How many Google reviews does a business need before the site writes itself?
This update was drafted on a schedule by the AI I build with, from real project notes — part of the vibecoding experiment this blog documents.
I left a loose end in the post about building a website out of a business's Google reviews. I said the method falls apart below about twenty reviews and moved on. Someone was going to ask what the actual number is, so let me answer it properly, because the honest answer is more interesting than a number.
There isn't one. Or rather, the count is the wrong thing to be counting.
Here's the rough shape of it as I actually experience it on the web side of Cadence Studios. Under about twenty reviews, you're reading a handful of individual opinions and calling it a pattern. Somewhere in the twenties to fifties, a voice emerges — you can tell how customers talk about the place, and the one or two loudest themes hold up. Past fifty, the recurring nouns get stable enough that I'd put them on a homepage. Past a couple hundred, you start getting actual segmentation: different customer types, different occasions, objections that separate out cleanly instead of blurring together.
But I want to be careful, because those are my working thresholds from doing this, not findings from a study, and I'm not going to dress them up as more than that. More importantly, I've seen a business with four hundred reviews where the method worked badly and a business with thirty where it worked beautifully. So the count clearly isn't what's doing the work.
Four things are, and these are what I actually check.
One: how many of those reviews describe the business that exists today. This is the big one and almost nobody applies it. A restaurant with four hundred reviews that changed hands in 2024 is being described mostly by ghosts. The staff people name are gone. The dish everyone raved about is off the menu. If you weight for recency — and I'd say the last eighteen to twenty-four months is the honest window for most local businesses — that four-hundred-review business might have sixty usable reviews, and now it's a mid-sized sample, not a large one. Count the recent ones. That's your real denominator.
Two: does anything actually repeat? The whole method rests on repetition. A specific noun, a person's name, a detail mentioned unprompted across independent reviews — that's the signal, and one mention is an anecdote no matter how good it sounds. My working bar is that a theme needs to show up in at least three independent reviews before it earns a place on the page. If you run through the whole pile and nothing clears that, volume won't save you. You have noise, and a bigger pile of noise is still noise.
Three: are there any reviews that aren't five stars? This is the test people find counterintuitive, and it's the one I'd defend hardest. Five-star reviews tell you the business is good. They don't tell you what to answer. The three-star review is where someone says "loved it, but booking was a mess," and that sentence is worth more to a homepage than fifty variations of "great service." A business with thirty reviews all at five stars has given you tone and nothing else — no objections, so no idea what doubt the page needs to resolve above the fold. I'd genuinely rather have twenty-five reviews with a handful of threes and fours in them than a hundred perfect ones.
Four: how much did people actually write? Twenty reviews averaging four sentences each will produce a better site than a hundred reviews that say "Great!" and "10/10." Star ratings with no prose contain almost no information about a business — they're a mood, not a description. The substantive reviews are the corpus. Everything else is a number.
Put those together and the practical gate I use is something like: roughly twenty or more recent, substantive reviews, with at least one theme repeating three times, and at least a few reviews that aren't five stars. Not a formula, more a shape. If a business clears it, the reviews can carry the positioning. If it doesn't, they can't, and pretending otherwise is where this goes wrong.
Which brings me to the part I actually think matters, and it's not about reviews at all.
The model will never tell you there wasn't enough. Hand it six reviews and ask for a website and you will get a website. A fluent, confident, well-structured website full of positioning that reads like it came from somewhere. It came from six people and a lot of interpolation. Nothing in the output is marked "I made this part up" — that's not how any of this works. The page looks exactly as authoritative at six reviews as at six hundred.
So the refusal has to live in the process, before the generation, not in the output afterward. That's why the review count gets checked as a gate rather than as a note. Below the line, the answer isn't "generate it anyway and be careful" — it's a different method: build the structure, then get the copy from actually talking to the owner, and let the few reviews you have supply tone and nothing else. No inferred differentiator. No claim that rests on three people. Same instinct as the hard approval gate on the outreach side — the machine can do the reading, but somebody has to decide whether there was enough there to read.
There's a version of this problem that has nothing to do with websites, incidentally. Any time you point a model at a small pile of real data and ask it to find the pattern, it will find one. Confidence in the output is not correlated with sufficiency of the input, and that decoupling is, I think, the single most expensive thing to learn about working this way. I ran into the identical wall on the trading side, where an edge built on forty trades looks precisely as convincing on the screen as an edge built on four thousand.
Status, as always: this runs, it's live, no client outcomes to show you and no conversion numbers to quote. I'm not inventing one.
The thing I'd leave you with is that "how many reviews do I need" turned out to be a proxy question. The real one is: is there enough here that I'd be repeating what customers said, rather than composing something plausible on their behalf? You can usually feel the answer after ten minutes of reading. The count is just a fast way of guessing before you start.
This one's auto-drafted from my notes on a schedule. If a number isn't in the notes, it doesn't show up here — I'd rather leave a blank than make something up.