Connfig

How to Use AI to Find Leads Without Getting a Made-Up List

Connfig TeamOctober 31, 2026ai-saleslead-generationguides
A map with blue pins connects to a chat bubble with a sparkle, which leads to three ranked contact cards; the top card has a check mark and a magnifying glass beside it

TL;DR: Ask a chatbot for "50 companies to sell to" and it fills the gaps with plausible guesses, including email addresses that don't exist. Give the assistant your ideal customer profile and real company data instead. Let it plan searches, score results against your rules and draft research notes, then check every fact before the first email.

To use AI to find leads, split the work in two. A real data source, such as Google Maps or a B2B database, supplies the companies. The AI does the thinking around them: it turns your ideal customer profile into searches, scores each result against rules you wrote and drafts a short research note for the best ones. You check the facts before anyone gets an email.

Why does a chatbot alone give you a bad lead list?

A hallucination is a confident answer that isn't true. OpenAI's research on why language models hallucinate (2025) says standard training and evaluation reward guessing over acknowledging uncertainty, and that rare, specific facts are where models go wrong. The email address of a 12-person dental clinic is exactly that kind of fact.

HubSpot's guide to ChatGPT for sales prospecting (updated 2026) gives the same warning: the model can hallucinate company details and misidentify decision-maker roles. Without web search or a connected tool, it also answers from training data, so a clinic that closed last spring can still appear on its list.

The cost lands on your domain. Addresses that don't exist bounce, and a run of hard bounces hurts a sender's reputation quickly, above all on a new domain that is still warming up.

Sales teams already see data as the limit. In Salesforce's State of Sales report (February 2026, 4,050 sales professionals), 55% said they use AI for prospecting, and 51% of sales leaders with AI said disconnected systems are slowing their AI initiatives.

Step 1: Write your ICP so the assistant can apply it

An ideal customer profile (ICP) describes the companies most likely to buy from you and stay. HubSpot calls it the foundation of all ChatGPT prospecting output. Write it as rules the assistant can check against one row of data:

Offer: website and Google profile setup for local clinics
Fit: dental and physiotherapy clinics in Manchester and Leeds
Quality: Google rating 4.0 or higher, at least 30 reviews
Signal: no website, or an unclaimed Google profile
Exclude: chains with more than 5 locations

Each line maps to a field you can actually get: category, city, rating, review count, website and profile status. Traits you can't read from a data row, such as "growing fast", belong in the research step. Our B2B ICP template shows how to build these rules from your best past deals.

Step 2: Let AI plan the search and a real source return the companies

AI is good at the planning work here. Ask it to expand your category into the terms people type into Google Maps ("dentist", "dental clinic", "orthodontist"), to split a region into cities and to estimate how many results per term you need. Then run the search on a real source and give the assistant the export.

You can also skip the copy and paste. MCP (Model Context Protocol) is an open standard that lets an AI assistant call outside tools. With Connfig's MCP server, Claude can start a Google Maps search itself, wait for it to finish and list the businesses that have an email address. A prompt like this works:

Search for dental clinics in Leeds, 50 results per term. When it's done, list the ones with an email address and a rating of 4 or higher.

Every company in the answer comes from the scan and traces back to a Maps listing you can open.

Step 3: Score and research with rules you write

Give the assistant your scoring table and ask for a score plus a one-line reason for each company:

RulePoints
Category matches the ICP+3
Rating 4.0+ with 30+ reviews+2
No website, or an unclaimed Google profile+3
Email address on file+2
Chain with more than 5 locations−5

Add one line to the prompt: "Use only the data I gave you. If a field is missing, write unknown. Never guess an email address or a person's name." A missing field then shows up as a gap to fill instead of a confident guess.

For the top 20, ask the assistant to read each website or the latest reviews, if it can browse, and to write one specific sentence you could open with. A clinic with a 4.8 rating and no website hands you that sentence. Our guide to cold emails that get replies covers what to build around it.

In Connfig, the Website: None and Google profile: Unclaimed filters in the Business Pool apply the signal rule before the list reaches the AI.

Step 4: Check every fact before the first email

HubSpot's guide calls human review of AI prospecting content mandatory before sending, and Salesforce found that 74% of sales professionals are focusing on data cleansing to get more from AI. For a prospect list, that work is short:

  1. Open the source. For each top-scored company, open the Maps listing or website and confirm the detail the AI used in its note.
  2. Keep only traceable emails. Delete any address that doesn't appear on the company's site or in your data, then run the rest through an email verification tool.
  3. Remove duplicates. Check the list against your CRM so a current customer doesn't get a cold pitch.
  4. Read every first line. Fix anything that sounds like a template or states something you couldn't confirm.
  5. Send from a prepared domain. Publish SPF, DKIM and DMARC, the DNS records that prove the mail is yours, and keep early volumes low.

FAQ

Can ChatGPT find leads on its own?

It can suggest company types and search terms, and with web search on it can point you to pages you can check. Treat any list it writes from memory as a draft, because names may be out of date and contact details may be invented. The reliable setup pairs the assistant with a data source and your own rules.

Which AI tool is best for finding leads?

The model matters less than what it can reach. An assistant that reads your export or calls a data tool over MCP works on real records. One that can't reach your data works on guesses, however good the model is.

To run step 2 from your own assistant, follow the MCP setup guide; the Free plan includes 25 businesses a month to test it.

Connfig'i ücretsiz deneyin

Lead bulma, CRM ve e-posta kampanyalarını tek platformda toplayın. Kart bilgisi gerekmez.

Ücretsiz Başla

Related Posts