ChatGPT for Sales Prospecting: A Source-Based Account Research Workflow

TL;DR: Use ChatGPT to organize evidence you supply, not to invent account facts. Give it approved sources, dates, an ICP, and a strict output schema. Review every claim, preserve unknowns, then move only approved domains into a documented GTM workflow for contact validation, CRM checks, and message previews.

How do I use ChatGPT for sales prospecting?

Use ChatGPT as a source-analysis workspace. Supply the pages or excerpts you approve, ask for a structured research table, and require every fact to retain its source and date. Keep hypotheses in separate fields, label missing evidence Unknown, and do not ask the model to invent contacts, work emails, technology use, CRM status, or buying intent.

After human review, pass only approved company domains and research notes into a system with documented discovery, enrichment, and CRM access. That second system can validate current roles and work channels, check ownership and protected states, and prepare message previews. Research, enrichment, drafting, enrollment, sending, and CRM writing must remain separate steps.

This boundary matters because a general assistant can reason over material it can access, but it does not automatically have access to your CRM, mailbox, product data, or paid prospecting sources. A polished answer is not evidence that a connection exists.

Build a source packet before opening ChatGPT

A source packet is the approved evidence set for one account. It should be small enough to review and complete enough to answer a defined business question. Start with authoritative first-party pages whenever possible, then add current reporting or regulatory material only when it contributes a specific fact.

• Company identity: legal or operating name, canonical domain, and any parent or subsidiary relationship
• Research question: the operational change or function you are trying to understand
• Approved sources: titles, canonical URLs, dates, and the relevant excerpt or section
• ICP definition: the fit criteria and disqualifiers the research must use
• Output schema: required columns, null rules, and the distinction between facts and hypotheses
• Action boundary: no contact creation, outreach, enrollment, sending, or CRM write

Do not ask ChatGPT to “research this company” without defining source and output constraints. An open-ended instruction encourages the model to optimize for completeness instead of traceability.

Use a copy-ready ChatGPT research prompt

Make the prompt enforce the evidence model. The goal is not a persuasive narrative. It is a table that a reviewer can inspect row by row.

ChatGPT research prompt:

Using only the sources and excerpts below, create an account research table with company, source title, source URL, publication or observation date, exact supported fact, relevance to our ICP, possible operating implication, unsupported assumptions, missing evidence, contradictory evidence, and next research question. Keep facts and hypotheses in separate columns. If a claim is not supported by the supplied source, write Unknown. Do not invent contacts, emails, technology use, budget, purchase intent, internal priorities, or quotes. Do not draft outreach yet.

Paste the ICP and source packet after the instruction. If you use ChatGPT deep research, the same rules still apply. OpenAI describes deep research as a way to search, analyze, and synthesize sources, but the user remains responsible for deciding which sources and conclusions are suitable for a sales workflow.

Ask the model to identify disagreement rather than reconcile it automatically. A company page, filing, and current job posting may describe the same function differently. Preserve the conflict and ask which source is newer or more authoritative.

Review the table before moving any data

The first review is a source check, not a copy edit. Open every decision-relevant URL and confirm that the cited location supports the wording in the table.

• Confirm that the source belongs to the right entity
• Check the publication, filing, or observation date
• Remove facts that cannot be located on the cited page
• Keep recurring and newly changed language distinct
• Preserve contradictions instead of blending them
• Replace unsupported confidence with Unknown
• Approve only the domains and facts permitted for prospecting

A source-backed row may still be irrelevant to the ICP. The reviewer should reject rows when the evidence is current but the account, function, geography, or business model does not match the sales motion.

Move approved rows into a documented GTM workflow

ChatGPT analysis and Unify MCP execution are two separate lanes. The inspected Unify MCP setup guide documents Claude Code, Cursor, and Codex as supported clients. It does not establish direct Unify MCP support inside ChatGPT, Grok, or Groq. Do not describe copying a public workflow into ChatGPT as a live connection to Unify.

In a documented client, our MCP workflow can run hosted research, discover accounts and contacts, perform bounded enrichment, check CRM context, and load results into a DataTable. A run may ask clarification questions, and those questions should be answered before it continues.

