How Amazon Sellers Actually Make Money (Full Guide)

I stumbled into this completely by accident

I needed a third product for my store

Spent three weeks doing the standard thing: Helium 10, viral search queries, watching what other sellers were promoting across social media

Found plenty of options. Every single one already had 50+ competitors Margins were dead before I even requested a sample quote

Eventually I got frustrated enough to just start dumping review data into Claude to see what came out

That experiment became a system. That system surfaced a product with 48% net margin that two of my regular supplier contacts had never considered sourcing

This is how I operate now

Why Trending Product Lists Are a Dead End

The reasoning feels sound: find what's already moving, sell that thing

The catch is that by the time a product appears on a trending list, you're already late

You're walking into a market that's been picked over. The sellers who extracted real profit got in months earlier they built review counts, locked in BSR, established supply chains. You're arriving at the tail end of the curve

I'm not saying you can never break into a competitive listing. But the upside is limited and the grind is relentless

Non-stop PPC babysitting, race-to-the-floor pricing, one defective batch away from losing everything you built

What I actually wanted was different

A product where demand was already proven people were buying *something* in that category but the existing version had obvious, documented flaws

Not some untested niche nobody cares about. A niche full of buyers who are making do with a mediocre product because nothing better exists yet

That's a specific problem. And it's one you can actually solve

Why Review Data Is the Real Product Intelligence

The richest source of product research on Amazon isn't the products themselves

It's the reviews. Specifically the 1, 2, and 3-star reviews on the dominant listings in a category

Those reviews are paying customers telling you completely free exactly what they expected and didn't receive. What cracked. What developed a strange odor. What stopped functioning after six weeks

Most sellers skim reviews to confirm general sentiment before listing

I started reading them as something else entirely: a product spec that nobody had built yet

The problem is raw volume. If the top 10 listings in a category each carry 300+ reviews, you're staring at potentially 3,000 individual data points. You won't read all of it. You'll scan, miss the real patterns, and walk away with vague impressions instead of actionable data

So I started piping it all into Claude

For pulling and organizing review data, I use a lightweight scraper built on browser automation:

apify/store

catalog of ready-to-deploy scrapers including Amazon review exporters, free tier available

https://apify.com/store

The Prompt Stack I Actually Use

Here's the real workflow. Not a cleaned-up template this is what I run

Prompt 1: Surface the gap

I pull 1 to 3-star reviews from the top 10 listings in a category. Paste into a text document, typically 300–600 reviews. Then I run this prompt:

ROLE: Product gap analyst

INPUT: Amazon customer reviews for [category]

YOUR TASK:
1. Identify the top 3 repeating complaints across all reviews
2. Flag which complaints no current product resolves
3. For each gap: is this a sourcing issue or a design issue?
4. Score each gap: (a) mention frequency, (b) difficulty to fix

IGNORE: delivery complaints, pricing complaints, seller behavior
FOCUS: product design, materials, mechanism, longevity

RETURN FORMAT: JSON only. No explanatory text.

The JSON format is intentional. When I asked for written output, Claude would hedge and soften everything. JSON forces it to make concrete commitments

For converting web pages and product listings into clean text Claude can process:

microsoft/markitdown

convert any file format to clean Markdown for Claude context, 40K+ stars

https://github.com/microsoft/markitdown

Sample output portable dog water bottle category:

{
  "gaps": [
    {
      "complaint": "Silicone trough warps after repeated dishwasher cycles",
      "frequency": "38% of 1-2 star reviews",
      "gap_type": "material",
      "fixable": true,
      "solution": "Food-grade PP or BPA-free Tritan interior component",
      "difficulty": "low — standard material substitution"
    },
    {
      "complaint": "Lock button sticks or stops engaging after 4-5 weeks",
      "frequency": "31% of 1-2 star reviews",
      "gap_type": "mechanism",
      "fixable": true,
      "solution": "Redesign as friction-fit twist lock, eliminate spring component",
      "difficulty": "medium — new tooling required"
    },
    {
      "complaint": "Capacity too low for large breeds, needs multiple refills per walk",
      "frequency": "22% of 1-2 star reviews",
      "gap_type": "sizing",
      "fixable": true,
      "solution": "Introduce 750ml variant alongside standard 350ml",
      "difficulty": "low — same mold geometry, scaled"
    }
  ]
}

Three distinct problems. Two of them straightforward to address

One of them the button mechanism appearing in nearly a third of all negative reviews across every major listing in the category

This isn't a bad product. It's a design decision the first manufacturer made that every other seller copied without ever questioning it

Prompt 2: Validate against the supplier spec

Once I found a supplier on Alibaba with a twist-lock design, I ran a second pass:

Here is the complaint data from the review analysis: [paste JSON]

Here is the product spec from a potential supplier:
- Material: BPA-free Tritan
- Mechanism: friction-fit twist lock, no spring component
- Capacity: 350ml / 600ml
- Certifications: LFGB (EU food grade)

TASK:
Does this spec directly resolve the top complaints?
Which complaints does it still leave unaddressed?
What questions should I ask the supplier before committing to samples?

