of these same assets belongs to Listing Content & Conversion Design.
Step 3 of 3-Step SEO Framework (Keyword Research → Build → Validate) is the one most sellers skip, and the framework's claim is that skipping it is one of the two ways listings fail. "Keyword indexing is one factor, but ranking is what actually drives sales." An optimised listing is necessary and not sufficient — you have to confirm Amazon agrees.
The verification runs in three stages, and the middle one catches the most common silent failure. A listing can contain every individual word of a target phrase scattered across its copy and still not be indexed for the phrase. Helium 10 Index Checker's "traditional index" column tests exactly that: whether the exact phrase returns your ASIN in Amazon's search results. The workflow is to upload the master keyword list with "remove duplicates" and "maintain phrases" checked, add any unchecked phrase into the listing verbatim, then recheck 15 minutes to an hour later once Amazon has reindexed. Then Helium 10 Keyword Tracker answers the next question — not "can it be found" but "where does it sit." It reports rank per keyword in tiers: top 10 and top 50, split organic and sponsored. Add the ASIN and master list and watch tier movement over time; one practitioner checks it daily specifically to know where a listing ranks before adjusting bids, rather than moving spend blind. Cross-referencing tracked organic rank against the keywords funded in launch campaigns is what tells you which keywords are actually converting ad spend into organic position and which are just absorbing budget — the reporting side of that lives in PPC Optimization, Analytics & Advanced Targeting.
Keyword rank isn't the whole visibility picture. Helium 10 Boost (Browsing Scenarios Monitor) monitors browsing scenarios — how and where a listing surfaces in related-item carousels and category browse, as opposed to direct keyword search — and a listing can rank well on search terms while quietly losing placement in browsing contexts. One practitioner checks it seven times a day.
Some of this validation can be pulled forward. Pre-Launch Test-Listing Relevancy Validation builds a placeholder test listing before committing the real SKU and UPC, then checks Amazon's Recommended Rank for the target keyword inside Cerebro. The documented sequence is: test listing → Recommended Rank check → a Vine batch (30 units, fixed regardless of price or category) → go live with the real SKU/UPC → start PPC immediately. The point is to discover a relevancy problem while it's still cheap to fix, rather than after the real listing and its Vine reviews are committed.
If the test listing shows weak relevancy, Search-Find-Buy Relevancy Signal (Small-Scale) is the fix offered: add the keyword to the title, drive traffic, and have one or two people search that keyword and buy the product so the purchase signal attaches to the term. The source draws the compliance line itself and it's worth quoting rather than paraphrasing: "to do like one or two just to get some relevancy signals, that is not against what Amazon says… [but] trying to get 20 people to do search find buy for a keyword and trying [to] rank for that keyword, that is against Amazon terms of service." This is one seller's reading of where the line sits, not Amazon policy — treat the risk accordingly, and see Reviews & Account Health for the same escalation logic applied to review acquisition.
Finally, the order of operations when a live listing is underperforming inverts everything above. Two-Step Conversion-Then-Indexing SOP — Brandon Young's sequence — says fix conversion rate first, then fix indexing, never the reverse. Without a strong conversion rate a listing can't rank, can't compete and can't be profitable, so any ad-spend increase before conversion is solid is spend poured into a listing that can't convert regardless of visibility. The underlying claim reframes PPC entirely: "the number one reason you run ads is to rank." Because organic rank is driven by ad strategy, the wrong strategy tanks rank fast and the right one restores it just as fast. And the diagnosis has to be root-by-root and keyword-by-keyword — finding where conversion is already strong (worth indexing effort), where money is being wasted (indexing effort is premature), and where to double down — never a blanket account-level spend increase.
Once a listing is live and branded, Amazon Search Query Performance (SQP) Report becomes the highest-value feedback loop available — described by its advocate as "possibly the most powerful data source on Amazon inside Amazon Seller Central that Amazon actually gives us access to." It is not the PPC search-term report: that one shows terms triggered by ads, while SQP shows organic, brand-versus-market performance for every search term shoppers used to find products in the category, ads or no ads. Access it via the hamburger menu → Brands → Brand Analytics → Search Analytics → Search Query Performance.
Pull the last full calendar month, at least two weeks old, so attribution has settled. For each search term the report gives a total (all sellers) and a brand-specific count at four funnel stages — impressions, clicks, cart adds, purchases — from which you compute brand share % at each stage. Two handling rules: ignore cart adds, treated as very difficult to affect and not diagnostic; and sort by brand purchase count rather than by Amazon's built-in Amazon Search Query Score. That score is Amazon's own blend of search volume and relevance to the product (1 is the best possible, lower is better), but since purchases are a big part of how Amazon computes it, sorting by purchases reproduces Amazon's prioritisation without a separate column — and purchases are "the data that we really trust." Then apply Midtail Keyword Prioritization deliberately: scroll past the top one or two terms, whose story you already know, because "where are the deep insights, they tend to come from the midtail keywords" — longtail has too little volume to act on, head terms carry volume with diluted relevance.
