Amazon’s Punitive Algorithms Outlined.
Seller Performance uses reward based and punitive based algorithms that act on Seller Accounts. Frontside work is visible on Amazon.com and its ROI is measurable.
In a punitive based environment, It’s a blend of intuition, experience and caginess.
The Acronyms and definitions that I use here are ours. We theorize that these algorithms exist and what their components are. We also speculate on the advancements that these algorithms that we’ll see in the next 90-180 days. We also have ideas on how these will work in the future. The future in Amazon language is 12-18 months.
We have no confirmation that anything that we say here is accurate or truthful.
BAB: Bad Actor Behavior AI
A tremendous amount of fraud is perpetrated on Amazon. It’s a global platform and criminal elements with significant capital and intelligence are a constant threat. BAB attempts to find then using signals from patterns of behavior. These can be patterns that match similar behaviors of previous BAB seller accounts. Any account can have BAB deployed at it. Many of the signals that Trigger BAB are common in large accounts, fast growing accounts, and recently suspended accounts. The algorithm is using a probability score and a seller maturity score to trigger an investigation or another event.
The most sophisticated forms of BAB are sellers using Reverse AI to generate adversarial data to alter the component weights within Amazon’s AI. These are organizations that write their own algorithms, and work quickly to offset the signals that trigger BAB. There are other low-tech forms of reverse AI.
TRM: Transaction Risk Management AI
The BAB algorithm can trigger this. The algorithm looks hyper-close at your AOV, Order Velocity, Buyers, CSX data, Fulfillment, Buyer/seller messaging to determine Amazon’s financial risk if you are in fact a Bad Actor. TRM and BAB enforcement generally bypass Seller Performance. The Business and Legal Teams own FINAL-WORD on enforcement and reinstatement decisions. They also make other decisions such as keeping any or all your money and products, and potentially suing you.
How Have we seen innocent sellers suspended for BAB or TRM? If your succeeding, you are generating signals that contribute to the probability score. We’ve seen successful sellers change the EDD (estimated day of delivery) simply because they were on religious observance and needed to add days to the EDD post-order. This behavior coupled with AOV, time on platform, buyer-seller messaging was enough to trigger a suspension. There are a multitude of behaviors that may seem benign to the seller, but they can trigger false-positive signals
In these suspension situations, where you are dealing with Business & Legal, ask yourself, do you want the Amazon Seller Association on your bench in moments like these? That seems like a good idea to me.
HnR: Hit & Run AI,
GSGB: Good Seller Gone Bad AI. Sellers Acquire maturity scores. Within those scores are degrees of elasticity that sellers can use to defraud Amazon in a short window. Because we have an inordinate amount of failing private labels it’s a common problem with a dedicated Algorithm.
SMB’s in general have a high fail rate, on Amazon, 75% of private labels, fail. These are the common seller profiles of HitnRun, GSGB. Other’s are virtual identities who have a HnR on their agenda from the start.
We believe that all three of these algorithms, (BAB, TRM and HnR) are fully deployed into every selling account. And that every selling account has a dynamic and a separate constant probability score. We speculate the dynamic probability scores push you into watch-list algorithms that are hypersensitive to certain account values.
Amazon has a 2% return rate across its platform. Its higher in apparels and shoes than it is in household or personal care products, but on average, most sellers will deal with a 2-6% return rate even in the most optimum delivery experience. We call it reverse logistics.
It’s important to remember that when a buyer purchases using Prime, Amazon also handles the returned product. For every million orders of 3rd party products, Amazon refunds, and accepts the return of a minimum of 20,000 products that have had a less than perfect order experience. <POE
This image represent a refund flow in a 1000 unit a day selling account. Whats important is that Amazon controls this funnel. Amazon receives it from the seller, they store it, they feature it for sale, they pick it, they pack it, they ship it, they customer serve, they process and receive returns, they grade returns, and they make final disposition on cause of the <POE
Examples of Non-seller <POE’s are: Damaged by carrier, Damaged in Amazon’s FC, Late delivery and other factors that are non-seller related. These are given reason codes that Amazon uses to score itself, its partner carriers, and all its systems and processes. Amazon’s Goal for each buyer is a perfect order experience. POE
Once the AI determines it is a seller based <POE, it then determines a reason code.
