Use local LLMs with lead data Export from Google Maps, then feed your AI

Feed your AI with real business data

Large language models are only as useful as the data you give them. Ask ChatGPT to "find dentists in Madrid" and it will guess. Hand it a clean file of real businesses you captured yourself, with names, phones, websites, ratings, and review text, and it becomes a genuine analyst.

That is the whole point of the Vonsel Chrome extension. You run a search on Google Maps, capture the results, and export them as CSV, Excel, or JSON. The file is the bridge between the map and any model you want to point at it. No API keys, no scraping code, no copy and paste.

In this guide we focus on the two tools most people actually reach for, ChatGPT and Claude Code, then cover Codex and local LLMs for anyone who wants private or scripted runs. For a deeper primer on the concept, see feed your AI with business data.

Export the right format from the extension

Open Google Maps, search for a category and area, and let the Vonsel pop-up capture the listings. When the capture finishes you choose a format. Each one suits a different kind of AI work, so pick with the destination in mind.

A useful rule of thumb: CSV or Excel when a human or a chat window reads the data, JSON when a script does. If you want the full breakdown of when each shines, read JSON vs CSV for AI analysis and the practical export leads for AI walkthrough.

ChatGPT: upload the file and ask

This is where most people start, and for good reason. Open ChatGPT, attach your CSV or Excel export, and treat the file as a table the model can read directly. Its data analysis tools can group, sort, and score without you writing a formula. See OpenAI's developer docs for how it handles structured input.

Good first prompts once the file is uploaded:

1. Segment. "Group these businesses by rating band (under 3.5, 3.5 to 4.2, above 4.2) and count how many in each have no website." That instantly surfaces the web design prospects hiding in your list.

2. Score. "Rank every row by how likely they are to need marketing help, using low review counts and missing websites as signals. Return the top 25 with a one line reason." Now you have a call sheet, not a spreadsheet.

3. Draft. "Write a short, specific first line for a cold email to each of the top 10, referencing something real from their data." You feed the AI, the AI hands back openers you can actually send.

For a step by step version of this, our upload a lead CSV to ChatGPT guide walks through the exact clicks.

Claude Code: run scripts over your JSON

When a chat window is not enough, Claude Code turns your export into a working dataset. Because it runs in your terminal and can read files, it shines with the JSON export where every business carries nested reviews, hours, and social links. Point it at the file and it writes and runs the code for you. The Claude Code agent handles the parsing so you never touch a library.

Things that are trivial in Claude Code and painful by hand:

For deeper recipes, see analyze leads with Claude Code. Together, ChatGPT for fast questions and Claude Code for scripted work cover the large majority of what you will do with a scraped list.

Codex and local LLMs for offline runs

Two more paths matter when privacy or automation is the priority. Codex works much like Claude Code, an agent in your terminal that reads the export and writes the enrichment scripts. If your team already lives in that toolchain, our enrich lead lists with Codex guide picks up there.

Local LLMs are the answer when you would rather not upload business contacts to a hosted service at all. A runtime like Ollama lets you run an open model on your own machine, then feed it the same CSV or JSON. The data never leaves your laptop, which matters if you handle regulated verticals or client lists under strict agreements.

Local models are smaller and slower than the hosted giants, so keep tasks focused: classify a column, summarize a batch of reviews, tag leads by niche. Do the heavy reasoning in ChatGPT or Claude Code, and reserve the local model for the runs you want to keep entirely offline. On the legal side, scraping public business data is generally defensible, but always read is it legal to scrape Google Maps before you build a pipeline on it.

The export is the means, the dashboard is the end

Running a scraped file through an LLM is powerful, but a chat transcript is not a sales system. Insights sitting in a terminal do not book meetings. This is why the extension is only the front door to Vonsel.

Everything you capture also flows into the Vonsel dashboard, a mapped CRM where those leads live on a map, not in a stray CSV. There, Sales, Reviews, and Email Intelligence do the analysis for you: reviews are mined for pain points automatically, and an AI email is drafted for each business using its own context. The value added, contextualized database is built as you scrape, not bolted on later.

So use ChatGPT and Claude Code freely to explore and experiment. When you are ready to actually work the pipeline, reopen a scrape to add more businesses or attach reviews to captured ones, and let the dashboard carry the leads from first line to first meeting. The AI is where you learn. Vonsel is where you sell.