I’ve spent the last two weeks reverse-engineering the methodology behind that claim. The source is a nearly five-minute video (281 seconds of ASR transcript) from a Douyin creator known in Chinese cross-border circles as “蟹老板” (Crab Boss). The content is dense—eight distinct dimensions, each with hard numbers attached. What follows is a structured breakdown of every dimension, the exact thresholds, and what they mean in practice for a refined-selection operation.

The Problem with “Feel-Based” Selection

Before the framework, the context matters.

Most small-to-mid-sized Amazon operators run what the Chinese ecommerce world calls rush-deployment: list everything, see what sticks, double down on winners. It’s high-volume, high-waste, and extremely sensitive to headcount. You need people for every listing, every update, every price change.

The refined-selection model is the opposite philosophy. Fewer products. More diligence on each. Lower review-count dependence. Higher hit rate per launch. The tradeoff is upfront analysis work: you have to be right before you commit inventory.

The promise is a consistently above 90% success rate. The question is what makes it reproducible.

The Eight-Dimension Framework

Dimension 1: Precise Traffic Keywords

The first dimension is foundational. Every product must have a “precise traffic keyword”—a keyword that accurately describes the product’s function, material, attribute, use case, or target user.

Examples: “wooden under-bed storage bin with wheels”, “40-lb metal mug”. These aren’t generic categories. They’re specific enough that a search returns direct competitors, not adjacent products.

Why this matters—four reasons:

1. Find direct competitors. Only precise keywords return a clean competitive set for analysis.
2. Simpler ad architecture, better conversion. Generic keywords scatter spend across unrelated placements.
3. Enable Helium 10 reverse-search. With precise keywords, you can use Jungle Scout or Helium 10’s reverse ASIN lookup to estimate traffic cost and project ROI.
4. Define the market. A “small but no keyword” product isn’t a blue ocean—it has no market. If customers can’t find you and you can’t find customers, there is no product.

The hard rule: No precise keyword = no product. Period.

Dimension 2: Operational Difficulty

This dimension measures how hard it is for a new, low-review listing to compete in a given subcategory. The metric is a ratio:

Low-review top-5 average sales ÷ Top-5 average sales > 20%

Translation: take the five best-performing listings with 30 reviews or fewer. Average their daily sales. Divide by the average daily sales of the top five overall listings (regardless of review count). If the result exceeds 20%, the subcategory is “new-listing friendly.”

What the ratio tells you:

Supplementary signals for operational difficulty:

• If low-review listings are already generating sales → market is not saturated
• If review-free “裸奔” (naked) listings are also converting → market has genuine demand gaps
• If FBM (Fulfilled by Merchant) listings are moving steadily → demand is broad and robust
• If no obvious price/review wars → competition is still healthy
• If the subcategory has no aggressive review acquisition → advertising can carry new products

Dimension 3: ROI and Capital Efficiency

The ROI threshold is ≥ 120%. Products below this line don’t get through the gate.

But ROI does two jobs here. It’s not just a profitability measure. It acts as a market health indicator.

First function: market health signal. If a large proportion of competitors have poor ROI, that subcategory is already overcrowded. High ROI across competitors means the market is still dividing profits healthily.

Second function: capital efficiency. For cross-border ecommerce, the cash conversion cycle is brutal. You pay suppliers in RMB, ship to Amazon warehouses in the US or Europe, and wait 30–60 days to collect USD. Every dollar tied up in FBA inventory is a dollar not working elsewhere.

The cost model Huang uses:

Total AMZ cost = Procurement + Freight (head freight) + FBA fulfillment fees + Platform commission

These are not overhead. They are embedded in the selling price—they come off the top before margin. Higher capital turnover efficiency directly multiplies effective return on investment.

The rule: A 120% ROI on a 90-day cash cycle beats 150% ROI on a 180-day cycle. Factor the cycle, not just the percentage.

Dimension 4: Differentiation and Supply Chain Fit

Differentiation is where most operators cut corners—and where most get caught.

Differentiation paths, in order of execution:

1. Product-level modification: Adjust specifications, materials, or packaging for a clear point of difference
2. Bundle strategy: Combine products in ways competitors haven’t
3. Variant planning: Offer color, size, or configuration variants that create listing depth
4. Repurpose the product: Find new user groups, usage scenarios, or use cases for existing products
5. Creative layer: Visual differentiation (images, lifestyle shots), copy differentiation (tone, framing), even review differentiation (encouraging specific review angles that attract a niche)

The last point is underappreciated. Platform algorithms respond to review content. A cluster of reviews that all mention a specific use case tells the algorithm something—and it allocates accordingly.

