Last updated · By Madalina Turlea5 min readCase Study · Prompting
One of the use cases we talk about a lot is using AI to personalize content infinitely. A good example is a lead magnet: a quiz on a website where someone fills in where they are in their journey of integrating AI into their product, anywhere from greenfield to very advanced, and then receives a report by email with insights and next steps that are personalized for them.
What AI changes about lead magnets
You probably know lead magnets from before LLMs. You filled in a form, and you got back something that was an aesthetically compiled email, assembled from whatever options you chose. It was static. Everyone who picked the same boxes got the same thing.
AI gives you the possibility to generate content that is infinitely personalized. Instead of a static, generic report, the output can be tailored to the team, their needs, and their current state, so the person actually finds themselves in it and it feels personal rather than assembled.
The challenge: personalization versus slop
When you personalize with AI, you are working on a spectrum. On one end is deep personalization, where there are always details specific to the person's situation. On the other end is delivering genuinely valuable insights, rather than sliding into generic AI slop.
We all know how the slop feels. It tries to make you feel seen, but it does not deliver any real insight. It is content generated for the sake of generating something, sometimes fabricated, that is not actually valuable or real. For a feature like this one, that is one of the biggest challenges: personalizing so the insights feel specific to the team, without producing generic AI answers or made-up content.
There are two failure modes to watch for. One is the model pulling in insights that do not belong or are not relevant, or making things up to fill the personalisation. The other is overfitting, taking the insights you do have and forcing them onto the specific use case the user filled in, whether they fit or not.
How to keep it real
The way to avoid the slop is to feed the model the real data you have, not the generic training data the model already carries. For this feature, that meant describing the actual benchmarks and insights built up over months of discussions with product teams and thousands of experiments, the real unlocks teams miss, and the platform features that help at each level. The model maps the person's quiz answers to that real data, rather than inventing plausible-sounding advice.
A useful test for a good answer: it surprises you with how accurate or insightful it is, or it says the thing you would have said to that person directly in a call. If the output reads like something anyone could have gotten from a generic chat tool, it has slipped into slop.
One more thing to design for, specific to a lead magnet: the report should not just inform, it should also make the person curious to try more, and it should give you enough data to understand where they are so you can actually help. That duality, real value plus a reason to take the next step, is what separates a personalized report that works from one that just looks personalized.
Enjoyed reading? Add us as a preferred source in Google to support us.
Define what good looks like for your AI feature, measure it on real data, and improve it with your domain experts in the loop, no code required. Teams have gone from 52% to 88% accuracy in under three weeks.
