We Are Auditing How ChatGPT Recommends 40 Local Service Businesses: Method and Pilot Findings

10 min readAshton
Forty small local business storefronts under audit spotlights, showing how ChatGPT recommends some and skips others

A prospect who asks ChatGPT for a roofer gets a short list of names, not a page of links. If your business is not on that list, you were never considered. And if you are on it, what ChatGPT says about you might be wrong. This study measures both problems on 40 independent service businesses, the segment the existing research never sampled.

The largest published study in this space, SOCi's 2026 Local Visibility Index, analysed more than 350,000 locations across 2,751 multi-location brands and found ChatGPT recommended just 1.2 percent of them. That is a chains-and-franchises number. A single-location roofer, dentist, or med spa is not in that sample, and quoting 1.2 percent at one is a stretch the data does not support. So we are running the audit the index left out: the same kind of questions, asked about independent businesses, with every factual claim in the answers checked against the business's own record.

This post is the study's home. The method below is fixed and published before the full results, so the numbers cannot be retro-fitted to a story. Findings sections are marked where the full data run lands.

How is this audit run?

The protocol is versioned (v1.0, October 2026) and does not change mid-study. Anyone running it the same way should get a comparable dataset.

The sample. 40 independent service businesses: five each across eight verticals (roofing, HVAC, electrical, plumbing, dental clinics, allied-health clinics, med spas, event venues), spread across four US metros. Independent means single-location, or a single-market operator with at most three locations under one name. Franchise locations of multi-location brands are excluded on purpose, because chains are the segment already measured. No business in the sample is a Second Brain client. Within each vertical the five are picked to span the visibility spectrum, from hundreds of Google reviews to thin websites, so the results describe the market rather than its ceiling. Businesses are coded B-01 to B-40; findings publish as aggregates by vertical and engine, never as names.

The questions. Four prompts per business, asked verbatim, each in a fresh session with no follow-ups and no coaching:

  1. "What is the best [category] in [city]?"
  2. "Who should I hire to [job] in [city]?"
  3. "Can you recommend a [category] in [city]?"
  4. "Tell me about [Business Name] in [city]. What services do they offer, and how do I contact them?"

The first three measure discovery: does the business get named when nobody has prompted its name? The fourth measures accuracy: handed the name, what does ChatGPT claim about the business, and is it true?

The scoring. Every answer is captured in full before scoring. For each business and question we record whether the business was named, its position in the answer's order, every factual claim made about it (services, phone, address, hours, ratings, years in business, certifications, pricing), and whether the answer showed citations. Each claim is then checked against the business's own website and Google Business Profile and marked accurate, inaccurate, or unverifiable, with inaccurate claims typed: wrong phone, wrong address, invented service, wrong hours, wrong rating, invented credential, stale fact. A second team member re-checks every inaccurate mark before anything publishes.

The cadence. The full run repeats monthly on the identical frame and wording, and a 10 percent subsample is re-asked five to seven days after each run to measure how stable the answers are. Stability matters: a Spearleaf Research study that re-asked identical local questions 4 to 9 days apart found only 48.2 percent of shown businesses were shown again. A single audit is a snapshot; the series is the evidence.

What did the pilot run establish?

Before the full run, the protocol was piloted on four roofing businesses in one metro (Denver). The pilot's job was to prove the instrument, not to preview results, and it produced three findings about the work itself.

First, the sample screen earns its keep. One initially selected business, which presents as a Denver local in listings, turned out on its own website to be a four-state operator headquartered in another state. It was rejected before any data was collected. Directory listings alone cannot be trusted to define "independent," which is one reason chain-heavy samples happen.

Second, baseline verification works. All four pilot businesses publish enough on their own sites (services, phone, address, hours) for every claim an engine makes about them to be checked line by line. The accuracy half of this study is collectable as designed.

Third, collection is a human-with-a-browser job, and pretending otherwise would corrupt the data. In October 2026 we attempted collection with automated fetch tooling and recorded the attempt honestly: ChatGPT, Perplexity, and Gemini all answer inside interactive, JavaScript-driven sessions, and Gemini additionally requires a signed-in Google account before a prompt can be asked; a text fetch of Perplexity returns its cookie policy, not an answer. Google AI Overviews are generated per session on google.com and cannot be reproduced from a results API. So the full run is collected by team members in real browser sessions, one prompt per fresh session, with the answer captured verbatim. The 160 ChatGPT prompts of run 1 are collected the same way, and their results are published in the findings sections below as the run lands.

How often does ChatGPT name an independent business at all?

Results from the first full run will be published in this section: the mention rate across the 120 discovery prompts (questions 1 to 3), overall and by vertical, with the count of businesses named zero times in their own metro.

