Why 1-star reviews and not customer interviews?

Do both — but 1-star reviews come first because they are free of the biggest bias in interviews. An interview subject is being polite to you and speculating about a future purchase. A 1-star reviewer already paid, already used the thing, and is angry enough to write about it unprompted. That is revealed preference plus a specific unmet need, in the customer's own words.

What does a full validation pass actually cost?

A complete run — trend history on your keywords, ~50 Reddit threads, five funded competitors, 200 Trustpilot reviews across them, 50 local businesses in your target segment, and verified contact addresses — costs about $1.42 at current per-result rates. You pay per result delivered, and nothing for a search that returns nothing.

What if the competitor has no Trustpilot page?

That is itself a signal, but do not over-read it. Check the parent company and the consumer-facing brand name separately — B2B tools are often listed under a different entity than the product name. If a category genuinely has no review presence anywhere, lean harder on the Reddit step, where complaints show up before review sites exist.

Isn't a pile of complaints just proof the market is bad?

The opposite, usually. Complaints require customers, and customers require someone already paying. A category with angry paying users and a funded incumbent is a market with money in it and a gap in it. A category with no complaints normally has no customers.

I Validated a Business Idea in 48 Hours by Mining 1-Star Reviews

Asking friends if they like your idea is not validation. The real signal is what people already pay for and still hate. Here is the exact flow — demand, unprompted complaints, funded incumbents, and the 1-star review clusters that become your product spec — run entirely from Claude via MCP.

Every founder has had this conversation. You describe the idea. Your friend says “oh that’s actually really cool.” You leave feeling validated. Nine months later you have a product nobody buys.

Nothing about that conversation was validation. Your friend was being kind, speculating about a hypothetical purchase, with none of their own money involved. You collected a data point with a bias baked into it so deep it is not worth having.

Here is the signal that is actually worth something: what people already pay for and still hate.

That signal exists, in enormous volume, in public. It is sitting in the 1-star reviews of the products already serving the market you want to enter. Those reviews are written by people who already opened their wallet, already lived with the thing, and are annoyed enough to write about it unprompted. They are not speculating. They are reporting.

A list of what paying customers hate about the incumbent is not market research. It is a product spec, written by your future customers, for free.

That is a data problem. So let’s treat it like one.

This post is the complete flow I would run before writing a line of product code. Seven tools, one Claude conversation, about a dollar and a half. No code.

The setup (two minutes, once)

The Mine Works MCP server puts our scrapers inside Claude as native tools. Claude calls them itself and reasons over the results in the same conversation — you never copy-paste a CSV.

Open ~/.claude.json (or Claude Desktop → Settings → Developer → Edit Config) and add:

{
  "mcpServers": {
    "themineworks": {
      "url": "https://the-mine-works-mcp.hatchable.site/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_APIFY_TOKEN"
      }
    }
  }
}

Grab your token from Apify → Account → Integrations. Restart Claude. Runs bill to your own Apify account at pay-per-result rates — there is no subscription on our end.

Now the flow.

Step 1: Is there demand — and is it growing or dying?

The manual version of this step is you, on a Tuesday night, staring at a Google Trends tab, eyeballing a squiggly line, and deciding it looks encouraging. It always looks encouraging. You are the least objective reader of that chart alive.

The problem is not the data — Trends is good data. The problem is that a single 12-month view of a term you already love tells you almost nothing. What you need is direction over years, not a snapshot, and a comparison against something you are not emotionally invested in.

Google Trends Pro pulls that history as structured data Claude can actually reason over, rather than a picture you squint at.

Prompt: Use search_google_trends with keywords for my category and its two closest substitutes, geo “US”, timeRange “today 5-y”. Then run it again with timeRange “today 12-m”. Compare the two: is the 12-month direction consistent with the 5-year trend, or is the recent movement noise? Tell me plainly whether this category is growing, flat, or dying, and say so even if the answer is dying.

Two views matter here. The five-year view tells you whether you are entering a category or exiting one. The twelve-month view tells you whether anything changed recently. A term that is up over twelve months but down over five years is a dead category having a moment — that is the trap, and it is the single most common way founders misread this chart.

Run at most five keywords per call — that is the tool’s ceiling, and it is Google’s, not ours.

Cost: $0.015 per trend report. Two reports is three cents.

Step 2: What are people complaining about when nobody is asking?

