Ranking & AI search
How Amazon Works · C1
A9/A10 → COSMO and Rufus: what Amazon rewards, how AI shopping changes listings, and what still moves rank.
How Amazon decides what to show — the classic ranking factors that still move rank, what COSMO and Rufus change about that, and how sellers are rewriting listings for AI shopping.
Amazon search now runs as two layers. The classic engine (A9) still builds the pool of findable products and ranks them by keyword relevance × sales performance — which is why an established brand went from organic rank #71 to #2 in five days by selling ~150 units against one keyword [L03]. On top sits an AI layer — COSMO's semantic model and Rufus (now merged with Alexa+ into "Alexa for Shopping") — that reads the whole listing, images and reviews included, and picks a winner for a shopper's mission, not a keyword. Amazon claims Rufus users are 60 % likelier to buy [L96]; it reaches ~100 million shoppers [L04]. Neither layer replaces the other; a listing must satisfy both.
Two layers, one page
A9 is Amazon's search-ranking system. Its aim is to make Amazon money by surfacing whatever is most likely to sell, so it weighs keyword relevancy (is the listing indexed for, and does its text match, the words typed) and sales performance (click-through, conversion, velocity) [L43] [L47]. Sellers fully control only relevancy [L43]. One practitioner models rank per keyword as performance × relevancy: relevancy scores 1 only when the phrase appears in exact form, ideally at the start of the title (plural or reordered variants earn less); performance is CTR, conversion and revenue; each click or purchase gives partial credit across related keywords, unlocked in proportion to relevancy [L79]. Organic rank is a trailing leaderboard of who sells for that term, not who is most topical [L47].
Where sources disagree — is there an "A10"? FBA Elite refers to "A9/A10" as the legacy engine [L53]; Andrew Bell, working from Amazon patents and science papers, says there is no A10 and never was [L04]. Trust Bell — the "A10" label was always a seller-community name, and both sources describe the same mechanism.
COSMO is Amazon's semantic layer: a knowledge graph built from queries, purchases and reviews that infers why people buy — the canonical example links "shoes for pregnant women" to slip-resistant, low-heel, arch-support purchases even when a listing never says "pregnancy" [L16] [L53]. Rufus is the shopper-facing assistant that sits on that graph. The clean framing: A9 asks "does the listing contain the words the customer typed?"; COSMO asks "does this product solve the problem the customer described?" [L53]. Keywords still matter for indexing and for building the candidate pool; COSMO changes how that pool is evaluated [L53] [L02].
Where sources disagree — does COSMO replace A9? Helium 10's AI-content episode says COSMO "replaces" A9 [L75]; FBA Elite, BDS and Orange Klik say it sits on top of A9 [L53] [L02] [L96], and a Helium 10 guest argues Rufus and A9 share one backend — Rufus builds "search query plans" and pulls candidates through the product search API [L65]. Trust the layered view: it fits the case where 1–2 purchases against a keyword moved both the A9 result and Rufus's answer almost simultaneously [L65].
