Generative engine optimization · AI recommendation thresholds
ChatGPT recommends 1.2% of local businesses. Your star rating is the cutoff that decides which ones.
SOCi published its 2026 Local Visibility Index in March, pulled from 350,000-plus locations across 2,751 multi-location brands and 120-plus visibility metrics. Deep in the AI section is a comparison that finally puts a number on something we'd been circling for months: ChatGPT names a local business in 1.2% of the queries it measured, Perplexity does it in 7.4%, and Gemini in 11%. Google's local 3-pack, the old standard, still shows up in 35.9% of those same searches. Google mentions a business roughly thirty times as often as ChatGPT does for the exact same kind of search.
The gap itself isn't what caught our attention. What we cared about was the mechanism SOCi found behind it, because it matches almost exactly a pattern we'd already spotted in our own testing and couldn't fully explain.
The rating isn't a ranking signal here, it's a cutoff
SOCi's report says ChatGPT's recommended locations average 4.3 stars. Perplexity sits around 4.1. Gemini lands closer to 3.9. Read quickly, that sounds like Yelp or Google Maps: more stars, better placement. But the report draws a harder line underneath it. Locations hovering near 3.4 stars, paired with a review response rate under 5%, don't get ranked lower in AI answers. They get erased from them entirely. The report frames this as a confidence threshold rather than a ranking slope, a business either clears the bar the model needs to speak its name, or the model skips it and picks someone else.
We found a second data point pointing the same way. PushLeads, an SEO firm tracking its own client set, puts the entry point for competitive categories at 30 reviews and a 4.3-star average, climbing to 100-plus reviews in packed markets. Different method, different sample, and it still lands on the same shape: a floor you clear rather than a hill you climb.
What that explains in our own numbers
Back in July we ran 50 real wellness and spa businesses through GPT-4o and Perplexity Sonar, searching the way a real customer would: "best day spa near me," "best med spa near me," across five categories. Every business had a live Google Business Profile, a genuine review history, and a normal local footprint, no empty shells among them. Two-thirds, 66%, came back invisible, named by neither model. Med spas were the worst category: zero of ten cited by either engine.
We described that outcome as binary at the time, a business is either in the answer or it isn't, with nothing in between. We didn't have a clean explanation for why the drop-off was so sharp instead of gradual. SOCi's threshold framing gives us one. If AI recommendation really does run on a confidence cutoff rather than a ranked list, a business sitting just below that line looks identical to one that's nowhere close: both register as zero. A day spa averaging 4.6 stars and a med spa averaging 3.9 don't read as "somewhat different" to the model. They read as two completely different answer types, present or absent.
It also points at a lever most owners in our test set weren't touching. Review count and star average get attention because they sit right on the listing page. Review response rate gets ignored, since it's a habit nobody thinks to check rather than a number anyone sees. If SOCi's 5% threshold holds up beyond their own sample, a business could carry a respectable star average and still land on the wrong side of the cutoff because nobody's replying to reviews.
Where the two datasets don't quite meet
We want to be straight about the seam here. SOCi's numbers come from multi-location brands, chains and franchises with a profile at every address. Our 50 businesses were independent local operators, one location each, no franchise backing. The mechanism could differ between those two groups even if the outcome shape looks the same. We also don't have star-rating data attached to our own 50 businesses, so we can't say our med spa's zero-of-ten came specifically from a 3.4-star average. We can only say the binary pattern we measured fits a cutoff model rather than proving one exists. SOCi is vendor research rather than peer-reviewed science, and a report built to sell visibility software has reason to make its own gap look dramatic. Treat the 1.2%, 4.3, and 3.4 figures as directional, the same way we'd treat any single-source benchmark.
What actually changes for an owner reading this
If AI recommendation runs on a threshold instead of a gradient, the fix isn't "get slightly better reviews." It's finding out which side of the line you're currently on, because a business at 4.1 stars with solid response habits and a business at 4.1 stars that never replies to anyone could sit on opposite sides of the exact same cutoff. A star average alone won't tell you that. Neither will guessing.
Running your own name through ChatGPT and Perplexity answers the only question that matters: are you in the group AI is willing to name, or not. That's what our free check does in under a minute, no review audit required first.