ChatGPT prompts for lead lists Feed your AI with real business data

Data first, prompts second

Everyone wants ChatGPT to write cold emails and score their pipeline. The part they skip: an AI model has no idea who your local prospects are unless you hand it a real list. The prompt is the easy half. The data is the half that decides whether the output is any good.

That is where a Google Maps capture comes in. The Vonsel Chrome extension pulls business names, addresses, phone numbers, websites, categories, ratings, and review counts straight from a search, then exports them as a clean file. That file is what you feed the model. If you have never done a capture, our how to scrape Google Maps guide walks through it in a few clicks.

The extension is the means. The end is a worked list: deduped, scored, segmented, and ready to act on, whether you do that inside your Vonsel dashboard or in a chat window with your favorite model.

Export the right file, then upload it

From a finished capture, the extension gives you three formats. Pick based on the job, not habit. Our note on CSV, XLSX, and JSON export options covers the trade-offs, and there is a deeper piece on JSON vs CSV for AI analysis.

In ChatGPT, drag the file into the chat or use the upload button, and the model reads it directly. For CSV under a few hundred rows you can also paste the raw text. Full walkthrough in how to upload a lead CSV to ChatGPT. Whichever route you take, always tell the model your exact column names up front so its output lines up with your file. See the OpenAI prompt engineering guide for why specific instructions beat vague ones.

Five ChatGPT prompts that earn their keep

1. Clean and dedupe. "Here is a CSV of businesses with columns name, address, phone, website, category, rating, reviews. Remove exact and near-duplicate rows, normalize phone numbers to E.164, fix inconsistent capitalization in name and address, and return a cleaned CSV plus a short summary of what you changed." More on this in clean lead data with ChatGPT.

2. Score against your ICP. "Score each business 0 to 100 as a prospect for a web design agency. Weight: no website (40), rating below 4.0 (20), fewer than 30 reviews (20), category relevance (20). Add a score column and a one-line reason, then sort descending." The rubric is what makes it repeatable. See score leads with ChatGPT.

3. Segment by niche and area. "Group these leads into segments by category and by city district. For each segment give a count, the average rating, and a one-sentence outreach angle." Great prep before you split the work.

4. Draft outreach. "Write a three-line cold email for the top 10 rows. Reference their category and the fact that they have no website where relevant. Keep it under 70 words and avoid hype." Pair it with writing cold emails with ChatGPT from scraped data.

5. Find the patterns. "Summarize this list: what categories dominate, which areas are underserved, and where the best-rated competitors cluster." A quick way to turn rows into a market read. Deeper flow in analyze lead lists with ChatGPT.

When the list is big, reach for Claude Code

ChatGPT is perfect for a few hundred rows in a chat window. Once you are at thousands of businesses, or you want the same treatment every week, move to the terminal. Claude Code runs against your actual export files on disk, writes a reusable script, and saves the cleaned output where you want it.

Point it straight at the file: "Read leads.csv from the Vonsel export. Dedupe on name and address, add a boolean has_website column, score each row 0 to 100 for a solar installer selling to homeowners, and write the result to leads_scored.csv." Because it can run code, it handles ten thousand rows without truncating, and it shows you the script so you can run it again next month. Our analyze leads with Claude Code and Claude Code prompts for lead databases guides go further.

This is where JSON exports shine. Hand Claude Code the JSON version of your capture and it can keep nested review data attached to each business, then build a small dashboard or map from it without you touching a spreadsheet.

Which one for which job

Not either or. Many people prototype a prompt in ChatGPT, then hand the finished logic to Claude Code to run at scale. If you want to compare them head to head, read ChatGPT vs Claude for lead analysis.

Codex and local models

If you already live in an IDE, coding agents like Codex can enrich a lead file the same way Claude Code does, straight from the editor. That path is covered in enrich lead lists with Codex.

Prefer to keep prospect data off third-party servers? A local model on your own machine can run the same clean-score-segment prompts. It is slower and blunter than a frontier model, but nothing leaves your laptop. See use local LLMs with lead data. Whatever you choose, remember these are business contact details: our note on whether scraping Google Maps is legal is worth a read before you go to scale.

From capture to a CRM you can work

AI analysis is powerful, but a chat transcript is not a workflow. Once ChatGPT has scored and segmented your list, you still need somewhere to track status, plan visits, and send outreach. That is the job the Vonsel dashboard does: a mapped CRM with Sales, Reviews, and Email Intelligence, plus an AI email drafted per business and contextualized, value-added databases built around each prospect's pain points.

The extension also lets you reopen a previous capture and add more businesses, attach reviews to captured records, and keep a history of every scrape, so the AI-worked list keeps growing instead of going stale. For the raw-data-versus-usable-database angle, compare notes with Google Maps data vs bought lists.