What still moves rank
| Lever | Mechanism | Numbers practitioners give | Refs |
|---|---|---|---|
| Indexing | A keyword must appear once, anywhere (title, bullets, description, backend) to be indexed; the exact phrase must be present, not just its words | Backend field: 2,500 chars total, 500 per line; recheck indexing 15–60 min after an edit | [L43] [L47] |
| Title placement | Title carries the most weight; earlier words weigh more; intact phrases beat scattered words | "First five words" rule; keep the head phrase inside the first ~80 chars for mobile (~75 % of traffic) | [L43] [L84] [L16] |
| Repetition | Over-repeating a term hurts | ≤2 uses of a keyword in the title; >3 exact repeats anywhere in the copy can de-index the ASIN for that term | [L53] [L84] [L93] |
| Click-through | An Amazon Science paper is cited putting CTR ahead of conversion as a ranking input; main image, title, price and reviews are the CTR levers | #1 organic slot gets ~25 % of clicks / ~19 % of sales, #2 ~12 % / 8 %, #3 ~8 % / 5 % (2 M search terms, Jan 2026); top-of-search ad CTR ~9 % vs 1.3 % rest-of-search vs 0.2 % product pages | [L16] [L07] [L28] |
| Conversion + velocity | Amazon compares your CTR and conversion on a term against everyone else ranking for it — proven largely via PPC sales, "relative not absolute"; three drivers of organic rank: search-then-buy, ads on high-CTR placements, ads on high-conversion placements | CPR = units to sell in an 8-day window to reach page 1; every keyword has a "rank ceiling" set by how well the products above you convert on it | [L56] [L19] [L47] [L93] |
| Honeymoon | New in-stock listings get a temporary rank boost with no sales history | Start PPC the minute the listing is live; finish images/copy before launch | [L47] [L44] [L45] |
| Reviews | Rating acts as a floor, depth as a tie-breaker | 4.4★ is effectively the floor for positions 2–8; median ~7,700 reviews at #1 vs ~4,000 for #2+; only ~21 % of #1 products were cheapest and ~30 % most-reviewed — #1 is a "fit" decision | [L04] |
| Stock and delivery | Stockouts crater rank because Amazon deprioritises listings that can't promise fast delivery | Keep ~3 months of inventory | [L19] |
| Prime badge | Mobile app auto-applies a Prime filter, so non-Prime listings lose impressions silently | >70 % of buyers on mobile; badge lifts conversion 5–10 % | [L86] |
| Price | Discounts and "was price" resets buy velocity | Raise price ~10 % at a time to reset the strike-through baseline, then cut for a deal | [L19] [L47] |
Two implications. PPC's main job at launch is rank, not profit — a single-keyword exact-match campaign run deliberately at 65 % ACoS was the top sales driver in a $1 M launch because it bought organic position [L28] [L93]. And the fix order for a stalled listing is conversion first, indexing second: spend on a term whose page-one competitors out-convert you burns budget under a ceiling you cannot pass [L93].
The #71 → #2 case, and the sequencing it proves
A five-figure-a-month brand targeted "birthday gifts for women" (~340,000 searches/month; the top three products take ~39 % of clicks; only ~0.8 % of searchers buy anything). It sold ~150 units in five days (~30/day), moved from rank 71 to rank 2 by day four, stopped the ads after day five, and held #2 [L03]. The lesson is sequencing, not "buy rank". Intent keywords ("retinol under eye patches" — the shopper knows what they want) convert and get ranked first; research keywords (gifting, "best X for Y") are an awareness play that only works once reviews and conversion history exist — earlier is "an expensive way to light your money on fire" [L03]. One smaller relevancy nudge recurs: 1–2 real people searching the keyword and buying ("search-find-buy") is treated as legitimate; organising ~20 people to do it is a terms-of-service violation [L64].
Reading the scoreboard
The search page itself is the scoreboard [SS-03]: who holds the paid slots, who owns the organic ones, and what badges (deal, Choice, Best Seller) frame each click.
Three checks, in order: indexed (phrase present) → found (Helium 10 Index Checker or equivalent shows the ASIN for that phrase) → ranking (Keyword Tracker shows top-10/top-50 organic and sponsored) [L43]. Most sellers stop at publishing and never verify [L43].
The richest free scoreboard is Search Query Performance (SQP) in Brand Analytics (Brands → Brand Analytics → Search Analytics; needs Brand Registry — verify availability on .ae). Per search term it gives total and brand-specific impressions, clicks, cart adds and purchases; pull the last full month at least two weeks old, sort by brand purchases, read brand share down the funnel [L38]. Impression share ≫ click share is a thumbnail/title problem; click share ≫ purchase share is a listing problem; share that holds or rises means the term deserves more traffic [L38]. Brand clicks ÷ brand impressions is the only way to get organic CTR [L38]. Compare your ad conversion on a term against SQP's average competitor conversion to decide whether to push spend or accept the ceiling [L93]. Mid-tail terms carry the insight; head terms are high-volume/low-relevance, long-tail too thin [L38]. It is also how BDS measures any listing change — weekly CTR, conversion and position against category averages [L16], one variable at a time, two to three weeks per test [L02].