The supply chain checkpoint is non-negotiable. Before committing to a product, you must verify:

• A capable factory exists and can produce the required specifications
• The factory can achieve the required quality consistency at the target cost
• The MOQ (minimum order quantity) aligns with your first-shipment budget

If any of these fail, the product goes to a “backup list.” It may be a good product but an un-executable one under current conditions.

Dimension 5: Entry Timing

Amazon subcategories have peak and off-peak seasons. Entry timing is the fifth dimension, and the standard rule is:

Enter 1–2 months before the peak sales season.

The logic: new listings need time to accumulate reviews, build listing authority, and develop ad campaign history. By the time peak traffic arrives, the listing should already be positioned. Riding the wave is fundamentally different from trying to establish yourself in it.

The forward-looking seasonality rule: Operators evaluate products based on the next three to five months, not current-season data. A product that’s doing well right now—particularly if it has low review counts and high sales—may be a seasonal or event-driven spike (Cinco de Mayo, July 4th, Teacher Appreciation Week, Mother’s Day). These peaks are already missed by the time the data appears.

Seasonal/event products specifically: If you want to play a seasonal product, you need to enter the market 4–5 months in advance. Research, selection, sourcing, shipping, and listing all have to happen before the window opens.

Dimension 6: Seasonal, Event, and Trend Products

Building on the timing dimension: the system flags whether a product is seasonal, event-linked, or trend-driven.

The detection signal: Low review counts + high sales volume + sudden appearance in rankings. This combination almost always means the product is riding a current event or trend (Cinco de Mayo, Independence Day, back-to-school, etc.).

The rule for event products: Research 4–5 months ahead, select 4–5 months ahead, ship 4–5 months ahead. The refined-selection model can do event products, but only with a long enough runway.

The warning: Event products that look attractive in real-time data are almost always already past entry point for the current cycle. The data lag is structural—by the time you see the spike, the window is closed.

Dimension 7: Policy and Compliance Risk

This dimension is where operators lose everything.

Compliance risk has three categories:

1. Patent infringement Standard but non-negotiable. Clear the product with a patent search before committing.

2. Copyright infringement The TRO (Temporary Restraining Order) minefield. Copyright protection in the US extends broadly:

• Books, publications, literary works
• Artwork, paintings, illustrations
• Film/TV IP, character names, film/TV titles
• Software, programs
• 3D design files, photographs, sculptures

If your product or its listing includes any of these without a license, you are at risk. Forcing a settlement from a Chinese company via TRO is a known tactic among rights holders.

3. Platform policy violations This is more dangerous than IP infringement because it can result in immediate store suspension or termination. Examples include:

• Electronics with safety implications
• Disallowed product categories (certain weapon categories)
• Products restricted for certain seller types
• Products involving personal safety, property safety, or financial safety

Products that touch platform-restricted categories need to be evaluated before design finalization, not after.

The workflow: Product design finalization → IP clearance → compliance review → listing. Not the other way around.

Dimension 8: Inventory Quantity and FBA Warehouse Layout

The final dimension manages inventory risk.

First-shipment quantity calculation:

Reference a direct competitor with:

• ≤ 30 reviews
• Low ad spend
• No listing merge history
• No heavy ACoS (Advertising Cost of Sales) campaign

Take that competitor’s average daily sales. Use 50% of that figure as your first-shipment quantity.

This is deliberately conservative. It leaves room to reorder based on actual market response rather than guessing.

Seasonal demand coefficient adjustment: The baseline figure is then adjusted based on:

• Where the product sits in its seasonal cycle at launch
• Historical sales data for the same product category in prior years
• Demand volatility coefficient derived from historical data

The goal: avoid both overstocking during low season and stockout during peak. Neither is free.

The FBA multi-warehouse layout strategy:

Sales inventory     → Primary fulfillment warehouse
Backup inventory   → Secondary warehouse, holds reserve stock
Replenishment inventory → Pre-positioned for fast restock

Splitting inventory across multiple warehouses:

• Prevents single-warehouse incidents from taking down the listing
• Reduces stockout risk from unexpected demand spikes
• Provides flexibility to shift allocation based on sales velocity by region

The equilibrium rule: Conservative first-shipment quantities make it easier to hit the equilibrium point where sales volume and ROI are both positive. Aggressive first shipments work against this equilibrium—overstock absorbs storage fees while understock misses peak conversion.

