We turned a paid SaaS tool into our own working product. The interesting part isn't the tool. It's how we got there — the problem, the decisions, and the plumbing.
We found a paid tool, getaiscorecard.com, that audits a Google Business Profile and scores it the way AI engines like ChatGPT and Google's AI Overviews "see" a local business. We wanted our own version — one we control and can put our name on.
The first thing the AI did wasn't write code. It found the catch: Google's official API only hands back 5 reviews per business — but the entire product depends on the full review history. That one constraint quietly shaped every decision that came after.
A handful of choices up front mattered more than any single line of code.
Prove the engine works before building accounts, logins, and billing. Walk before you run.
Cheap, global, and one place to host the page, the database, and the minute-long scan itself.
We picked the one provider that gives all the reviews, the owner replies, and the local competitors — and tested it with a live call before betting on it.
Some scores are plain arithmetic. The judgment calls — "are these reviews specific?", "how does this owner handle criticism?" — go to Claude. We don't use AI for what a formula does better.
This is the part people skip. It's also the part that matters.
Before any code, we had the AI research the current API docs and pricing — then critique its own plan to catch mistakes. It built on verified facts, not guesses. (It caught two wrong endpoints and a scoring bug before we wrote a line.)
A Google Maps key with the right API switched on, a reviews-data account, a Claude key — each one tested with a live call before it went into the app, and stored as an encrypted secret, never in the code.
Paste a link → identify the business → find 3 real local competitors → pull the full review history → score it → write the recommendations. A system with steps that can each fail and recover on their own.
The AI kept notes between sessions — our decisions and the gotchas we'd already hit — so we never had to re-explain the project or relitigate the same calls.
Building with AI isn't one perfect prompt. It's a loop: try it → watch what reality does → diagnose → fix — with a human steering.
Worked on our machine, failed live. The AI read the error, traced it to a single invisible character accidentally added when the key was uploaded, and fixed it.
A scan crashed for one business. The AI pulled the live logs from Cloudflare, found we were hitting a hard limit by checking the data source too many times in one shot, and re-architected that piece to be patient and durable instead.
The benchmark pulled junk for a marketing firm. The AI tested the maps API live, realized the business category was too generic, and rewrote the matching to find real competitors.
The skill that matters now isn't typing code. It's knowing what to build, what to decide, what to verify — and how to keep the AI honest.
The repeatable part is the method: take a real problem, ground the AI in real information, connect it to real tools, and iterate against reality.
So the question isn't "can AI do this?" — you already know it can. The question is the one worth talking about tonight: what would you point this at?
It's live right now — paste a Google Business Profile and watch it score in about a minute.
Open the live scorecard →