Type "List 20 lawn-care companies in Raleigh with the owner's email" into ChatGPT and you will get 20 lines that look exactly right. Some of those companies will be real. Some will be real but in another city. A few will not exist, and every email address will be a guess at a pattern. That is the whole problem with using ChatGPT for lead generation, and it is worth understanding before you spend a week on a list it made up.
ChatGPT is good at thinking, writing and sorting text you give it. It is not a source of leads, because it has no live company database, it cannot confirm that a person works where you think they do, and it cannot send anything. Use it to describe your buyer, to find search terms, to draft first lines from facts you paste in, and to draft replies. Do not use it to supply the names. This page gives the prompts, shows the problem, and then shows how to check anything it tells you.
What ChatGPT is good for in lead generation
Four jobs, all of which work on text you already have. Turning a vague idea of your customer into a precise description. Producing search strings you can run yourself. Drafting a first sentence from a fact on a company's own page. Drafting a reply to a message someone sent you. In each, ChatGPT works from your input, and you check the output against reality. That is the pattern to keep: you supply the facts, it supplies the wording.
Prompts that describe your buyer precisely
A vague buyer gives you a vague list, whatever tool builds it. This prompt forces specifics. Fill in the brackets with real details about customers you already have, with no names.
"I run [a payroll service for small businesses in North Carolina]. My three best customers are [a 12-person lawn-care company, a 20-person dental practice and a 9-person HVAC contractor]. Describe the buyer in one paragraph. Then list five job titles that usually sign, five events that make them need this in the next 90 days, and five phrases they would use to describe the problem in their own words."
The output is a draft. Cross out anything that does not match what you have seen. The "events" list is the most useful part, because each event (a first hire, a new location, a missed deadline) is something you can search for. If you do not yet have a clear picture of your buyer, what is an ideal customer profile and how to find your ideal customers help first.
A second prompt turns that description into searches you run yourself:
"Give me ten Google search queries and five job-title searches I could use to find [small lawn-care companies with 8 to 25 employees] in [Wake County]. Do not list any company names."
Prompts for first lines and replies
This is where ChatGPT earns its keep, with one rule. Paste the real text and tell it to use only that.
"Below is text from a company's About page. Write three possible opening sentences for a cold email, each under 20 words, that mention one specific fact from the text. No flattery. Use only the text I pasted. If a fact is not in the text, do not use it. [paste text]"
The line "use only the text I pasted" is the guardrail. Without it, the model fills gaps with plausible details, such as a founding year or a recent award that never happened. For more on writing first lines without spending an hour on each, see personalisation at low volume.
For replies, paste what the person wrote and ask for options:
"Here is a reply from a prospect: [paste]. Draft three replies, each under 60 words. Answer their question directly, do not oversell, and end with two specific times I could talk. Do not invent prices or facts."
You will still edit these. The saving is the blank page, not the judgment.
The invented-company problem
Language models produce text that is likely to follow, not text that is checked. Ask for companies and it produces the shape of a company: a believable name, a believable city, a believable domain. Ask for emails and it produces the common patterns (firstname@company.com) with no knowledge of whether the mailbox exists. A model with web access can look things up, but even then you are trusting a summary, and it still cannot confirm that a specific person's address works.
Take Anita, who runs a payroll service in Raleigh and asks for 20 lawn-care companies. In a test like hers, the results will vary by model and by day, but the pattern is steady. Some names check out. Some exist in the wrong place. Some cannot be found at all. The danger is not that the list is bad; it is that it looks finished, so it goes straight into your outreach tool and your sender reputation pays for the bounces. Sending to invalid addresses harms deliverability, which is hard to repair.
Verifying anything it names
If ChatGPT gives you a company or a person, run this check before anything is added to a list. It takes about three minutes per company, which is roughly an hour for 20, or $40 of your time at $40 an hour.
- Search the company name and city. Does a website or a map listing appear?
- Open the site. Does the address match? Is the business still operating (recent news, a working phone number)?
- Find the person on the company's own page or a professional profile. Does the title match?
- Take the email from the site or a source you trust, not from a pattern. Run it through an email verification check.
- Record where each fact came from. A name with no source stays out.
If you are checking more than a few dozen, the checking is the job, and it is the reason to use a tool built for it. AI lead generation for local businesses covers what the better tools do differently.
Where a tool with live data is needed
You need a source that looks people up when you ask and checks that they exist. That means a lead database or a worker connected to one. In Rocketship, you describe your buyer in plain words, see the first 25 matching people by name before spending a credit, and add the ones you like. Launch includes 50 leads a month. It then emails them from your own Gmail, follows up and books the meeting, which ChatGPT on its own cannot do. If you already live in Claude, the same job runs in a chat through the Rocketship connector for Claude, which how to find clients using Claude walks through.
A simple split of the work
ChatGPT: your buyer description, search strings, first-line drafts, reply drafts. A tool with live data: the list of real people, their contact details, the sending, the follow-up and the booking. You: the judgment about who fits, the edits to every draft, and the conversation. Keep to that split and ChatGPT saves an hour or two a week. Ask it to be the database, and it costs you a week and a sender reputation.
When you have a verified list and need to know how many are enough, the best way to find B2B leads for a small company compares the ways to get one, and cold email examples shows what to do with the opening sentence once you have it.
Rocketship finds the buyers, emails them from your own Gmail, follows up, and books the meeting on its own. Plans that get you customers start at $24.99 a month (Launch), and every monthly plan starts with a 7-day free trial, card required.