SQP Funnel Stage Diagnostic (Thumbnail vs. Listing Problem) is what you do with the shares, and it localises the problem precisely:
One worked contrast from a single brand's two products — leak-proof ice bags versus ice storage bags — produced opposite diagnoses side by side: one product's healthy click share masked a listing-content failure, while the other was converting well down-funnel and simply needed more traffic. The same diagnostic also works within one ASIN, comparing thumbnail CTR across semantically distinct terms: for a wooden phone holder, if "travel wooden phone holder" clicks at roughly half the rate of "desktop wooden phone holder," the main image is failing to communicate the travel use case to that specific segment — which calls for a targeted primary-image split test aimed at that segment, not a generic refresh.
Organic CTR Back-Calculation from SQP extracts a number Amazon never surfaces anywhere in Seller Central. Ad CTR is available in the advertising console; true organic CTR is not. Export SQP to a spreadsheet and add a column dividing brand clicks by brand impressions per term. That is described as the only available way to get the figure. Track it over time per priority term to catch organic thumbnail and title drift independently of what paid placement is doing.
SQP also closes the loop back to keyword strategy. In the Essential Candy case, the report showed generic head terms like "peppermint" and "hard candy" capturing only ~0.05% of the available search volume for the ASIN — hard evidence the listing was chasing terms it had no path to winning — and that finding is what redirected the keyword strategy toward higher-intent, lower-competition terms like "nausea" and "morning sickness candy." The same data feeds Helium 10 Listing Builder, which now imports SQP for the ASIN being edited automatically, folding real converting-query data straight into the drafting step.
A growing share of product discovery now happens inside AI-generated answers rather than on a results page, and that runs on different signals than either Google ranking or A9. AEO (Answer Engine Optimization) is the practice of getting a brand cited and recommended inside those answers — ChatGPT, Gemini, Rufus (Amazon's AI Shopping Assistant). The reported decision inputs are brand-authority statements, mentions across 10+ credible sites, superlatives ("most cited," "leading"), and source credibility, with major media outranking Reddit — though Reddit and Quora presence still feeds citations. Source mixes differ by engine: ChatGPT leans on earned media (~40%) and DTC/marketplace listings (~23%), while Walmart's Sparky and Rufus lean even more heavily on earned media and affiliate review sites. The tactic list is concrete — fast-turnaround press releases and earned placements, a Reddit posting SOP, guest posts and comparison-site listings, FAQ/glossary/comparison-chart content, schema markup and product cards, an llms.txt file and an LLM-oriented sitemap — plus AI-specific rank tracking so progress is measurable the way rank trackers measure Google.
Two boundaries keep this honest. There's a citation gap: ranking well on Google doesn't guarantee AI citation, which is why AEO is treated as a separate workstream rather than a byproduct of SEO. And there's a ceiling — AEO can shape a narrative but can't override facts major media has already established about a brand, such as country of origin or a controversy. AI is reportedly easier to influence where it has less built-in knowledge, which structurally favours niche and private-label products over mainstream brands.
GEO (Generative Engine Optimization) is the parallel discipline aimed at generative search platforms — ChatGPT, Claude, Gemini, Perplexity. It runs on per-platform citation logic and training-data lag, since each vendor's retrieval and ranking pipeline evolves independently of Google's, and its structural tactics overlap with AEO: one clear claim per page, answers front- and back-loaded, evidence-backed claims rather than bare assertions. Practitioners describe a transitional window — estimated at roughly 6–8 months as of 2026 — during which the same piece of content can serve both SEO and GEO, after which the two may need separate treatment.
The input that appears to matter most is reviews, and here Amazon sellers face a specific structural problem. Review-Driven AI Shopping Visibility reports that AI shopping engines increasingly decide what to recommend from the depth and substance of customer reviews, pulled from public sources — Reddit, Yelp, Google — not just a brand's own site. Per Stackline, ChatGPT users alone make over 84 million shopping queries a week, and Fireclay Tile's CEO credits reviews as "1 million percent" driving their AI visibility. But Amazon has quietly blocked OpenAI's bots from crawling Amazon content, including reviews, so ChatGPT-style assistants can't see Amazon reviews at all. That inverts the usual logic: a deep Amazon review moat, the thing this whole business builds toward, is invisible to the engines, forcing brands to duplicate review-gathering effort off-Amazon. Tactically, brands are optimising for substance over count — Paco, a dog food brand, offers $20 off after a customer's third order in exchange for a review, waiting two weeks after a one-time order or three full reorder cycles for subscriptions before asking. The Amazon-side rules and risks of review acquisition are Reviews & Account Health.
How much of this to fund, and where, is a judgement call — which is what Agentic Commerce Autonomy Tracker is for. Credited to analyst Scott Wingo, it ranks AI shopping assistants along research → find → buy: research means the assistant answers product questions without surfacing specific listings; find means it actively surfaces and links to listings (product cards); buy means it can complete the transaction on the shopper's behalf. Track which direction a given platform is moving to decide where AEO investment is worth it — and re-check periodically, because movement isn't one-directional. A platform can retreat down the spectrum (removing a buy button) as readily as it advances. Broader off-Amazon channel expansion is Scaling, Omnichannel & Brand Growth.