The Synthesis: Low-Review-Dependent, Differentiated, Precise-Keyword, Compliant, Cyclical

The eight dimensions converge into a single operational philosophy:

Low review dependence + micro-differentiation + precise keyword targeting + compliant variation + cyclical timing

The “low review dependence” part is critical. Most new Amazon sellers assume they need reviews to compete. The refined-selection model argues the opposite: select products where the market structure doesn’t punish low reviews, and reviews become secondary to keyword relevance, listing quality, and ad structure.

The framework is designed for the operator who wants to build a business that doesn’t require massive headcount. Fewer products, more analysis upfront, lower operational risk per unit.

Key Data Thresholds Summary

FAQ

Q: What is the most important dimension in Amazon refined-selection product selection?

A: Precise traffic keywords is the foundational dimension—everything else depends on it. Without precise keywords, you cannot identify a clean competitive set, build a targeted ad architecture, or verify that a real market exists. No precise keyword = no viable product, regardless of how good the other seven dimensions look.

Q: How do you determine if an Amazon subcategory is new-listing friendly?

A: Calculate the ratio of average daily sales for the top 5 low-review (≤ 30 reviews) listings divided by the top 5 overall listings in the subcategory. If this ratio exceeds 20%, the subcategory is friendly to new listings. A ratio below 10% indicates a review wall that will be difficult for new products to penetrate.

Q: Why is 120% ROI the minimum threshold for Amazon product selection?

A: The 120% ROI threshold (1.2x return on ad spend plus all embedded costs) accounts for the long cash conversion cycle in cross-border ecommerce, where capital can be tied up for 30–90 days. A product delivering 150% ROI but requiring 6 months to convert cash may underperform a 120% ROI product with a 45-day cycle when capital efficiency is factored in. The threshold also acts as a market health filter—subcategories where most competitors can’t hit 120% are typically overcrowded.

Q: How does entry timing affect Amazon product selection success?

A: Entering the market 1–2 months before peak season allows time for review accumulation, listing weight building, and ad campaign history development before high-traffic periods arrive. Products are evaluated based on their performance potential 3–5 months forward, not current real-time data—which often reflects seasonal or event-driven spikes that are already past the entry window.

Q: What are the three categories of compliance risk in Amazon product selection?

A: (1) Patent infringement—design and utility patents. (2) Copyright infringement—the widest risk category, covering artwork, publications, film/TV IP, character names, software, 3D files, and photographs. (3) Platform policy violations—which can result in immediate store suspension and are more dangerous than IP infringement because they carry no warning period.

“Before RPA, we had 170 people doing maybe 100 million RMB in revenue,” Huang Xufeng told me over a video call from his office in Shenzhen. “Today we have 50 people and we’re doing 300 million.”

The math doesn’t add up the way most people assume. This isn’t a story about replacing humans with machines. It’s about understanding which human work is fundamentally automatable, and which requires actual judgment—and then building the infrastructure to separate the two.

Huang is a 14-year veteran of cross-border e-commerce. He sells on platforms across 187 countries, manages roughly 300,000 SKUs (of which about 30% have ever generated an order), and has been through enough cycles to know what “scaling wrong” looks like. “I lost tens of millions in my early years,” he said. “Because I was young and felt invincible. I wanted to try everything.”

The transformation began in April 2023, when he saw a video about RPA—a robotic process automation tool—and bought a license the next day.

The Three Bottlenecks That Were Killing the Business

Huang breaks down the cross-border e-commerce operation into three layers that every player in this space has to manage: sourcing and selection, content creation and listing, and publishing and post-sale. Each layer was its own kind of hell.

Bottleneck 1: Human Error and Scalability

The first problem was foundational. Every platform listing required 60+ steps—screenshots, field entries, pricing calculations, logistics estimates. Huang had built a 500-page SOP document to train new employees. Three hours of self-study and they could do the job.

Except they couldn’t. Not really.

“The SOP solved training speed, but not accuracy,” Huang explained. “If someone mistyped a price, that was a direct loss. We carried those losses. Management overhead was enormous.”