Reason Code Policy Violations.
USN: Used as New. Products listed in a condition of New cannot have dents, dings, fingerprints, missing parts, open box, missing manufacturer warranty and/or registration cards, packaging abrasions,
NAA: Not as Advertised. The product detail page is claiming something that is demonstrably inaccurate.
NasD: Not as Described. The product detail page has a red shoe, and you get something closer to pink.
EXP: Expired. Same as NEW. New assumes not expired.
MatDiff: Materially Different. All apply. A buyer orders a pack of 6 deodorants and gets 5, or a buyer orders a towel and gets a hockey stick. These both get hit with MattDiff. Mattdiff is a big wide sweeping algo’ that has advanced very quickly in recent months. The example that I gave is actually not true any longer. We see the Algo’s doing more reasoning and using more human logic. We think they’re using NODE-Specific logic. In this case the ALGO’s would grade the , Got 5 materially different, punitively lower than the , expecting towel and got a hockey stick materially different”.
CR: Inauthentic AI A buyer has authenticity concerns. Over 3% in a 50 order sample of 10,000 orders can and does trigger. Hyper-sensitive trigger.
CCrT+: Counterfeit True Positive AI. The brand, its agents, or even a buyer, can use a verified order to trigger a T+ CCr.
PQ_P: Product Quality AI You buy a folding table and put 20 pounds on it and the leg breaks. You buy a bird feeder , and after 60 days in the Sun the label or coloring is gone. You buy an expensive item with high expectations and it fails your expectation. Even if you bought it 6 months ago, even if Amazon or the seller will not refund you because 6 months ago. To the seller, Product Quality complaints are never tolerated by amazon. Even the buyer may not have been refunded because it was six months, Amazon can and frequently does raise PQ_Product Quality Policy Violation and expects a seller to answer for any POE it determine. The age of the defect is irrelevant from a policing point of view.
PQS: Product Quality Safety. Any negative physical reaction, any side effect, any increased danger: electrical, choking hazard, fire hazard. Any adverse reaction to the ear, nose, eye or throat. The signals that trigger PQS are different within Browse Node’s. The Browse node of 13213828011, Infant Dental Care, has its own dedicated deployment. It works off of data specific to that node. The team that handles this is well trained in this vertical. They have access to FDA Data, Recall Data, Public Health Care Databases, Ingredient expertise and more. Because as the Algorithms have gotten better at identifying node specific data, it makes smarter and quicker decisions. As a result Seller Performance is able to funnel it to capable teams. It works, its gotten better at lightening speed in the last 6 months. I believe Product Safety is a Major investment for them. No stone unturned in pursuit of healthy perfect order experience.
CoC: Code of Conduct AI.
Virtually every seller abuses Amazon systems. None more than the Private Label Sellers and their faux-experts. Catalog abuse is rampant. Amazon and Seller Performance catches up to everything though. Many sellers and just as many “experts” are currently being face-planted by CoC. And its about to get a whole lot worse. Easily, 65% of private labels that joined Amazon the last 24 months, will not be here 12 months from today. Did you catch this?
Code of Conduct AI is fully deployed into every seller account in the same way that BAB, TRM and HnR are. Its constant, and triggered by probability scores centered on the front side, such as reviews and variation abuse. But also in the back end when dealing with automated reimbursement systems.
CoC is not something that can be explained without highlighting the mechanisms that trigger the policy violations. The areas where CoC focuses are shown here.
Caution: If you are using any type of automated system for reconciliation. Anything at all, returns, removals, FBA refunds, Inventory, and more, you are throwing of BAB and TRM signals very close to the Mendoza line already. Be very careful with automated process that center on reimbursement or fees. Its among the most easily triggered BAB signal.