The Problem With Amazon SEO Advice
The Amazon seller ecosystem is full of confident claims about "how the algorithm works." Blog posts describe ranking factors with decimal-point precision. Consultants sell courses explaining the inner workings of systems they have never seen. Forums debate whether "A10" weights external traffic at 3x or 5x.
Almost none of these claims have a primary source.
Here is what most sellers don't realize: Amazon has published dozens of research papers through Amazon Science that describe their actual search and ranking systems. These papers are written by the engineers who build and maintain those systems. They are peer-reviewed, presented at top-tier conferences like SIGIR, KDD, and CIKM, and they describe real production architectures — not speculation.
We read them. Here is what they say, and where the SEO blogs get it wrong.
Myth: "A10 Replaced A9"
This is the most widespread myth in the Amazon seller community. The story goes that Amazon's "A9 algorithm" was replaced by a newer "A10 algorithm" sometime around 2019–2020, with A10 supposedly giving more weight to organic sales, external traffic, and seller authority.
There is no "A10." There never was. Amazon's search subsidiary was called A9.com, incorporated as a company name. No Amazon publication, API document, patent filing, or researcher has ever described an algorithm called "A10" or "A11." Search the entirety of Amazon Science, arXiv, ACM Digital Library, or any academic database — it does not exist.
The actual system is a continuously evolving machine learning pipeline that changes incrementally through thousands of experiments per year. The SIGIR 2016 paper "Amazon Search: The Joy of Ranking Products" describes the architecture as a learning-to-rank system with gradient-boosted decision trees. By 2022–2023, papers describe BERT-based semantic embeddings and LLM-scale retrieval. The system evolved continuously — it was never "replaced" by a versioned successor.
When an SEO blog tells you "A10 now weights X differently than A9 did," they are presenting their speculation as fact about a system that does not exist. The real evolution is far more interesting and better documented.
Myth: "Keywords in the Title Are the Most Important Ranking Factor"
The traditional Amazon SEO playbook says to pack your title with every possible keyword because the algorithm is essentially a keyword-matching engine. Get the right keywords in the title, and you rank.
What the research actually says: Amazon's retrieval system moved well beyond keyword matching years ago.
Amazon Search uses a multi-stage retrieval architecture. The first stage retrieves candidate products from the catalog, and the second stage ranks them. The retrieval stage is where keywords historically mattered most — if your listing didn't contain the search term, it wouldn't be retrieved.
Amazon's retrieval stage now uses semantic embeddings trained on transformer models. The 2022 paper on HISS (Hybrid Inference with Semantic Search) describes a production system combining SBERT and DSSM with knowledge distillation, enabling the system to understand the meaning of queries rather than just matching tokens.
The 2023 paper "Web-Scale Semantic Product Search with Large Language Models" goes further, demonstrating that LLM-scale models can retrieve relevant products even when there is zero keyword overlap between the query and the product listing. A customer searching for "something to keep my coffee hot at work" can be matched to an insulated travel mug listing that never contains those exact words.
Does this mean keywords don't matter at all? No. Keywords still provide important retrieval coverage, especially for long-tail and specific queries where the semantic models have less training signal. The Shopping Queries Dataset (ESCI), released as part of KDD Cup 2022, shows that Amazon categorizes query-product relevance into four levels: Exact, Substitute, Complement, and Irrelevant. Having strong keyword coverage helps ensure you are retrieved for Exact-match queries.
But the days of keyword matching as the primary retrieval mechanism are over. Semantic relevance — whether your listing genuinely describes the product a customer is looking for — matters more than whether you stuffed the exact search phrase into your title.
Myth: "Sales Velocity Is the #1 Ranking Signal"
The SEO consensus says that sales velocity — how many units you sell per day — is the single most important ranking factor. Get more sales, rank higher. Rank higher, get more sales. A virtuous flywheel.
What the research actually says: Amazon's ranking system is multi-objective, and purchase rate is one signal among many.
