Enrich lead lists with AI ChatGPT, Claude Code, and Codex, fed with real data

Feed your AI with data, not vibes

Ask any model to "give me 50 dentists in Madrid with emails" and it will invent a plausible list. The names look real, the phone numbers are fiction. Language models are built to predict text, not to know the current phone number of a clinic. That gap is why enrichment starts with a real source of truth.

The fix is simple: capture the businesses first, then hand that structured file to the AI. When a model reads a column of real ratings, review counts, categories, and websites, it stops guessing and starts reasoning. This is the whole premise behind feeding your AI with business data.

The Vonsel extension is the means to get that file. You browse a Google Maps search, capture the results, and export them. The AI is where the enrichment happens. The extension is the source, your AI tool is the workshop, and the Vonsel dashboard is where the finished leads go to work.

Export a clean file from the extension

Open Google Maps, search for something like "roofers in Austin" or "hair salons in Barcelona", and let the Vonsel pop-up capture the listings. Every business comes with name, address, phone, website, category, rating, and review count. You can pick the format that suits your AI tool from the CSV, XLSX, and JSON export options.

Rule of thumb: use CSV or Excel when you plan to upload the file to ChatGPT, and use JSON when Claude Code or Codex will read it, because nested fields survive better. If you are new to this, the export formats for AI guide covers when each one shines.

A clean, consistent header row matters more than volume. Well-structured columns are what let a model line up "rating" against "review count" and reason about them. Raw text dumps produce raw text answers. Structure in, structure out.

ChatGPT: upload the file and prompt it

For most people, ChatGPT is the fastest route. Open a new chat, attach your exported CSV or Excel file, and describe what you want. Its data-analysis mode reads the spreadsheet directly, so you can ask questions in plain language and get a downloadable enriched file back. OpenAI documents the underlying behavior in its text generation guide.

Good starter prompts once the file is attached:

• "Score every business from 1 to 5 on how likely they need a new website, using rating, review count, and whether a website is present."
• "Add a column that tags each lead by size: solo, small, or established, inferred from review count."
• "Flag businesses with a strong rating but very few reviews. Those are underexposed and easy to pitch."
• "Draft a one-line, personalized cold email opener for each row based on its category and city."

The workflow is covered end to end in uploading a lead CSV to ChatGPT and in analyzing lead lists with ChatGPT. Keep batches under a few hundred rows per chat so the model stays accurate, then merge the pieces.

Claude Code: enrich a whole folder on your machine

When one file turns into fifty, ChatGPT's chat window becomes a bottleneck. Claude Code runs in your terminal, so it can read a folder of exports, loop through thousands of rows, and write the enriched results back to disk. It is the difference between asking a question and running a repeatable pipeline.

Drop your JSON or CSV exports into a folder, open Claude Code there, and instruct it directly: "Read every file in ./leads, add a lead_score and a website_gap column to each, and save the results to ./enriched." Because it works on local files, nothing gets truncated and you can rerun the same instruction on next week's scrape. The details live in analyzing leads with Claude Code.

This is also where you can build reusable prompts. Save a well-worded enrichment instruction once and apply it to every future export. A starter set is collected in Claude Code prompts for lead databases.

Codex and local models for scripted or private work

Some teams want enrichment wired into code rather than run by hand. The Codex CLI fits here. Point it at your exported file and ask it to write a small script that cleans phone formats, deduplicates by address, and appends a scoring column. You review the script, run it, and now you have a repeatable enrichment step that lives in your repo.

If the data is sensitive or you simply prefer to keep it off the cloud, a local model works on the same exported file. Nothing leaves your laptop, and the JSON export stays the exact same input. See using local LLMs with lead data for that path. The common thread across ChatGPT, Claude Code, Codex, and a local model is the same clean file underneath.

From an enriched file to a working pipeline

Enrichment is not the finish line. A scored CSV sitting on your desktop still has to become conversations and meetings. That is the role of the Vonsel dashboard: a mapped CRM where your captured businesses live, with Sales, Reviews, and Email Intelligence layered on top. The AI enrichment you did upstream becomes the segmentation and priority you act on downstream.

Inside the dashboard, an AI email is generated per business, and databases are contextualized with the pain points surfaced in reviews. So the "website_gap" flag your model produced maps neatly onto a real outreach angle. A raw export is a starting point; a value-added database is the finished product, as argued in value-added database vs raw CSV.

One more habit worth keeping: scrape responsibly and only work with business contact data. A quick read of whether it is legal to scrape Google Maps keeps your enriched lists on solid ground before you email anyone.

Do not let the model invent the data