How to ask an AI assistant which channel brings your best leads

Yes, if it can see each lead's source and whether the lead became a customer. It cannot work out what your CRM (the system where you keep contacts and deals) never recorded.

Imagine a CRM with 200 leads from the last six months, where 140 say "unknown" in the source field. This is an invented example. The assistant will still produce a ranking, but it covers only the 60 leads that have a source, and those 60 may look nothing like the other 140.

If most of your leads show "unknown", the fix is to record the source at the form. Tools that track lead source do this, and TrackLeadSource, a lead source tracking tool, records where a visitor came from and which pages they viewed. It passes that with the form enquiry to your CRM or spreadsheet, once the form fields are connected to your CRM or sheet. Each enquiry then arrives with its source filled in, so the assistant has something to rank.

This guide covers asking an assistant to analyse your own leads. Visitors who arrive from ChatGPT are covered in the guide on tracking ChatGPT traffic and leads.

I have not run these prompts against a live CRM for this guide. The prompts are examples to adapt, and the CRM details come from each vendor's own help pages, read on 29 October 2026.

Each lead needs three things: where it came from, which pages it viewed, and what happened to it in sales. The first two come from your website tracking. The third comes from your team.

The UTM parameters guide explains source, medium and campaign in more depth. The CRM lead source guide covers keeping those fields consistent once they reach the CRM, and the lead quality guide covers what counts as a good lead.

Your team adds the sales outcome, because the CRM or sheet is where that lives. The website form tracking guide covers how the source reaches your forms.

Start with the export. It takes a few minutes, needs no approval and works with any CRM. Move to a connector when you ask the same questions every month on a large CRM.

A connector is a built-in link between your CRM and the AI assistant, made by the CRM company. With an export, you download the leads as a file and give the file to the assistant.

Some connectors follow your own CRM permissions, so the assistant sees only what your login could see. The HubSpot and Attio pages describe this. If your CRM offers a read-only option, choose it for analysis.

Most CRMs have an Export button on the leads or deals list. Ask for one row per lead with these columns: source, medium, campaign, created date, status, won or lost, deal value, and pages viewed. If the file includes names, emails or phone numbers, delete those columns in a spreadsheet before you upload it, unless the question needs them. Check that your assistant plan accepts file uploads, as plans differ.

If you need help from colleagues, two short messages cover it.

To your CRM admin: "Could you give me read-only access to our leads and deals for an AI analysis, or turn on the CRM's AI connector in its read-only option?"

To whoever manages the website: "Please record each visitor's source, medium, campaign and pages viewed on our contact forms, and send them to the CRM with each enquiry."

Ask for numbers per channel, ask for the number of leads behind each rate, and tell the assistant to say what it cannot tell from the data. Add one rule to every prompt: "Answer only from the data I gave you. If something is not in the data, say 'not in the data'."

Here is a tiny invented sample of what an export could look like. A real file will have hundreds of rows.

Prompt 1: compare channels by customers

Here are my leads from the last 6 months with columns: source, medium, campaign, created date, status (open, qualified, won, lost) and deal value. For each source and medium, give the number of leads, how many became customers, the percentage, and total deal value. Rank them by customers. List every group with fewer than 20 leads separately and say which conclusions are too weak to rely on.

It needs source, medium, campaign, status and deal value. A useful answer is a table with the lead count beside every percentage, and a separate list of small groups. To check it, add up the won leads for two channels yourself and compare.

Prompt 2: find the enquiries that go nowhere

Which sources sent the most leads with status "lost" or no activity in 60 days? For each, list the pages-viewed values that appear most often, and tell me whether those leads share a first page they land on or a path that the won leads never took.

This needs the pages-viewed text in each record, which is why recording it at the form matters. A useful answer names a pattern, such as lost leads from one source all stopping at the same page. Treat a pattern as a lead to investigate, because page views do not show what a person intended.

Prompt 3: propose a budget move

I spend about [amount] a month across Google Ads, Meta Ads and email. Using only the data above, which channel would you spend less on and which more, and what extra data would you need before I act? Do not assume anything about cost that I have not given you.

A useful answer lists what is missing, such as cost per channel or how long it takes a lead to buy. If it comes back with invented costs, the last sentence of the prompt was left out.

The numbers 20 leads and 60 days in these prompts are example thresholds. No fixed number makes a group big enough to trust, so pick a floor that suits your volume and use it everywhere.

A worked example of the decision

Say the assistant reports this for the last six months (invented numbers):

Google sent fewer leads but six times as many customers. One customer from Meta is thin evidence, so a sensible next step is a small shift. Move perhaps a tenth of the Meta budget to Google for one month, then compare the new figures before moving more.

Five CRMs have an official vendor page for connecting AI assistants: HubSpot, Pipedrive, Close, Attio and Microsoft Dynamics 365 Sales. Most use MCP (Model Context Protocol), a standard way for an AI assistant to connect to another tool. The table is as of 29 October 2026, and statuses such as beta or preview change often, so read the vendor's page before you rely on one.

Salesforce and ActiveCampaign appear to have connector pages too, but I could not open them on 29 October 2026, so check them directly. I found no official Freshsales page. For any other CRM, search its help centre for "MCP" or "AI connector".

If you use HubSpot, the HubSpot UTM guide shows how the source reaches the contact record, which the assistant then reads.

The only provider policy I read for this guide is Anthropic's. I did not read OpenAI's terms, so nothing in this section covers ChatGPT. Beyond that, safety depends on the account type and your company's rules: use a paid team or business plan, not a free personal account, and send only the fields the question needs.

Anthropic's privacy page (dated 18 August 2026, read 29 October 2026) says that by default it does not use inputs or outputs from its commercial products, such as Claude for Work and the Anthropic API, to train its models. It lists two exceptions: feedback you submit, and data you choose to let it use. Check the terms for your ChatGPT plan yourself.

Three habits cut the risk. Take names, emails and phone numbers out of exports when the question is about channels. Use a read-only connection where the CRM offers one. Ask whoever owns your contracts and data rules before you connect a live CRM. This section is not legal advice.

Check the arithmetic against the CRM, look at the number of leads behind each figure, then change one thing and watch the result.

Start with the sums. Pick two channels and count their won leads in the CRM. If your total differs from the assistant's, ask it to show which rows it counted.

Next, read the sample sizes. Two customers from five leads is 40%, which is thin evidence. Treat small groups as "not enough yet", and give every lead the same time to close before you compare. The lead quality guide explains how long to wait.

Then settle which source you are counting. A person who finds you through search and returns through an ad has two sources, and the lead attribution guide explains how to choose between the first and the latest. Leads marked direct or unknown deserve their own line, and the guide on missing UTM parameters covers why they appear. If you ask people how they heard about you, keep that answer in a separate column. The survey options guide covers wording.

When the answer holds up, move a small part of the budget and watch the next month before moving more.

TrackLeadSource does not run the AI or score leads. Setup is a one-time job for whoever manages your website, who adds a small piece of code once, and connecting the form data to your CRM or sheet usually goes through Make.com or Zapier. Our guides to adding source fields cover Contact Form 7, WPForms and Gravity Forms.