Counting, Not Scrolling: A Formula-First Workflow for Amazon Review Mining

The graveyard

Every Amazon seller I know has the same two files sitting somewhere: an export of a few hundred reviews nobody has time to read, and a Search Term Report that got downloaded, glanced at, and abandoned.

There are SaaS tools that will happily charge $99+/month to fix this. What most of them do, underneath the dashboard, is show you your own data back with charts. The data was never the problem. The counting is the problem — and counting is free.

So here's the workflow I run instead: a spreadsheet that counts, and one AI prompt that explains. In that order.

Counting first, AI second

Review mining is a counting problem before it's an AI problem.

Paste 200 reviews into a chatbot and ask "what do buyers think?", and you get a summary that sounds plausible and ranks nothing. It will spend as many words on the one-off complaint as on the problem mentioned forty times, because it has no way to know which is which.

Formulas don't have that failure mode. A COUNTIF doesn't care which review was written most passionately. So the order of operations is: formulas rank, then AI explains the top of the list. Everything below follows from that.

The worksheet: one log, one formula layer

The log is boring on purpose — one row per review:

• Date
• Star rating
• Review text
• Pain-point tag — one short label you assign: lid hard to open, sizing, packaging, instructions
• Sentiment — positive / mixed / negative, which I assign while I'm tagging anyway; it takes two seconds and catches the sarcastic 5-star review a star count never will

On top of that sits a summary block of plain formulas:

• COUNTIF over the tag column → mentions per tag
• AVERAGEIF → average stars per tag, which separates "genuinely unhappy buyers" from "happy buyers with one gripe"
• COUNTIFS → the 1–2★ share per tag
• COUNTIF over the sentiment column → set totals

That's the whole machine. The one real design decision is the tag taxonomy. Tag the failing thing or the broken expectation, not the emotion. Too broad — quality, bad — and everything collapses into mush. Too narrow — lid hard to open with wet hands — and every tag has exactly one mention and nothing ranks. A good tag is the size of a fix: something you could act on in the product, the listing, or the instructions.

The prompt layer: where AI actually earns its place

Once the sheet says which two or three tags matter, I paste the review text for those tags into this prompt. It's the exact one I use — brackets are the parts you replace:

You are helping me analyze customer reviews for my own Amazon product. I will paste real review text below. Work ONLY from the text I provide — do not invent product features, specifications, or buyer complaints that are not in the reviews.

Product: [product name and type]
Reviews (all for the pain-point tag "[tag]"):
[paste review text]

Return:
1. The root cause behind these complaints, in one or two sentences — what specifically is failing or being misunderstood?
2. Split the cause: (a) product issue, (b) listing/expectation issue (the listing promised or implied something the product doesn't do), or (c) usage issue (buyers using it in a way the instructions don't cover). Quote the review phrases that support your split.
3. For (b) and (c) only: draft the exact listing or instruction change that would prevent this complaint — a rewritten bullet, a clarified size/spec line, or an added usage note. Use only facts from the reviews and my product details above; add no new claims.
4. One thing I should NOT change, based on what happy reviewers in this set said they liked.

The split in step 2 is the entire point. Only one of the three causes needs a factory, a re-order, or a tooling change. Expectation issues get fixed tonight, in the listing, for free. And the "work ONLY from the text I provide" line is not decoration — without it, the model will invent specifications your product doesn't have and complaints nobody wrote.

One full run, on fictional data

To show the loop end to end, I'll use the demo set that ships with my toolkit: 18 fictional reviews for a fictional 32 oz insulated bottle. Every number below comes from that set.

The summary block ranks the tags like this:

Plus three single-mention tags. Set totals: 3.4★ average, 7 positive / 3 mixed / 8 negative.

Reading those reviews one by one, I'd have remembered the 1★ "leaks in my bag" — it's vivid. The count says it's a single mention, while four separate tags tie at two. And the average-stars column separates those four instantly: two are real 2.0★ problems, one is a 4.0★ annoyance, and one — cup holder, at 3.0★ — smells like an expectation issue, because both reviewers praise the bottle while complaining: "Quality is good, but…" and "great bottle, wrong diameter."

Running the prompt on that tag's two reviews, the split comes back (b) listing/expectation, quoting "which was the whole point for my commute" and "wrong diameter for a standard car cup holder." The drafted fix is a single bullet:

Size check before you buy: this is the wide-body 32 oz. It does not fit standard car cup holders. It's built for desks and gym bags — if your commute is cup-holder-or-bust, this is the wrong bottle.

And step 4 earns its keep: the prompt warns not to shrink the bottle chasing this complaint, because the 5★ reviews are praising exactly the capacity and the insulation — "still cold at the end of my 10-hour shift," ice still inside after 12 hours in a hot car. The fix is a sentence in the listing, not a new product.

Small numbers, fictional bottle — but that's the loop at any scale. The counts tell you where to look. The prompt tells you what kind of problem it is. Most of the time, the fix is copy, not inventory.

The other two stations

The same count-first logic runs the rest of my monthly pass:

Competitor teardown. A rival's listing is public data you're allowed to analyze. I use an 8-section fill-in template — title structure, bullet angles, pricing log, their review weaknesses, image map, keyword coverage, differentiation verdict, change log — because a template gets finished and a blank page doesn't. Their 1-star clusters, counted the same way, are your differentiation angles in the buyers' own words.

PPC search-term triage. The Search Term Report is the review graveyard's twin: downloaded, never counted. Count spend per term, count conversions, and the report splits itself into waste to negate and winners to harvest into exact match. Same principle as the reviews — the data is already in your account. It just hasn't been counted.

Guardrails, stated plainly

Three lines I don't cross, and the prompts are built to hold them:

• Everything runs on data you paste in. The prompts instruct the model to work only from that text — no invented specs, no invented complaints.
• Nothing in this workflow writes, edits, or solicits reviews. It only reads reviews that already exist. That line is deliberate — it's the one part of "AI for Amazon" that can end a business.
• No outcome is promised. This produces analysis and drafts. Whether a bullet rewrite helps is your call to make and your data to judge.

The file

I packaged my version of this loop as a one-time toolkit: the review-mining worksheet (.xlsx + .csv, formulas prebuilt — the 28 of them that produce the summary block above), the 8-section competitor teardown template, and 50 prompts in five groups (listing, review mining, competitor teardown, PPC, buyer-message replies), each with a worked fictional example. A free 5-prompt mini-pack is included, so you can test the format before paying anything.

Personal is $39, Commercial (for VAs and agencies, up to 5 seats) is $79. One-time, no subscription: https://promptandskillco.gumroad.com/l/spjfod