RETURN: list of remaining issues + 5 specific supplier questions.

Claude flagged that the 600ml maximum still didn't address the large-breed capacity complaint

It also noted that LFGB is an EU food safety standard and I should verify whether it satisfies FDA requirements for US market sales

Both things I might have caught later. Neither one I would have caught before my first supplier message

Prompt 3: Supplier outreach

I used to write these myself. My messages got replies. Claude's get better replies, faster more specific framing, cleaner structure, reads like an established operation

Write a professional Alibaba supplier inquiry for this product:

Product: portable dog water bottle, twist-lock BPA-free Tritan, no spring mechanism
My requirements:
- 1,200 MOQ for initial order, potential 3,500+ on reorder
- BPA-free Tritan, need LFGB and FDA documentation
- Capacities: 350ml, 600ml, 750ml
- Custom logo engraving + matte forest green colorway
- Samples required before any bulk commitment

Key questions to include:
- Sample lead time vs. bulk production lead time
- Whether they can reinforce the twist-lock housing wall thickness
- Whether carton specs are compatible with Amazon FBA inbound requirements

Tone: professional and direct. Do not signal this is a small operation.

Sent to five suppliers. Four responded within 72 hours

One already had 750ml tooling completed for a European client. Samples arrived in 22 days

What the Numbers Looked Like

Samples arrived. Ran them through testing. Twist-lock engaged cleanly after 300+ cycles. No material odor after multiple dishwasher runs Approved the bulk order

Unit cost (1,200 MOQ)                $1.30
Custom packaging + logo              $0.25
Sea freight to FBA warehouse         $0.51
Amazon FBA fulfillment fee           $3.40
PPC spend (month 1 estimate)         $1.65
─────────────────────────────────────────
Total landed cost per unit         $7.11

Sale price                      $14.99
Net per unit                    $7.88

Month 1 target: 550 units
Projected net: ~$4,300

48% net margin isn't extraordinary for physical products in isolation

For Amazon FBA with paid traffic, it's genuinely strong. The category benchmark for comparable products runs around 20–26%

The difference is everyone else is still selling the spring-loaded version and competing on price to stay visible

I'm selling the one that doesn't jam after a month

Review velocity has been healthy because the product actually solves what buyers documented in their complaints. That effect compounds over time

Part 5 The Complete Workflow

STEP 1  Select a category with real purchase volume
        Pull 300–600 reviews, 1–3 stars only, top 10 listings
        Claude prompt: extract gaps, receive JSON output

STEP 2  Take the gap to Alibaba
        Search for suppliers already solving that specific failure
        Shortlist 4–6 with specs that match

STEP 3  Claude validation pass
        Input complaint data + supplier spec
        Receive unresolved issues + supplier question list

STEP 4  Claude writes outreach messages
        Send to shortlist simultaneously
        Evaluate replies on speed, documentation, flexibility

STEP 5  Order samples, test against original complaint checklist
        Approve or push for iteration
        Negotiate MOQ, branding, lead time

STEP 6  Claude writes the Amazon listing
        Copy angles built around the problems you solved
        Launch PPC on exact-match terms pulled from complaint language
──────────────────────────────────────────────────────────────
Active work:          3–4 days total
Claude subscription:  $20/month
Starting capital:     ~$2,000–2,500

For running the full Claude workflow as a repeatable automated system:

anthropics/claude-code

official CLI for running Claude as an agentic workflow tool

https://github.com/anthropics/claude-code

For connecting Claude to external tools like scrapers and supplier lookups:

anthropics/model-context-protocol

MCP, the open standard for integrating tools with Claude

https://github.com/anthropics/model-context-protocol

Where the Real Advantage Is

This isn't a secret

Review mining has existed in ecommerce strategy discussions for years. The actual edge is in execution depth and speed

Before, doing this properly meant hours of manual reading, building spreadsheet tags for complaint categories, trying to detect patterns across thousands of disconnected data points

Most sellers ran a stripped-down version and accepted the shallow output

"Buyers mention durability issues" is not actionable. Which component? At what frequency? Is it a material decision, a design flaw, an assembly problem?

Claude processes 500 reviews in under two minutes and returns structured, specific, committed output

That changes which tasks are actually worth doing. The concept isn't new. The version of it you can now execute is simply more thorough than anything that was realistic to attempt before

The competitors still selling the spring-lock design aren't oblivious

They just haven't run the analysis. Their 2-star reviews accumulate and they respond by dropping price to compensate

That gap is still sitting there

It's not really about the water bottle

GIF