AI Patent in Real-Time Lead Scoring
In 2018 I built the proof of concept for a patented system that predicts, in real time, how likely each anonymous visitor is to convert, and changes the page's content for that visitor based on how they behave. I'm one of three inventors on the patent.
One website, shown the same way to everyone
Think about the last marketing website you visited. The person before you, and the person after you, saw the exact same page: the same hero image, the same features in the same order, the same call to action. The site had no idea who any of you were or what you wanted. In 2018, that was simply how the web worked, and mostly it still is.
The team I was part of asked a different question. Ad networks were good at guessing which ad to show a person, but nobody was changing how the page itself was built for each visitor, and nobody was tracking how close a given visitor was to converting. We wanted a site that could estimate that from behavior alone and act on it while the person was still browsing.
We built it. The method is covered by US patent 10,951,685 B1, "Adaptive content deployment," with me as one of three inventors. My part was the proof of concept: the system behind the demo, running on the production website of one of the largest apartment operators in the US.
Score the visitor, then reshape the page
The system does two things at once, continuously, for every anonymous visitor.
First, it keeps a running likelihood to convert. Every action a visitor takes, a page view, a click, how long they linger, the path they take, is fed to a model that raises or lowers a single score: the probability that this person is about to act. The score moves in real time as they browse.
Second, it adapts the page to what that behavior reveals. Not just which ad to drop into a fixed slot, but the content and layout of the page itself: which images, which features surface first, what the icons mean, what the map highlights. And when the score crosses a threshold, the site stops waiting and reaches out directly.
The adaptive loop
How a behavior becomes a number
The score is a weighted sum. The system turns raw browsing into structured signals (did this visitor view pricing, look at a specific unit, open the contact page, come back for a second session), and a model assigns each signal a weight based on how strongly it predicted conversion in the past. Add up the weighted signals and you get a probability.
Illustrative weights the model assigns to actions
Because the weights come from real conversions and the model is retrained on a schedule, the score reflects what actually predicts a lease rather than a guess. As a visitor moves through the site, the number climbs step by step, with larger jumps on higher-intent actions.
One visitor's likelihood, action by action
The same page, rebuilt for who you are
The score decides when to act; the adaptive content decides what to show. The proof of concept followed three visitors through the same apartment website, and the site changed for each one.
A default visitor with no strong signals saw the standard page. A visitor who arrived from a pet-friendly ad saw the page lead with pet imagery, pet amenities pulled to the top, and the neighborhood map highlighting dog parks. A visitor whose early behavior signaled a parent saw images of families, schools surfaced first, and the map highlighting nearby schools. Even the small save icon changed: a plain heart for the default visitor, a paw print for the pet visitor, a graduation cap for the parent.
- Save icon
- Heart
- Top amenity
- Standard order
- Map highlights
- Nothing special
- Imagery
- Generic
- Save icon
- Paw print
- Top amenity
- Pet perks
- Map highlights
- Dog parks
- Imagery
- Pet-friendly
- Save icon
- Graduation cap
- Top amenity
- Top schools
- Map highlights
- Nearby schools
- Imagery
- Families, play areas
Because the pet visitor and the parent visitor saw the content that mattered to them, their likelihood to convert climbed higher, in fewer actions, than the default visitor's did. When each crossed the threshold, the chat was tailored too: a pet-rent offer for one, a note about nearby schools and an invitation to tour for the other.
From a demo to a live system
The proof of concept was convincing enough that it grew into a real system running on the company's production website, scoring actual visitors as they browsed. As a visitor moved deeper into the site, the system knew more about them and the likelihood climbed, from an anonymous machine on the home page to a well-understood prospect by the time they favorited a specific unit.
Likelihood rising as the journey deepens
This was 2018. There were no on-demand GPUs and no off-the-shelf personalization platforms. It ran on ordinary CPU servers, with the data pipeline, the model, and the real-time scoring built by hand. What is now sold as website personalization and predictive lead scoring was, then, a patent application and a working demo.
Why it held up
The core idea is straightforward: read a visitor's behavior, estimate where they are headed, and change the page and the timing of outreach to match. Behavioral scoring plus content that adapts per visitor is common now; most personalization and predictive-lead-scoring tools do some version of it.
In 2018 it was early enough to be worth a patent. Most of the work that made it accurate was in the data and the scoring, not the model.