Do AI recommendations actually send customers? What the data shows
Partly yes, and anyone who answers that question without a “partly” is overclaiming. What’s solid: large, measured numbers of people now use AI to find local businesses. What’s genuinely hard: proving that a specific mention produced a specific customer, because the path from answer to phone call runs through steps nobody’s analytics can see. This post separates the two, because the honest version is still a strong case.
What we can measure and what we can’t
We can measure how many people use AI to find local businesses, and we can measure whether AI mentions a given business. We cannot cleanly measure how many customers a mention produced, because the common path runs mention, then branded search or a direct call, and the mention leaves no fingerprint on either. Treat every confident attribution claim in this industry accordingly.
This is the correlation post of the basics series in miniature: honest measurement in this space is its own discipline, and the vendors who skip the honesty are the ones to walk away from.
The usage numbers
Three numbers, each with a name and date attached. 45% of consumers say they have used AI like ChatGPT to find a local business, up from 6% a year earlier1. 900 million people use ChatGPT every week2. And behavior inside classic search is shifting the same direction: in Pew’s March 2025 browsing data, visits where an AI summary appeared ended in a click on a regular result 8% of the time, versus 15% without one3.
What these numbers support: the asking is happening at scale, and the answers absorb attention. What they don’t support: any specific count of customers delivered to any specific business. We’re not going to pretend otherwise, and the section on referral traffic below is where most write-ups quietly start pretending.
What referral traffic from AI looks like in analytics
Thin, honestly. Some businesses see visits tagged from AI domains in their analytics, and industry write-ups regularly claim those visitors convert well. We’re not citing those claims because we can’t verify their methods, and this post only carries numbers with names on them. What we can say structurally: click-through referrals are the minority path. An AI answer is a recommendation read aloud, not a results page, and a mention doesn’t need a click to do its work. If your measurement of AI’s value is your referral-traffic report, you’re reading the smallest slice and calling it the pie.
The invisible path: mention, then branded search, then a call
Framed plainly as a hypothesis, with the signals that support it: the likeliest common path is that a customer asks, hears two or three names, and then searches the name they liked or calls directly. Supporting signals: that’s how radio, word of mouth, and every other spoken recommendation has always converted; branded-search volume is a thing owners can watch move; and the click data above shows attention staying inside answers rather than flowing out through links. What would confirm it properly is longitudinal data joining mentions to branded-search movement, which nobody, including us, has published yet. When our tracking can say something real about this, it will say it with dates on it.
What tracking shows instead of guessing
The measurable layer is the mention itself: whether you’re named, on which questions, against which competitors, over time. That’s a real number with an honest definition, it moves when the underlying sources move, and it doesn’t require anyone to believe an attribution story. Our methodology spells out exactly what gets measured and how, including what we refuse to conclude from it. Our opinion, defended across this whole post: in a market where the last mile can’t be proven, the honest buy is measurement of the mile that can be, and vendors selling proven-customers-delivered are selling the part nobody can see.