A go-to-market experiment for $0.32: what 901 bad Google reviews told me
In 2024 a personalised email cost us US$0.17 to US$0.29 to produce. In 2026 I read 901 bad reviews across 10 industries for about $0.32. Market research a small business could never afford is now a cheap experiment, and the costly part has moved to judgment.
Before I picked which Singapore businesses to approach, I wanted evidence of where they were losing customers. So I read their worst Google reviews: 901 of them, across 10 industries. The data cost about $0.32.
Two years ago I couldn't have run that experiment for the price of a coffee. I know, because I was pricing a similar product at the time.
What this cost us in 2024
At Mindreader we built a tool called LinkLeads. It read a prospect's public profile and recent news, then wrote a personalised email for a salesperson to send.
My notes from January 2024 have the unit costs. Each email cost between US$0.17 and US$0.29 to produce, and about half of that was the news search. Leaving the news out halved the cost. So 100 emails cost us up to $29 before we charged anything. The case for it was labour: customising an email by hand takes 10 to 20 minutes, so someone sending 100 a week gets back at least eight hours.
LinkLeads wasn't the only attempt. Because Mindreader sat so close to sales and communication, I also built two lead-finding tools around the same time. Leadlight was meant to be a search engine for leads. Its partner tool wrote personalised messages using recent news about each person. My note from January 2025 called them "demos + mvps while we wait for the agents to get good enough". That was an accurate description of where the tools were.
That maths was sound. It was also expensive enough that every experiment needed a business case first. You didn't test an idea on 900 prospects to see what happened.
The experiment
The question was simple. Which local service industries lose the most customers before the job starts, through slow replies, late quotes and people giving up?
The method had three steps:
Find the businesses
A Google Maps search for each of 10 industries, at $0.00175 a search. I kept open businesses with at least 30 reviews.
Collect their worst reviews
Up to 40 reviews per business, lowest-rated first, at $0.0035 per business. That gave 901 one and two-star reviews with enough text to read.
Tag the complaint
A small AI model read 682 of them and picked the main problem: slow reply, late quote, price surprise, scheduling, follow-through or poor work.
What it found
Poor work was the top complaint almost everywhere, and no system of mine fixes that. The useful signal was the share of bad reviews from people who never became customers, because nobody replied or quoted in time.
| Industry | Lost before the job started | Price surprise |
|---|---|---|
| Interior design | 41% | 5% |
| Renovation | 39% | 4% |
| Electrician | 38% | 16% |
| Plumber | 26% | 33% |
| Movers | 21% | 30% |
| Cleaning | 18% | 5% |
| Aircon | 12% | 23% |
One customer waited a month for a reply and was then told the firm had no time. Another asked for a quote in June and still had nothing in August. Those people never got far enough to complain about the work.
The first pass had a weakness: it sampled only the top-ranked firms, which are mostly large. So I ran a second pass on renovation and interior design in 12 housing estates, sorted by firm size. It tagged 462 bad reviews across 71 firms, 49 of them small, for about $0.41. For small firms the gap was lower, 25% to 27%. Later passes put small plumbers and electricians at about the same level, with far more price surprises on top.
Where the cost moved to
The data was nearly free. The judgment wasn't.
The model over-counted. It tagged rude replies and complaints after the job as "gave up before hiring". So I graded every firm by hand before it went into my lead list. On the first day, 38 small firms went onto my lead list and I dropped 17 others, including one where the "customer" was a subcontractor chasing an unpaid bill.
The experiment also shows its limits when you run it again. A later pass on aircon and cleaning firms sampled 80 businesses for about $0.36 and produced one weak lead. Their bad reviews are mostly about poor work, so the method has little to find.
And reviews undercount the problem. Most people who never hear back don't write a review. The real gap is probably larger than these numbers.
What changed between 2024 and 2026
In 2024, the unit cost decided which ideas we could afford to test. In 2026, it barely registers. What I pay for now is an hour of my own attention reading the results and throwing out the wrong ones.
That changes what a small business should do first. You no longer need to guess which customers to chase, or buy a list. You can look at public evidence of where an industry is failing its customers, for less than a dollar, and then decide.