Survey answers are performances. Someone knows they are being studied, so they give you the reasonable-sounding version of their opinion — the one that makes them look like a thoughtful consumer.

Reddit is where the same person says what they actually think, at 1am, to strangers, because something broke and they are annoyed. Nobody is watching. That is exactly why it is useful.

The manual version is you reading threads for six hours and remembering the three that confirmed your idea. That is not research, that is confirmation bias with extra steps.

Prompt: Use scrape_reddit with mode “search” and a query describing the problem I think I am solving. Pull maxResults 50. Then run it again with mode “subreddit” on the two subreddits where my target customer actually hangs out. Read every thread. Do not summarise them one by one — cluster them. What are the recurring complaints, ranked by how often they appear? Separately, list any complaint that appears only once but is unusually specific, and say why it caught your attention.

The clustering instruction is the whole step. Fifty individual thread summaries are useless — you cannot hold them in your head. Five ranked complaint themes across fifty threads is a finding.

Note the second half of that prompt too. Frequency tells you what the market feels. The specific one-off is often the person who has thought about this hardest.

Cost: $0.003 per post. Fifty posts is fifteen cents.

Step 3: Who already serves this market — and did someone fund them?

Founders treat competitors as a threat to be explained away. “Nobody’s doing this” is almost always a research failure, and when it is true it is usually bad news: it means someone tried, it did not work, and they left.

What you want to know is not whether competitors exist. It is whether someone with a diligence process has already looked at this market and put money in it.

Prompt: Use search_crunchbase to find companies in my category. I want total funding, last round, and the date of that round for each. Then tell me what the funding pattern implies about the market: is money flowing in, or did it stop two years ago?

Funding date matters far more than funding total. A category where the last meaningful round closed three years ago is a category the market already judged. A category with recent rounds means someone with more diligence budget than you has concluded there is money here.

Note: Crunchbase is a browser-based scrape and takes longer than a single request can wait. The tool hands back a run_id and Claude collects it with get_run_results a few seconds later. You will see it do this automatically — that is expected, not an error. search_google_maps in step 5 behaves the same way.

Pick five competitors out of this. Those five are the input to the step that matters.

Cost: $0.0133 per company. Five companies is seven cents.

Step 4: The core step — what do their paying customers hate?

This is the one. Everything above it was setup.

The manual version of this step barely exists, which is why almost nobody does it. You would have to open a competitor’s review page, sort to the bad ones, read a few hundred, and somehow hold the pattern in your head across five companies. Nobody does that. So the richest source of product truth on the internet just sits there.

Trustpilot Reviews pulls the reviews with their star ratings attached. Claude does the part you were never going to do.

Prompt: Use scrape_trustpilot on each of these five competitor domains, maxResults 200 across them. Then filter to 1-star and 2-star reviews only and cluster them into recurring complaint themes. For each theme: how many reviews mention it, which competitors it hits, whether it is a product problem or a service problem, and quote two reviews in the customer’s own words. Rank the themes by frequency. At the end, tell me which single theme you would build against and why.

Read that output carefully, because the important distinction is in it.

Service complaints — slow support, bad billing, hard to cancel — mean the incumbent is badly run. That is an opening, but a shallow one. They can fix it by hiring, and they eventually will.

Product complaints — the thing structurally cannot do what customers need — mean the incumbent is architecturally stuck. That is a moat you can walk into. If four out of five competitors have the same product complaint, you have found a constraint the whole category shares, and that is a business.

The instruction to quote the reviews verbatim is not decoration. Those quotes are your landing page copy, your cold email opening line, and your first user interview script. Your customers already wrote your positioning for you, in their words, while angry. That is better copy than you will write on your own.

A caution: a review corpus skews toward extremes. People write reviews when they are furious or delighted, rarely when they are fine. That skew does not invalidate the complaints — a real problem is real regardless of who reports it — but it does mean you cannot read the ratio and conclude anything about how the median customer feels. Use this for what is broken, never for how many people think so.

Cost: $0.005 per review. Two hundred reviews across five competitors is a dollar.

Step 5: Who exactly would I sell to?

You now have a hypothesis with evidence under it. The next question kills more startups than any of the above: can you name fifty specific people who would buy it?

Not a persona. Not a TAM slide. Fifty names.

If the answer is a local or SMB market, Google Maps Search turns your abstract segment into a real list.

Prompt: Use search_google_maps for the business type that has the problem I identified, location set to my target market, maxResults 50. Give me name, website and category for each.