COSMO: from matching words to matching intent
What the AI layer reads is reconstructed from Amazon papers, patents and vendor tooling — treat specifics as practitioner claims, not documentation [L16] [L96]. The consistent picture:
- Escalating match tiers. Word meanings → semantics → inference → personalisation, so two shoppers can see a different "#1" for the same term [L75]; "Project Amelia" reportedly reorders titles per shopper, weighting the first ~80 characters [L16].
- Fifteen semantic relations. COSMO checks a listing against ~15 relations (what it is, used-for-function, used-for-event, used-for-audience, used-in-location, used-with…); a title should hit about 5, the full listing all 15 somewhere. A live audit of a beard-oil listing found it missing event, audience and location [L16]. ZonGuru's framing is the same "knowledge graph" of attribute nodes, with "used in location" the one most often missing [L96].
- Data sources. Catalog data, reviews, Q&A, curated external publications (reportedly not Reddit/Quora) and customer behaviour, retrieved live (RAG) and routed to different models via Bedrock depending on whether the query is a lookup or research [L96]. Missing structured attribute fields cause exclusion before scoring, not a penalty [L04]; the left-hand filter tick-boxes on a results page are effectively COSMO's attribute checklist [L53].
- Images and reviews are read. BDS demonstrated Amazon's own OCR (Rekognition) reading text on images and packaging [L16]; Helium 10 auditors describe Rufus scanning images for keywords and mining reviews for sentiment, and even advised dropping a secondary keyword list into blank space on a main image [L71] [L69].
Where sources disagree — do images index? Crescent Kao says A+ images "aren't indexed by search" [L47]; My Amazon Guy says Amazon is moving away from alt text toward crawlable text [L89]; BDS and Helium 10 say the AI layer OCRs image text [L16] [L71] [L75]. Reconcile: text in images may feed the AI layer, but nothing you need for A9 indexing should live only in an image. Put every fact in crawlable text and use image text to reinforce it.
Rufus: the mechanism and where it shows up
What it is. A generative shopping assistant in the app and on desktop that answers natural-language questions, compares products, tracks price history, adds to cart and — via "auto buy" price triggers and "buy for me" — completes purchases [L53] [L96] [L09]. In 2026 Amazon merged it with Alexa+ into "Alexa for Shopping" and began licensing it through AWS [L01] [L06]. Adoption figures vary: 250 M users in 2025 and $10 B projected 2026 sales [L96]; 300 M+ users and ~$12 B incremental 2025 sales [L09]; 38 % adoption [L13]; 3.5× Black Friday conversion and 66 % of purchases in one study [L07]. Take the direction, not the decimals.
Where it appears (US — verify for .ae). Search-bar autocomplete, Q&A snippets in results, the chat box, questions under the main image, a pre-purchase cart question [L75]; suggested prompts on results pages and an AI "What are customers saying?" summary on detail pages [L53]; "Ask Alexa" predicted questions [L89]; and for broad queries ("men's skin care" vs "men's lotion") a curated set labelled "researched by AI" [L07]. Rollout is uneven — mobile-only in some markets [L75] — so check what amazon.ae renders.
How it decides. A9 retrieves; the AI layer interprets the shopper's mission, budget and context, weighs page evidence, and picks [L02] [L04]. A single prompt fans out into sub-searches by room, style, budget, material, recipient, use case — the "query plan" — so eligibility across that family matters more than rank on one term [L02] [L04]. It answers from the whole page (title, bullets to #10 for brand-registered sellers, description, A+ text, lifestyle images) plus the open web — an outlet's review copy can win the citation over your listing [L04]; its most-cited off-Amazon sources are earned media (~43 %) and affiliate review sites (~40 %) [L07]. It surfaces negative reviews and cross-brand comparisons [L75] and 1-month/3-month/1-year price history [L04] [L65]. It learns from outcomes: claim a benefit the product lacks, get returns, and it stops recommending you for that intent [L53]. It re-indexes listing changes slowly and its questions change rarely — a monthly audit is enough [L75].
