How to Create An AI Lead Magnet
A guide for email marketers
A while back I was brainstorming ways to use AI tools like GPT for marketing purposes. It turns out that a lot of the obvious ideas around lead generation, market research and automated outreach aren’t quite possible yet, but probably will be soon when LLM providers start adding agent functionality.
One thing I did come up with– and even build a prototype of– is the idea of using AI to offer people a personalized lead magnet.
For those who don’t know, a lead magnet is a free information product you offer people in order to entice them to join your mailing list. “Sign up for your free five-day graphic design course” or something like that. You can also use this for other purposes, which I’ll get to.
So, here’s how you can use AI to send people a personalized report, action plan, recommendation list, or whatever kind of thing, and then use that for marketing purposes.
Step one: create a survey using Google Forms. This survey serves two purposes: the information it provides will be used to generate a prompt for GPT which in turn will generate the user’s personalized thingamajig, but also you want to use this form to learn whatever it is you want to know about the user.
This form also needs to collect their email address, and ideally at least their first name as well.
Step two: Create a workflow using either Zapier or Make. Zapier is a lot easier to use, but Make costs a lot less. For something like this, where each person is only using the thing once, I recommend Zapier.
Use this workflow to connect the Google Form to the OpenAI API and turn the responses from the survey into a prompt for GPT-4, send that prompt in to the API, and then receive the response.
Step three: In your Zapier workflow, connect a Gmail account and use it to email the response from GPT to the user. You can customize this email to include a standard statement from you to the user (using their first name if you got it) before and after the GPT response, so the email can begin like “Hey Steve, it’s John. Here’s your customized workout plan.”
Step four and maybe also five: Add steps in Zapier to add the user’s information to a Google Sheet and/or your mailing list.
You can use the information you got from the form to segment your mailing list. For instance, using my prototype, you could segment a fitness mailing list by age, sex, level of training experience, training goal, or favored body part.
Now that you know this is the end point, take another look at step one: you want to design that form to capture whatever information you would want to know about your audience.
Aside from using this as a lead magnet to draw in new email leads, you could also use it to segment people who are already on your newsletter.
Alternatively, the form itself could be the point: maybe you just want to do some market research, and offering people a customized report of some kind is simply a way to entice them to take the survey and improve your response rate.
A few final notes here:
First, because this uses the GPT API rather than ChatGPT, you can set a limit to how long you want the responses to be, and it can be longer than what ChatGPT normally gives you. Longer responses will cost more– the prototype cost me about a nickel per use, including the monthly price of Zapier.
This also means you can’t rely on testing this in ChatGPT– you can start there, but before taking it live you need to do a lot of test runs using the whole system– GPT API, e-mail and everything– in order to see how it works.
You also want to consider how bad the hallucination issue hits you. Sometimes it’s very obvious to the user, like if it’s generating a local guide for tourists and it says the beach is on the east side of Los Angeles. Other times it’s not noticeable, like if you tell it to generate a free weight workout, and it includes some machine exercises but the user didn’t realize you asked for free weights.
Some of this comes down to the subject you choose– LLMs hallucinate more on some subjects than others. They tend to do well on creative tasks where there are few wrong answers, precisely because this masks the hallucination issue. Some of it also comes down to structuring the prompt and the Google form to both make instructions very clear, and make any failure to follow the prompt properly non-obvious to the user.
As a final note, I recommend putting a line in the prompt telling it to skip the preamble where it recaps the prompt at the beginning of its response, and just dive straight into the response. I don’t know if this goes for other LLMs, but GPT at least has this annoying habit of starting every response with something like “Sure, here’s an evidence-based weight training program for an intermediate-level male who wants to focus on his shoulders and sleeps 6 hours a night and blah blah blah….” Not only is this a bit tedious in general, but it would look weird to the user in this case since they haven’t seen the actual prompt that GPT is responding to.
Bottom line: test this a lot more than you think you should before you take it live. If you do though, this has a lot of potential as a way to drive mailing list engagement while delivering personalized information to your audience.