We Checked 36 Businesses on Five AI Platforms. The Platforms Barely Agree.

Ask ChatGPT to recommend a dental practice and it names one set of businesses. Ask Gemini the same question and it names a different set. Ask Copilot, Perplexity, and Google AI Mode and you get three more answers that overlap less than you would expect.

Most businesses have checked at most one of these. Many have checked none.

Over the summer I measured how five AI platforms respond to buyer-intent questions relevant to 36 real businesses, most of them owner-operated professional service firms in a metro area in the southern United States, and recorded whether each platform actually named the business in its answers. This article reports what that data shows, what it does not show, and what happens next, because this is the start of an ongoing study, not the end of one.

The businesses in the study are identified by code and vertical only. They did not consent to being named in published research, so they are not named. The three national brands used as a benchmark are public companies and are named.

How the measurement works

Each scan poses a set of buyer-intent queries across five AI platforms: ChatGPT, Gemini, Perplexity, Copilot, and Google AI Mode. For each platform, the scan records whether the business was named in the responses, and in what share of topics it appeared.

Coverage percentages are a coarse instrument. On a default scan they move in 20-point steps, so this study leads with the binary question, present or absent, and treats coverage numbers as directional. One business in the sample runs on a larger custom query set, which is why its numbers are finer-grained. Presence and absence remain comparable across all 36.

The anchor finding: the platforms disagree about who exists

Nineteen of the 36 businesses were visible on at least one platform. Here is how those 19 spread across the five platforms.

Of the 19 businesses visible anywhere, 10 surface on two platforms or fewer. Six surface on exactly one. Only two surface on all five.

The single-platform cases are the clearest illustration. One coaching business surfaces only on ChatGPT. Another surfaces only on Copilot. A third surfaces only on Google AI Mode. One business in the sample reaches full topic coverage on one platform while returning zero on another, which is about as far apart as two measurements of the same business can be.

The practical consequence is simple. If you ask one assistant about your business and conclude you are visible, or conclude you are not, you have not measured anything. You have sampled one platform out of five that disagree with each other.

Nearly half are absent everywhere

Seventeen of the 36 businesses, 47 percent, returned zero across all five platforms. Not low coverage. Zero. No platform named them in response to any topic in the scan.

These are established businesses with websites, reviews, and customers. Their owners would mostly not predict this result, and in my experience running individual snapshots, they don't. I wrote about that gap between belief and measurement in August. This study puts a population-level number on it.

Similar rates, poor overlap

The disagreement is not explained by some platforms simply being stricter than others.

The spread from top to bottom is 16 points. If the five platforms were reading the same underlying evidence and applying different thresholds, the same businesses would tend to surface everywhere, with the stricter platforms surfacing fewer of them. That is not the pattern. Similar rates with poor overlap points at different platforms drawing on different sources.

That has a direct implication for anyone doing visibility work. There is no single lever that makes a business visible to AI, because there is no single AI reading from one place.

A sentiment score for a business AI cannot name is not a compliment

Sentiment scoring came up in the August article, and the study data makes the point sharper. Across the 17 businesses no platform could name, mean sentiment was 72.0. Across the 19 visible somewhere, 76.9. Five points apart, in a population whose scores run from 37 to 84.

A business that no AI platform can name still receives a comfortable sentiment score, because the score measures the tone of the answer to the question, not the platform's view of the business. The business is not in the answer. Read as a performance metric, sentiment on an absent business is false comfort, and it is exactly the number most likely to be quoted back to an owner as reassurance.

Presence first. Sentiment only means something once you are in the answer.

The national control: this is not a technology problem

Alongside the 36 local businesses, I ran the same scan configuration on three national brands in the real estate investment space: Kiavi, BiggerPockets, and Visio Lending.

All three surface on all five platforms. Two of them reach full topic coverage on Google AI Mode, the platform where local businesses surfaced at the lowest rates alongside Perplexity. The weakest of the three still outperforms the strongest local business in the sample on overall visibility.

Three domains is a control, not a sample, and it should be read that way. But it rules something out. The platforms are not failing to answer these kinds of questions. They answer confidently, with entities they can corroborate across many independent sources. National brands have years of press, citations, reviews, directories, and third-party references. Most local businesses have a website, a Google profile, and not much else machines can cross-check.

Absence from AI answers is not a technology problem. It is a corroboration problem. That distinction matters because corroboration is buildable.

What this study cannot tell you

Two limits are built into this first wave, and stating them is part of the point.

First, the 36 scans were not run on one day. They were run across seven weeks, from late June to early August, because each scan is frozen at the date it ran. Two businesses compared in this article may have been measured weeks apart, and some of what reads as platform divergence could be platform drift over time. Wave 0 cannot separate the two. That is the single strongest argument for running this longitudinally instead of publishing one snapshot and moving on.

Second, the sample is not random. These are businesses I selected to scan, weighted toward professional services and owner-operated practices in one metro. The study describes that population. It does not claim to describe all small businesses, and vertical-level patterns inside a 36-business sample are too small to publish as findings.

One more platform note for completeness: Grok appears in the tooling but returned no usable data for this wave and was excluded.

What happens next

This is Wave 0, the reconnaissance pass. Wave 1 runs on September 10 on a fixed schedule, and monthly waves follow. Because of a tooling constraint, the ongoing waves use a different measurement route that covers ChatGPT, Gemini, Perplexity, and Grok, so the longitudinal series starts its own clean baseline rather than pretending continuity with this wave. Each wave records its run dates, so no future wave can quietly repeat this one's seven-week spread.

In mid-September I will publish what moved between waves. If nothing moved, I will publish that. A study that reports stability is still reporting something: it would mean AI answers about these businesses are stable enough that the divergence pattern is structural rather than noise.

What to do with this if you run a business

One thing, before anything else: do not conclude anything from checking one assistant. Ask the same buyer-intent question, the one your customers would ask, on at least ChatGPT, Gemini, and Perplexity. Note who gets named. If you are absent from all of them, you have learned something your analytics dashboard will never show you.

And treat any reassuring score about your AI presence with suspicion until you know whether the platforms can name you at all. Presence is the measurement. Everything else is commentary on it.

If you would rather have this done systematically, the first snapshot is free. startsolutions.ai.

Start Solutions AI measures how AI platforms discover, understand, and recommend businesses, and publishes what it measures. The Platform Divergence Study is ongoing; Wave 1 results publish in mid-September.