The paper "Multi-Objective Ranking Optimization" from Amazon Science describes how Amazon's ranking system jointly optimizes for multiple objectives using stochastic label aggregation. The system does not simply rank by "most purchased." It balances relevance, purchase likelihood, and other business objectives simultaneously.
The SIGIR 2016 paper makes this point explicitly with a memorable example. Consider a search for "diamond earrings." The most-purchased product might be inexpensive cubic zirconia studs, but a customer searching specifically for diamond earrings would find cubic zirconia results irrelevant. The most-purchased product is not necessarily the most relevant product.
Sales velocity is a real signal. But it does not dominate. The ranking model balances it against relevance, price appropriateness, product quality signals, and category-specific behavioral features. A product with moderate sales but high relevance to a specific query can outrank a product with higher overall sales volume but weaker query-product match.
Myth: "Organic Sales Count More Than Ad Sales"
This myth is particularly persistent: that Amazon's ranking algorithm distinguishes between organic purchases and ad-driven purchases, weighting organic sales more heavily because they represent "real" demand.
What the research actually says: Amazon's published ranking architecture trains on all purchase data regardless of traffic source.
The ranking models described in Amazon's papers use purchase events as training signals and behavioral features. No published paper describes a mechanism that identifies the traffic source of a purchase and weights it differently. A purchase is a purchase in the training data.
What the research does describe is position bias correction. The paper "Learning to Rank in the Position Based Model with Bandit Feedback" addresses the fact that products shown in higher positions get more clicks and purchases simply because they are more visible — not because they are better. Amazon's models correct for this position advantage, ensuring that products ranked first don't get permanently boosted just because more people see them.
This is a meaningful and important correction, but it is correcting for position visibility — not for whether the customer arrived via an ad or an organic listing. The 2024 paper "Sponsored is the New Organic" (arXiv:2407.19099) further describes how sponsored and organic results increasingly share the same ranking infrastructure, with models jointly optimizing placement decisions across both result types.
Myth: "Verified Reviews Count More for Ranking"
Many SEO guides claim that verified purchase reviews carry more weight in the ranking algorithm than unverified reviews, and that you should prioritize getting verified reviews specifically for ranking purposes.
Amazon has not published any mechanism that distinguishes verified versus unverified review weight in search ranking. Review rating and review velocity are documented ranking signals, but no published research describes the ranking model giving different weights to reviews based on verification status.
What is true is that unverified reviews are more susceptible to Amazon's manipulation detection and removal systems, which are separate from the search ranking system. Amazon has published extensively on fake review detection (these are trust and safety papers, not search ranking papers). Unverified reviews are more likely to be flagged and removed, which affects your review count and rating — but this is a content moderation outcome, not a ranking weight distinction.
Focus on getting more reviews and maintaining a high rating. Whether they are "verified" matters for fraud detection, not for ranking weight per se.
Myth: "You Need to Stuff Keywords to Rank"
Keyword stuffing — cramming every possible search term into your title and bullet points — remains standard advice from many Amazon SEO practitioners. The logic: more keywords equals more retrieval opportunities equals more ranking opportunities.
What the research actually says: Keyword stuffing is actively counterproductive in modern Amazon Search.
Keyword stuffing hurts conversion rate, and conversion rate is a dominant ranking signal. Amazon's ranking models use behavioral features drawn from a pool of roughly 150+ features, selecting approximately 20 features per category through feature importance analysis. There is no "keyword density score" in the published feature sets.
Highly stuffed titles are harder to read. They look spammy. They reduce buyer confidence. This hurts click-through rate and conversion rate, which are primary inputs to the ranking model. You are trading a marginal retrieval benefit (which semantic search is already handling) for a measurable conversion penalty.
The research-backed approach: write listing copy that is semantically clear and conversion-optimized. Use natural language that accurately describes your product. Put relevant search terms in backend keywords where they aid retrieval without hurting the customer-facing experience.