The math: when you have 200 stores, each potentially carrying 3,000 products, you’re managing hundreds of thousands of SKUs. At 1,000 new listings per day, the operational risk was constant and compounding.

Bottleneck 2: High Employee Turnover

The work was basic. Input numbers, click buttons, follow steps. It required no judgment, no creativity, and offered no growth path. Employees came, learned, and left—sometimes within weeks. Every departure meant lost institutional knowledge and renewed training costs.

“Ideally you’d want people who could grow into the system,” Huang said. “But the work didn’t allow it. The machines weren’t ready yet to take over, and the humans were stuck doing machine work.”

Bottleneck 3: Decision Lag

When revenue spans 187 countries and hundreds of thousands of products, data is supposed to be king. But the data was slow. Really slow.

“Before our AI dashboard, if the boss wanted to know why a store’s numbers dropped today, the process was: check ERP, find the order anomaly, forward to the relevant manager, manager investigates, traces back to the specific listing, identifies the cause, reports back up the chain.”

Huang pauses. “Minimum decision cycle: four to six hours. And if the boss was busy that day, maybe it got addressed tomorrow. Or never.”

The Solution: A Three-Layer Automation Stack

Huang’s fix wasn’t a single product or a single “AI solution.” It was a layered approach, each layer solving a different problem.

Layer 1: RPA for Repetitive Operations

The breakthrough came from an unlikely place—a video his junior college classmate sent him about a Japanese RPA tool called UiPath. Huang bought it the next day and started figuring it out himself.

But here’s the part most companies miss: he didn’t just hire an RPA specialist to handle it all. Instead, he had all 160 employees get UiPath certified.

“We made everyone study,” Huang said. “Watch videos, practice, take the exam. About 9 hours per person. Because I realized the bottleneck wasn’t the RPA developer. It was that non-IT people couldn’t articulate what they needed.”

By certifying everyone, he suddenly had 100+ people who understood the boundaries and possibilities of the tool. “Suddenly we had 100 people who could actually submit useful automation requests. That’s when it started working.”

The first wave: automate the repetitive, standardized tasks that had high turnover and high error rates. One click, automated operation, running 24 hours. The effect was immediate—operational load dropped significantly.

Layer 2: AI Dashboard for Real-Time Decisions

The second layer addressed the decision lag. Huang used RPA to pull data from all platforms into a multi-dimensional spreadsheet, then built a BI dashboard that automatically pushed alerts to WeChat.

Not “you can check the dashboard.” Push. “Look at this listing. Its conversion rate dropped. Fix it.”

The difference between a traditional BI system and an AI push system is subtle but massive: passive monitoring vs. mandatory notification. “Before, you had to actively look. Now, whether you look or not, it comes to you. You have to see it.”

Decision cycle: from 4-6 hours to real-time.

Layer 3: Workflow Systems for Product Generation

The third layer was more sophisticated. Huang built an internal web interface that accepts a single product photo and generates a complete set of listing images—white background, size charts, model shots, detail shots.

“Different stages of the workflow use different models,” he explained. “Some stages are just background removal—any free tool handles that. Other stages need more complex generation. In total, maybe 2-3 RMB per finished set of images.”

The whole system is integrated with Feishu (Lark), so staff can upload a photo and get back a complete set of ready-to-use product images within minutes.

The Org Structure Change: From Vertical to Horizontal

The automation didn’t just change the work—it changed the organizational structure.

“Before, we had about 20 platforms, each with its own leader vertically downward. If I wanted to launch 20 platforms, I’d need 20 leaders.”

Now: one leader per function (selection, publishing, post-sale) across all platforms. Because RPA handles platform-specific execution, and AI dashboards handle cross-platform visibility, you don’t need a dedicated human for each platform anymore.

“Within our system, we only need to抓住几个点—the key leverage points. Selection is still human. After that, once the product is chosen and assigned to a platform, the rest is automated.”

The IT Team Structure

Huang’s internal IT organization is unusual. It has three tracks:

1. RPA BP (Business Partner): Ensures program stability across operations.

2. Database: Breaks down data silos, connects platform APIs, builds the company’s own data infrastructure.

3. AI Selection Officer: A dedicated role—budget capped at 20,000 RMB/month—whose job is to research emerging AI tools, test them, and produce structured reports on what works and what doesn’t.