Ads inside Rufus (US — verify for .ae). Sponsored Products/Brands "Prompts" auto-generate questions from your detail page and answer them with your ad inside Rufus; the beta was free and auto-enrolled — one account: 35 prompts across 18 questions, ~5,000 impressions, 41 clicks, $0 spend, four orders worth ~$513 in six weeks [L11] [L53]. A multi-touch attribution toggle now credits every ad that assisted a purchase [L02].
How sellers are adapting listings for AI search
The channels converge on one rule — keep the A9 foundation, add an intent layer — and differ on how far to go.
- Rebuild the title as a noun-phrase stack, ~75 characters. Head noun + material + use case + constraint + proof, readable by a human ("dimmable metal floor lamp for reading nook"), not 200 characters of jammed keywords [L02] [L04]. Amazon's own move to a 75-char item name + 125-char "item highlights" field (US rollout July 2026 — verify for .ae) points the same way; Amazon's AI split preserved most keywords but broke "coffin gift box" into "gift box", so keep phrases intact [L63] [L91].
- Constraint/proof phrase in the first two bullets, audience/occasion line somewhere. Agents filter on hard constraints before comparing; most listings omit who the product is for [L02]. Map every feature to an outcome chain ("304 stainless steel → what does that give me?") [L53] [L04] [L14].
- Answer predicted questions in crawlable text. My Amazon Guy's "AEO" phase: type your top phrase into search, read the AI-generated questions, answer each in a bullet or an A+ FAQ module; after edits one ASIN started appearing for two more prompts [L89]. Helium 10's workflow harvests Rufus's follow-up questions and turns them into an FAQ image, objection callouts, care-instruction and dense-spec images, and an "us vs them" slide [L75].
- Fill everything. Every attribute field, up to 10 images and 10 videos, storefront pages; register the brand to control "product truth" [L75] [L04].
- Audit against the 15 relations, prioritise with RICE, change one thing at a time so SQP can attribute it [L16].
- Build off-Amazon evidence. Earned media, comparison content, mentions on 10+ credible sites feed the AI layer, which is easier to influence in niches it knows little about [L07] [L96].
Where sources disagree — how urgent? ZonGuru claims un-updated listings slide after ~6 months [L96] and BDS frames COSMO literacy as the price of visibility [L16]; FBA Elite says keep the keyword foundation and layer intent on top [L53]; BDS itself notes only 6 % of Gemini conversations carry purchase intent [L01]. Trust the moderate view: the intent work is cheap, also lifts conversion, and loses nothing if AI adoption on .ae lags the US.
What this means for SHIO
- Do pick the exact head phrases shoppers type — "filtered shower head", "shower head filter", "shower filter for hard water" [SS-03] — put the strongest verbatim in the first five title words, and verify each phrase indexes and is found; ≤2 uses per phrase in the title, ≤3 anywhere [L43] [L84] [L93].
- Do write for COSMO's relations: what it is; function (removes chlorine, counters hard water, boosts pressure); audience (hair loss, dry skin, sensitive scalp, families); location (UAE apartments, hard-water areas); used-with (replacement vitamin-C/carbon cartridge); event (moving in). Five in the title, all across the listing; fill every attribute field (material, flow modes, hose thread, filter life) [L16] [L04].
- Do put a hard-constraint/proof line in bullets 1–2 ("fits standard ½-inch hose, tool-free", "activated carbon + vitamin C, replace every ~N months") and a care-instruction image — filter-life complaints are exactly what Rufus surfaces [L02] [L75].
- Do mine the results-page filter tick-boxes and competitors' "What are customers saying?" summaries for attribute language; if Rufus is live on .ae, ask it your own listing's questions monthly [L53] [L75] (US — verify for .ae).
- Do sequence: intent keywords first, reviews to a 4.4★+ base, then research/gift terms; go live only with finished images and A+ so the honeymoon isn't wasted; ads from day one [L03] [L04] [L47].