What the Research Actually Confirms Matters
Based on what Amazon's own researchers have published, these are the high-impact signals within a seller's control, roughly ordered by evidence strength:
- Semantic relevance of listing copy — Your listing needs to genuinely match what customers are searching for. This drives retrieval (getting into the candidate set) and relevance scoring.
- Conversion rate — The dominant behavioral signal in ranking. Everything about your listing that affects whether a customer buys after clicking — images, price, reviews, copy quality, A+ content — feeds into this.
- Review rating and velocity — Higher ratings and a steady stream of recent reviews are documented ranking features across multiple papers.
- Price competitiveness — Price is a feature in ranking models and a major conversion driver. Being significantly overpriced relative to substitutes hurts on both fronts.
- Prime eligibility (FBA) — Prime badge affects conversion rate substantially, and FBA products have documented ranking advantages in multiple analyses of Amazon's systems.
- In-stock rate — Out-of-stock products cannot be purchased. Inconsistent availability disrupts the behavioral signals the ranking model depends on.
- Category accuracy — Amazon's ranking models are category-specific, with different feature weights per category. Being in the wrong category means being ranked by the wrong model with the wrong features.
The Algorithm Evolution Timeline
The real story of Amazon Search's evolution is documented in published research. Here is what actually happened, based on primary sources:
| Period | Development | Source |
|---|---|---|
| Late 1990s | Collaborative filtering ("customers who bought...") | Amazon's foundational recommendation patents |
| 2010s | GBDT learning-to-rank with behavioral features | "The Joy of Ranking Products" (SIGIR 2016) |
| 2018–2020 | BERT-based semantic embeddings for retrieval | "Semantic Product Search" (Amazon Science) |
| 2020–2022 | Multi-objective listwise ranking optimization | "Multi-Objective Ranking Optimization" (Amazon Science) |
| 2021 | Seasonal relevance signals in ranking | "Seasonal Relevance in E-Commerce Search" (CIKM 2021) |
| 2022–2023 | LLM-scale semantic search in production | "Web-Scale Semantic Product Search with LLMs" (2023) |
| 2023–2024 | Whole-page joint optimization (organic + sponsored) | "Sponsored is the New Organic" (arXiv:2407.19099) |
| 2024 | RL-powered query reformulation and understanding | "Exploring Query Understanding for Amazon Product Search" (arXiv:2408.02215) |
Notice what is absent from this timeline: version numbers. No "A9 to A10" cutover. No sudden algorithmic shift. Just continuous, incremental improvement driven by machine learning research — the way real production systems evolve.
The Real Picture Is Better Than the Myths
The actual Amazon ranking system is more sophisticated and more nuanced than any SEO blog suggests. It is not a keyword-matching engine with a sales velocity multiplier. It is a multi-stage, multi-objective ML system that combines semantic understanding, behavioral prediction, and continuous experimentation.
The good news for sellers: you do not need to reverse-engineer a fictional algorithm. You need to create listings that are semantically clear about what you sell, priced competitively, backed by genuine customer satisfaction, and consistently in stock.
That is not a secret. It is published research. The papers are free to read. And the strategy they imply — build a great product, describe it accurately, keep customers happy — is the one strategy that will never be disrupted by an algorithm update.
Because it is exactly what the algorithm is optimizing for.
- There is no "A10 algorithm." Amazon Search is a continuously evolving ML pipeline, not a versioned product. Claims about "A10" have zero basis in any published source.
- Keyword matching is no longer the primary retrieval mechanism. Amazon uses semantic embeddings and LLM-scale models that can match products to queries with zero keyword overlap.
- Sales velocity is one signal among many. Amazon's ranking system is multi-objective, balancing relevance, purchase likelihood, and other factors simultaneously.
- Ad-driven and organic sales are treated identically in ranking models. No published research describes weighting purchases by traffic source.
- Keyword stuffing is counterproductive. It hurts conversion rate, which is a dominant ranking signal. Write semantically clear, conversion-optimized copy instead.
- The signals that actually matter — semantic relevance, conversion rate, reviews, price, Prime eligibility, stock availability, and category accuracy — are all well-documented in published Amazon research.