The value here is not the list. It is the forcing function. A segment you cannot enumerate is a segment you invented. If you cannot fill fifty rows, the problem is not distribution — it is that your market is a category, not a customer.

Cost: $0.0015 per place. Fifty places is eight cents.

Step 6: Go talk to them

Validation ends with a conversation, not a document. Everything so far exists to make that conversation worth having.

Prompt: Use find_website_contacts on the domains from the Maps list. Then run verify_emails on everything it finds — I only want addresses that pass MX and SMTP.

Website Contact Finder crawls the homepage, contact, about and team pages and pulls emails, phones and social profiles. Email Verifier then tells you which are real, which are role-based (info@, contact@ — a queue, not a person), and which are catch-alls.

Sending to a dead address is worse than not sending. Verify first.

Prompt: Write a short outreach note. Do not pitch anything — I have nothing to sell yet. Open with the specific complaint theme we found in the 1-star reviews, phrased the way the reviewers actually phrased it, and ask whether that matches their experience. Under 100 words. No “quick question,” no “hope this finds you well.” Just the problem and a real question.

This note lands differently than a founder’s cold email normally does, and the reason is not the writing. It is that you are opening with a problem the recipient genuinely has, described in language taken from people exactly like them. You are not asking them to imagine your idea. You are asking whether a thing you already know is broken is broken for them too.

That is a question people answer.

Cost: $0.001 per domain, $0.001 per email verified.

Step 7: Watch it instead of remembering to check

The whole thing is a prompt, which means it can be a recurring one.

Prompt: Re-run the Trustpilot pull on the five competitors and compare against last month. Any new complaint themes? Any theme that has gone quiet — meaning someone shipped a fix?

A theme going quiet is the most valuable output this flow produces. It means an incumbent just closed the gap you were about to build into. Better to learn that from a review diff than from your own launch.

What it actually costs

Real per-result prices, for a full validation pass:

Step	Tool	Rate	Volume	Cost
Demand	`search_google_trends`	$0.015/report	2 reports	$0.03
Unprompted complaints	`scrape_reddit`	$0.003/post	50 posts	$0.15
Funded incumbents	`search_crunchbase`	$0.0133/company	5	$0.0665
The 1-star mine	`scrape_trustpilot`	$0.005/review	200 reviews	$1.00
Name the buyers	`search_google_maps`	$0.0015/place	50 places	$0.075
Contacts	`find_website_contacts`	$0.001/domain	50 domains	$0.05
Verify	`verify_emails`	$0.001/email	50 emails	$0.05
			Total	$1.4215

A dollar forty-two. You pay per result delivered — a search that returns nothing costs nothing, and a blocked or failed run is never charged.

Worth noticing which line is the expensive one. Seventy percent of the spend is step 4, and that is correct. Everything else is context. The reviews are the evidence.

The honest part

This does not tell you whether to build the thing. It cannot, and anything that claims to is selling you something.

What it tells you is whether the problem is real, whether people are already paying to have it solved badly, and what specifically they hate about the current answer. Those are facts. Whether you are the right person to fix it, whether you can reach those customers economically, whether the fix is technically possible at a price they will pay — none of that is in a review corpus. That is judgement, and it stays yours.

Three more things this will not do.

It will not find a market nobody has entered. By construction, this flow reads incumbents. If your idea is genuinely novel there is no review corpus to mine, and you are back to talking to humans — go do that.

It will not protect you from confirmation bias unless you let it. If you prompt Claude for reasons your idea is good, you will get them. The prompts above are written to invite a “no” on purpose. Keep them that way. The value of this whole exercise is concentrated entirely in its ability to kill your idea cheaply, and you can defeat that by asking nicer questions.

And it will not replace the conversation in step 6. Reviews tell you what broke. Only a person tells you what they would pay to fix it.

What it removes is the part that was never the real work: not knowing whether anyone cares, and having nothing specific to open with when you finally ask. Those are lookup problems. They have always been lookup problems. The evidence has been public the whole time — it was just spread across five review pages and fifty Reddit threads that no founder was ever going to read.

Now it takes an afternoon and costs less than a sandwich. And a killed idea for $1.42 is the best trade available to anyone starting a company.

Tools used: Google Trends Pro · Reddit Scraper · Crunchbase · Trustpilot Reviews · Google Maps Search · Website Contact Finder · Email Verifier

All of them are available as one MCP endpoint.