How we measure who ChatGPT recommends

These rankings aren't opinions, votes, or paid placements. They are a record of what ChatGPT actually answered when we asked — the same way a homeowner would — who to see.

The five steps, in plain language

1. 1

   We ask the questions real homeowners ask.

   In each metro we asked ChatGPT for a pool resurfacing company 4 different ways — everyday phrasings like "Who is the best pool resurfacing company near me?", "My pool plaster is flaking — who should I call to replaster it?", and "Who should I hire to resurface my pool?". No trick prompts, no leading questions. 12 metros × 4 questions = 48 checks for this edition.

2. 2

   We record every company it names.

   Each answer gets logged exactly as given. We then match every recommendation against the 553 U.S. companies in our dataset. When ChatGPT names a business with multiple locations, we credit it once — one winner per named business, so no company is double-counted.

3. 3

   We publish the count, not a number one.

   We tested this. Running the identical questions on the identical list a second time, the same day, changed which company came out on top in more than a third of markets — and in most markets the top spot was already shared by two to five of them. What did NOT move was the count: a company named in most checks was named in most checks again. So the page shows how many of the 4 checks named each one. Companies with the same count are shown together, alphabetically, because the measurement does not separate them.

4. 4

   Named once doesn't make the list.

   Companies named in exactly one check reappeared less than half the time when we re-ran the same questions. That is a coin flip, not a recommendation, so we keep the data and don't publish the name. Two mentions came back 86% of the time; four came back every time. Each metro page shows only the companies ChatGPT named more than once. We never publish a bottom list and never name who didn't make it.

5. 5

   We re-run it quarterly.

   The same benchmark, the same way, every quarter. Companies that improve their AI visibility can appear in the next edition — and the run date is printed on every page. This edition: August 2026.

This edition

Measured August 21, 2026 on ChatGPT (gpt-4o, with web search): 12 metros × 4 question phrasings = 48 checks. No other AI system is included in these numbers.

One difference from most boards in this network, stated plainly: every question here was asked twice, not once. Only 34% of the company-and-question pairs named on the first ask were named again on the second. Because asking twice finds more names, a company has to be named in at least 3 of the 4 questions — on both asks — to be listed here, where most of our boards list at 2.

What this is — and what it isn't

Being on this list means ChatGPT recommended that company when we asked, in this edition, in that metro. It is a measurement of AI visibility — not a judgment of the quality of anyone's work, and not a claim that unlisted pool resurfacing companies are worse at what they do. AI answers also change over time, which is exactly why we re-measure quarterly and date every edition.