AI prompts for prospecting: prioritise a new-business list

Before you start: what is in the file?

An Auspector export is a CSV or XLSX file with one row per record — a place, not a company. The column headers are in English whatever your interface language: name, category, address, locality (city), area (neighbourhood), phone, website, plus hours (opening hours) and lat/lng (coordinates) if you select them when exporting. Every row also carries a stable id (definitions: what is in a record).

• No email, no names. There is no email column and no named-person column. If an assistant comes up with an email address or an owner’s name, it did not come from your file — treat it as unverified (why we don’t sell email addresses).
• The date is a filter, not a column. To work on one day’s or one week’s new records, set the arrival date in the search before you export. It is the day the record reached us, not the day the business opened.
• Smaller files work better. OpenAI’s help centre recommends one record per row and clear column names, and says a large or complex file may not be analysed in full; if a result looks incomplete, split the file (OpenAI · Data analysis with ChatGPT). Ask for the calculation behind any total.

Eight prompts

1. Set the context (paste this first)

I sell [product or service] to [type of business].
Good fit for us: [criteria, e.g. category,
no website on record yet].
My territory: [city, neighbourhoods].

Attached: an export of newly opened businesses.
Columns: id, name, category, address, locality,
area, phone, website, hours, lat, lng.

Rules for this whole conversation:
- Use only what is in the file.
- If a value is missing, write “unknown”.
- Do not invent emails, names or facts.
- Keep the id column in every table.

What it does: Sets the rules once, so the later prompts can stay short. Paste it at the start of a new conversation, with the file attached.

Check: That it repeats the columns correctly and that its row count matches your file.

2. Rank by fit

Score each row for fit with our ideal customer:
3 = strong, 2 = possible, 1 = weak.
Use only category, area, website and hours.
Give a one-line reason for each score.
Return a table: id, name, category, area,
score, reason. Sort by score, then by area.

What it does: A first ranking against your own criteria, with a reason you can check.

Check: Spot-check five rows: does the reason rest on something that is actually in the row? A reason that mentions size, reviews or the owner did not come from the file.

3. Group by district for a route

Take the rows scored 3 or 2.
Group them by locality, then by area.
In each group, order the stops by lat and lng
so that each stop is close to the previous one.
At most [8] stops per day.
Return: day, stop, id, name, address.

What it does: Turns the shortlist into day-by-day routes for a field team.

Check: The route against the real addresses: a distance worked out from coordinates is a straight line, not travel time.

4. Draft an opening line that mentions the opening

For the 10 highest-scored rows, write one
opening line each for a [phone call / visit].
Mention that they have just opened and link
it to [our offer] in one sentence.
Do not state an opening date. Do not claim
anything that is not in the row.
Friendly, plain, under 25 words.

What it does: One personal line per business that refers to the opening.

Check: No invented facts or dates. The file tells you a business is new to our database, not the day it opened: “you’ve just opened” is safe, “you opened on Monday” is not. For the whole call, see a cold-call script for new businesses.

5. Research one business before a call

Search the web for [name], [address].
Answer in five bullet points:
- what they sell, and to whom
- whether an opening was announced (link)
- what their website says
- anything that makes [our offer] timely
- any sign that they are not open
Cite a source for each point.
Write “not found” instead of guessing.

What it does: A short briefing on one business before you call. It needs web search switched on.

Check: Open every cited source. OpenAI itself advises checking that a source supports the answer and when it was published or updated (OpenAI · Searching the web with ChatGPT).

6. Flag records to verify

List the rows we should check before contact,
with the reason:
- phone missing, or unusual for the country
- the same phone on several rows
- category unclear or not a fit
- website empty (ask; do not assume none)
- address without a street number
Return: id, name, issue.

What it does: A to-do list of the records that need a check before anyone contacts them.

Check: Duplicate numbers are not always errors: chains and service-area trades (plumbers, locksmiths, cleaners) often appear on several records with one number. What an empty field does and does not mean is covered in fill rates in business lists.

7. Summarise a day’s new records

This file holds one day's new records.
Count the rows by category and by area.
Show the 10 largest categories: how many rows,
how many with a phone, how many with a website.
Mark the ones that match our ideal customer.
Use code for the counts and show the totals.

What it does: A one-screen overview of a day’s arrivals in your territory.

Check: That the totals add up to the file’s row count. ChatGPT can write and run code for an analysis, and OpenAI recommends reviewing the code, outputs and assumptions before relying on the result (OpenAI · Data analysis with ChatGPT).

8. Prepare the shortlist for your CRM

Return a CSV with exactly these columns:
id, name, phone, score, reason, next_step.
next_step is one of: call, visit, research, skip.
Do not change any id value.

What it does: A clean file to import into your CRM.

Check: That the id values are unchanged: the id is what lets a re-import update rows in your CRM instead of duplicating them.

Three rules that keep this safe

• Verify before you use it. Assistants can be wrong and sound sure of themselves; OpenAI recommends treating ChatGPT’s output as a first draft rather than a final source, and checking important information (OpenAI · Does ChatGPT tell the truth?). Check the number and that the business is open before you contact it.
• A person sends every message. Do not connect an assistant so that it calls, emails or messages businesses on its own. Drafts are drafts. Rules on contacting businesses without consent differ sharply by country, and the obligation sits with the sender — check your own jurisdiction.
• Follow your company’s data policy. Before you upload a customer or prospect file, check what your company allows into which AI tool. Settings differ by plan and workspace: on personal ChatGPT Free, Plus and Pro accounts, conversations can be used to train OpenAI’s models by default, and you can turn that off in Data controls (OpenAI · Data controls in ChatGPT). This is not legal advice.

What an assistant can and cannot find on its own — and why it does not replace a dated list — is covered in can ChatGPT find new businesses?