- Do be on Prime/FBA if amazon.ae mirrors the mobile Prime auto-filter (US — verify for .ae) [L86].
- Do pull SQP monthly if Brand Analytics is available on .ae; treat impression→click gaps as image/title work and click→purchase gaps as listing work [L38] (verify availability).
- Don't claim benefits the product cannot deliver (e.g., "cures hair loss") — returns teach Rufus to stop recommending you [L53].
- Don't rely on text-in-images or backend stuffing for indexing; put every fact in crawlable copy [L89] [L47].
- Don't assume Sponsored Prompts, "researched by AI" curation or the 75-char title format exist on .ae — the corpus only shows them in the US [L11] [L07] [L63].
The concept graph behind this chapter
While building the corpus, lore filed the ideas above as pages of a knowledge graph — each link opens that concept's own page, with its definition, the videos it came from, and how to apply it. The 36 most relevant of the subject's 385 published concept pages:
Amazon A9 Algorithm · Day Parting (Amazon PPC) · Amazon Search Term Report · Amazon Negative Keywords · Amazon Search Query Score · AWD (Amazon Warehousing and Distribution) · Rufus (Amazon's AI Shopping Assistant) · Amazon Listing Optimization: SEO vs. Conversion Optimization · Amazon Search Query Performance (SQP) Report · Amazon Seller 10-Step Launch Process · Cosmo (Amazon's Rufus-Powering Algorithm) · Amazon as the True Bottom of Funnel · "Amazon Escape" 23-Criteria Product Funnel · Five Reasons Amazon FBA Products Fail · Google vs. Amazon Search Intent (Informational vs. Transactional) · Amazon Brand Storefront as Ad-Free Conversion Destination · Amazon Generic Keywords (Backend Search Terms Field) · Branded Search-Lift & New-to-Brand Attribution for Off-Amazon Marketing · Fixed Bids (Amazon PPC Bidding Strategy) · Amazon Virtual Bundles (Buy Box Cross-Merchandising) · Advisor-Led vs. DIY Amazon Brand Sale (Fee Alignment & Risk) · Amazon Gated-Category Documentation Requirement · Amazon Product Life Cycle (Launch, Expansion, Harvest) · Master Keyword List & Listing Scorecard · Amazon-Only vs. Omnichannel Seller Skillset Framing ("Inside vs. Outside the Box") · Amazon Creator Connections (Influencer Affiliate Program) · Two-Tier Keyword List Strategy · Specialist-Partner Team Structure (Sourcing + Meta + Amazon) · PPC-to-Organic Rank Mechanism · Keyword-First Product Research (Starting from Keyword Demand) · Per-Keyword Rank Ceiling & SQP Competitor Conversion Benchmarking · Keyword Segmentation via Product Variant Listings · Helium 10 Keyword Tracker · Review-Before-Rank Launch Sequencing · Main Keyword & Organic Listing Validation Protocol · CPR (Units to Rank) Metric
Source legend
Not covered / open questions
- Rufus/Alexa for Shopping on amazon.ae. No source in this corpus discusses non-US or non-UK rollout; whether the assistant, "researched by AI" curation, Sponsored Prompts, price-history display or the 75-char title format are live on .ae is unverified — see [C8].
- Brand Analytics / SQP availability on .ae and whether Brand Registry there unlocks the same reports.
- How much weight COSMO's signals carry today versus A9's — every source describes mechanism, none quantifies the AI layer's share of traffic on a real listing; the only outcome numbers are vendor-sourced (ZonGuru) or single-account (My Amazon Guy's "+2 prompts").
- Whether image text is truly indexed by A9 (as opposed to read by the AI layer) remains contested across sources.
- Localisation for the AI layer — Arabic listing content, mixed Arabic/English queries, and whether COSMO's knowledge graph is marketplace-specific are not addressed anywhere in the corpus.
- Interaction with advertising — how Rufus curation affects Sponsored Products impressions is flagged as unresolved even by the sources [L09]; covered from the ad side in [C3].