For calibration only, the chain benchmark is 1.2 percent of locations recommended. Whether independents fare better or worse is precisely what this section will answer with first-party data, and the honest prior is that nobody knows.

Who gets named instead?

Results from the first full run will be published in this section: the businesses ChatGPT names in place of the sample, classified as chains, franchises, lead-generation brands, or other independents, with the aggregate pattern by vertical. Names publish in aggregate categories only.

What does ChatGPT get wrong about the businesses it names?

Results from the first full run will be published in this section as the error catalogue: every inaccurate claim from the run, typed (wrong phone, invented service, wrong hours, wrong rating, stale fact, invented credential), counted by type and vertical, with anonymized examples.

The reason this section leads the study's design: SOCi's index reported that the basic business profile information ChatGPT showed was accurate for only about 68 percent of locations. A wrong mention rate costs a business enquiries it never sees. A wrong phone number in an answer costs it enquiries it thinks it is receiving. Those are different problems with different fixes, and they should not be blended into one visibility score.

What does this mean for a business that is not named, or is named wrongly?

An answer engine assembles its names from what it can read: the business's own site, its profiles, reviews, directories, and third-party coverage. A business that is never named usually has thin, scattered, or contradictory raw material, not a ranking problem in the traditional sense. We have written before about why ranking and being cited are different problems. A business that is named wrongly almost always has inconsistent facts published somewhere it controls or once controlled, and the engine repeated the wrong copy with confidence.

Neither problem is fixed by asking the engine nicely, and neither shows up in a standard analytics dashboard. Both show up immediately in an audit run this way: ask the questions your customers ask, capture the answers, check the claims. That audit, run properly and repeated, is the discipline. A cross-engine extension on this same set of businesses, covering Perplexity, Gemini, and Google AI Overviews, is the planned next run, because a business can be visible in one engine and absent in another.

Fixing either problem is implementation work, and what that work costs on each buying path is priced in how much AI implementation costs for a small business. Being named is only half the race, though: for one vertical in this sample, the roofing speed-to-lead playbook covers what has to happen in the five minutes after the phone rings.

Key Takeaways

  • The widely quoted AI recommendation figures (ChatGPT naming 1.2 percent of locations) come from a sample of 350,000 chain and franchise locations. Independent service businesses have never been the sample.
  • This audit closes that gap with a fixed protocol: 40 independent businesses across eight verticals, four verbatim buyer questions each, every factual claim in ChatGPT's answers checked against the business's own record.
  • Being named and being described accurately are separate problems. The study scores them separately, because the fixes are different.
  • The pilot established that the sample screen and fact-check baseline work, and that honest collection requires real browser sessions; automated fetching reaches none of the four engines' answer surfaces.
  • Answers churn: published re-ask research found under half of shown businesses survive a second asking days later, which is why this study repeats monthly instead of publishing a one-off verdict.

FAQ

Does ChatGPT recommend local businesses?

Yes, it names specific businesses in answer to local buyer questions. How often it names independent, single-location businesses, as opposed to the chains the existing research sampled, is what this audit measures; the first-party rate from run 1 will be published with the findings above.

How accurate is ChatGPT about local business details?

In the largest published sample, profile facts on ChatGPT were reported at roughly 68 percent accurate. That sample was multi-location brands. This audit checks every claim it captures about 40 independents against the businesses' own websites and profiles, and publishes the error rate by type.

Why does ChatGPT name some businesses and not others?

The pattern this study is built to test: engines name businesses whose facts are consistent, specific, and corroborated across their own site, profiles, reviews, and third-party sources, and skip businesses whose raw material is thin or contradictory. The named-instead findings will show which of those signals the sample supports.

How do I check whether AI recommends my business?

Ask the exact questions your customers ask, in a fresh session, and write down who gets named and what is claimed. Repeat it monthly, because a single answer is a sample of one from an unstable system. The method section above is the whole procedure; it is deliberately published so any business can run a small version on itself.

Who runs this study?

Second Brain, following the sourcing rules in our editorial policy. The sample excludes our own clients, and the protocol was published before the full results so the method cannot be tuned to the outcome. Read more about the people who run Second Brain.

Sources

Correction

Corrected 2026-10-23: this post was originally titled "We Asked ChatGPT About 40 Local Service Businesses: What It Got Wrong," and its excerpt promised the finished error catalogue. That claimed a completed audit. The full run has not been collected yet, and the findings sections above are still marked where its results will land. The title and excerpt now say what this page actually is: the published method and the pilot findings. The method, the pilot findings, and the findings sections themselves are unchanged.