Unify run_agent prompt:

Using the approved company domains and source-backed research notes I provide, find up to two current people per company whose documented role matches the operating function in the research. Verify current employment and work email where possible. Check our CRM for account owner, customer status, open opportunities, recent activity, and protected states. Return a DataTable with company, domain, approved source fact, operating hypothesis, contact, title, role evidence, work-email status, CRM owner, CRM exclusions, confidence, and source references. Cap paid enrichment at 40 records. Preserve Unknown fields and contradictory evidence. Do not create CRM records, draft outreach, enroll contacts, or send messages.

The prompt specifies scope, output, budget, exclusions, and a no-write boundary. After the run completes, read the results and inspect the returned rows. Do not treat successful execution as proof that every row is correct.

Validate current roles and work channels

A source packet can explain why an account deserves research without identifying the right person. Contact selection should begin with the affected function, then verify current employment and reachable work channels.

• Choose a role because its documented responsibility matches the operating question
• Confirm the person currently works at the target company
• Verify work email only when the next approved step requires it
• Cap paid enrichment before the run begins
• Retain confidence and evidence date for each contact field
• Do not infer budget authority from seniority alone

Our AI agent research mechanics guide explains how planning, discovery, entity resolution, extraction, and verification fit together. The important point is that generated prose cannot substitute for current identity and source checks.

Check CRM ownership before drafting

Before a message is prepared, determine whether the account is a customer, an open opportunity, actively owned, recently contacted, opted out, or part of another motion. These states are not minor details. They decide whether the row can move forward at all.

Use one arbitration rule for conflicts. If a protected state, ownership disagreement, or open opportunity exists, hold the row until the accountable owner decides the next action. The model should not choose between competing teams.

The operating pattern in Prevent Duplicate Outreach Across Reps, Sequences, and Lifecycle Plays shows how suppression and ownership reduce overlapping contact.

Draft from approved facts only

Once a row passes source, contact, and CRM review, prepare a message preview. The message should cite or clearly refer to the approved fact, identify the function that may be affected, and ask a bounded question. It should not claim that the account is “struggling,” “actively evaluating,” or “ready to buy” unless a source actually supports that statement.

A useful preview has four parts:

• Observed fact: one source-backed change or initiative
• Reasoned relevance: why teams in that function sometimes revisit a process
• Validation question: whether the issue is relevant and who owns it
• Low-pressure next step: an invitation to compare notes, not a fabricated deadline

Ask for previews only. A preview sends nothing. An approved draft still sends nothing. Enrollment or sending is a separate execution step that requires explicit authorization.

Measure correction effort, not generated volume

A source-based workflow should make errors visible. Track whether reviewers can reproduce claims and how much work is required to correct the result.

• Rows with exact source and date retained
• Claims removed because the source did not support them
• Entity conflicts caught before enrichment
• Contacts rejected because employment or role was stale
• Protected CRM states found before drafting
• Unknown fields preserved instead of guessed
• Drafts accepted after human review
• Correction time per approved account brief

Do not use model output volume as a proxy for pipeline. A completed research table is not a qualified account, a draft is not a sent message, and a reply is not an accepted opportunity.

Frequently asked questions

Can ChatGPT find sales prospects?

ChatGPT can help structure sources you provide or sources it can access through a documented research mode. It should not be treated as proof of current contacts, work emails, CRM state, or buying intent.

What should I include in a prospecting prompt?

Include approved sources, dates, excerpts, ICP criteria, disqualifiers, required columns, null rules, and a no-outreach boundary.

How do I prevent ChatGPT from making up company facts?

Require exact source URLs and supporting locations, separate facts from hypotheses, preserve contradictions, write Unknown for unsupported fields, and review every important claim.

Does Unify MCP run directly in ChatGPT?

No inspected first-party setup guide documents direct Unify MCP support in ChatGPT. The current guide names Claude Code, Cursor, and Codex.

When should I use Unify after ChatGPT research?

Use a documented Unify client after approving the company domains and source facts, when you need bounded discovery, enrichment, CRM checks, DataTable results, or message previews.

Can the workflow send automatically?

Not from research or preview steps. Enrollment, sending, and CRM writing are separate actions that require explicit approval.

Start using Unify to turn approved research into verified, reviewable prospecting work without collapsing analysis into execution.

Sources

• Deep research in ChatGPT, OpenAI
• Getting started with Unify, Unify
• Unify Chat: Use Cases, Unify Knowledge Base
• How AI Agents Actually Research Prospects, Unify
• Prevent Duplicate Outreach Across Reps, Sequences, and Lifecycle Plays, Unify