Research · November 18, 2026

We asked ChatGPT for local leads.
Here's what it returned.

Five real prospecting requests, with web search switched on. The answers were fast and readable. They weren't lead lists.

Can ChatGPT build a lead list? Not in our test. For each request ChatGPT ran one web search, returned between 5 and 20 businesses, and then offered to "build a bigger list". It found an email for none of 20 Austin dentists, and 6 of the 20 phone numbers in its own table disagreed with the map listings it displayed underneath.

How we tested

On November 17, 2026 we sent five prompts to ChatGPT with web search forced on, through a standard answer-scraping API, and saved the full responses: the text, the tables, the business listings it showed and every source it cited. These are the kinds of requests a sales rep or agency owner would actually type:

  • Give me a list of 20 dentists in Austin TX with phone numbers, emails and websites
  • List roofing companies in Dallas with the owner's name and email address
  • HVAC companies in Phoenix with fewer than 20 Google reviews
  • Find me new restaurants that opened in Miami in the last 30 days, with owner names and contact info
  • Plumbers in Denver that don't have a website

Four prompts returned an answer. The plumbers request timed out, so the results below cover the other four.

What came back

RequestRowsEmailsMain sources cited
Dentists, Austin200 of 20Two dental directory sites
Roofers + owners, Dallas108 of 10, copied from directories and contact pagesA business directory, BBB, company sites, procurement PDFs
HVAC, few reviews, Phoenix5None requestedLead-list and contractor-license sites
New restaurants, MiamiAbout 8Contact routes, not emailsOpening trackers, reservation blog, local news

Finding 1: one search per request

Each prompt was rewritten into a single search, for example "Austin TX dentists phone email official website directory". ChatGPT then summarized the pages that one search returned. A list built from one search is only as complete as that search, which for a city the size of Austin is a small slice of the market.

Finding 2: small lists, no file

Answers ranged from 5 to 20 rows. Every one ended with an offer to build a longer list "with all the details", but there was no file to download, no way to filter, and nothing to send to a CRM.

Finding 3: emails only when someone else already published them

ChatGPT, sensibly, didn't invent email addresses. For the Austin dentists it returned none and pointed to contact forms instead. For the Dallas roofers it found 8 owner or office emails, all copied from a business directory, a company contact page or a public procurement document. Nothing was verified.

Finding 4: the data disagreed with itself

Under its dentist table, ChatGPT also showed a map-style block of business listings. For 6 of the 20 practices, the phone number in the table didn't match the number in the listing directly below it. That's nearly one in three where a rep sees two numbers and can't tell which is right.

Where the answers come from

The business facts (address, phone, rating, review count) came from a maps-style listing block. The lists themselves were mostly lifted from directory and lead-list websites, several of them new sites with very little Google search traffic of their own (by Semrush estimates). ChatGPT is summarizing other people's lists, not building one.

What a lead list actually needs

  • Coverage: many searches across listings, websites, license boards and registries, not one.
  • The right person: the owner or manager, not a contact form.
  • Verification: every email checked with the mail server before you send.
  • Consistency: one phone number per business, cross-checked against its source.
  • A file: a complete list you can export, filter and re-run next month.

That's the job LeadLists is built for. ChatGPT is still excellent for deciding who to target. Use it for the strategy, then hand the list to an agent built for it.

Method note: responses collected November 17, 2026 via a ChatGPT answer-scraping API with web search forced on, US English, desktop. Four answers, one run each. Results vary between runs; treat this as a snapshot, not a benchmark.