“This person tells me: ‘Here’s what Kling can do. Here’s what Sora can do. Here’s what it would cost to implement. Here’s the expected efficiency gain.’” Huang calls this “preventing blind trial and error.”

The 85% Mark and What’s Left

I asked Huang where he currently stands on the automation journey. “About 85%,” he said. “The remaining 15% is both a technical壁垒 and a mental one.”

His 2026 goal is 95%. The path forward is using tools like the recently released OpenAI Operator or similar enterprise-grade agents to build what he calls “an OS for cross-border e-commerce.”

“Imagine a single interface where I tell the AI: ‘I want to launch this product on 5 platforms, with these margin targets, and I want a report in 3 months.’ The system then autonomously handles: sourcing data, publishing listings, forming ad campaigns, tracking performance, and reporting back. If it misses the target, it self-corrects.”

The catch: this requires the company’s own foundational systems to be AI-ready first. Data cleaned. SOPs documented. Knowledge bases built.

“You can’t just throw an agent at a messy operation and expect magic,” Huang said. “That’s the mistake many small businesses make. They want the big lobster to fall from the sky and solve everything at once. It doesn’t work that way.”

The Mistakes Small Businesses Make with AI

When I asked about common pitfalls, Huang was direct: “Most business owners don’t want to start from the smallest thing. They want a big solution from day one. But AI adoption doesn’t work that way.”

His prescription: find the single most painful point in your operation, document it as an SOP, then use RPA to automate that one thing. “Find the feeling of breakthrough. Once you see the computer working for you, you’ll understand. The feeling is incomparable.”

He’s also honest about his own mistakes: “I once lost tens of millions because I tried to do everything at once. Don’t copy that. Find one pain point, solve it completely, then move to the next.”

The Future: Humans in the Right Roles

The deeper question Huang grapples with isn’t automation—it’s what humans should actually be doing.

“Every job in our company has multiple people, and each person thinks they’re the most important,” he said with a laugh. “But when I rotated through every role myself, testing each person’s methods, I found something interesting: which listing you write doesn’t actually determine if a product sells. The product itself, the keyword selection, whether the market is hot right now—these matter more.”

So what should humans do? “Judgment. Strategy. The things that require taste, instinct, and business sense.” Everything else? “Let the machines handle it.”

FAQ

Q: How did Huang Xufeng grow from 170 employees to 50 while increasing revenue from 100M to 300M RMB?

A: By implementing a three-layer automation stack: 1) RPA for repetitive operational tasks (listing uploads, data entry), 2) AI-powered push dashboards for real-time decision-making, and 3) internal workflow tools that auto-generate product images and listing content. This allowed the company to operate with fewer people managing far more products and platforms, scaling revenue 3x while cutting headcount by 70%.

Q: What is the single biggest mistake small businesses make when adopting AI?

A: They want a magical “one-click” solution from day one, without doing the foundational work first. Huang’s key advice: start with the single most painful process in your business, document it as an SOP, then use RPA to automate just that one thing. Find the feeling of breakthrough before trying to scale. Without clean data, documented processes, and AI-ready infrastructure, even the most powerful agent will fail.

Q: How does the cross-border e-commerce selection process actually work?

A: The company uses a data-driven approach: for each potential product, they calculate cost from 1688 + international logistics + platform fees, then determine if profit margin targets are met. Products are assigned to specific platforms (Amazon, AliExpress, etc.) based on financial核算体系和回款周期. One person can select about 150 products per day. Listings are then auto-published via RPA, with ongoing monitoring via AI dashboard.

Q: What is the AI Selection Officer role in Huang’s company?

A: A dedicated role with a 20,000 RMB/month budget to research, test, and evaluate new AI tools and models. The person produces structured reports on which tools can improve efficiency, at what cost, and with what expected ROI. This prevents the company from making blind investments in trendy but ineffective AI products. Huang believes this is critical because “going out and learning” has become a prerequisite for staying relevant in cross-border e-commerce.

Q: Why did Huang make all 160 employees get UiPath RPA certified?

A: Because RPA implementation fails in most companies not because of technical limitations, but because non-IT employees can’t articulate what they need. By certifying everyone, everyone understood the boundaries and possibilities of the tool. Suddenly 100+ people could submit useful automation requests instead of one IT person trying to extract needs from people who couldn’t explain them. This “全员考证” (everyone gets certified) approach transformed how the company identified automation opportunities.