We Taught AI to Spot Buyers Before They Buy
- Arif Khan
- Jan 6
- 3 min read
Updated: Jan 15
--author Arif Khan | founder zinzu.io
How behavior sequences reveal hidden conversion intent
Most analytics tools answer one question really well:
“Who already converted?”
We wanted to answer a harder and more valuable question:“Who is most likely to convert next, based purely on behavior?”
So we built an AI algorithm that compares how converted users behave versus non-converted users, not using demographics or guesswork, but by studying actual sequences of actions over time.
Here’s what we did, in simple terms.
Data Source
This analysis uses Google’s GA4 obfuscated sample ecommerce dataset for the Google Merchandise Store.
The data is publicly available. Details about the dataset can be found at the link below.
120,000 real users (scaled from the original dataset)
Each user performed multiple actions: views, clicks, cart events, checkout steps
Only 170 users actually purchased
The remaining 119,830 users did not
That imbalance is important.
Conversion is rare.
Which makes prediction hard.
Step 1: Study the Buyers First
Instead of starting with non-buyers, we did the opposite.
We deeply studied the 170 converted users to understand:
What actions they performed
In what order those actions happened
How long they spent between actions
Which actions repeated
Which actions mattered and which were just noise
This created a behavioral fingerprint of conversion.
Not a funnel. Not a dashboard. A sequence.
Step 2: Clean the Noise
Real user data is messy.
So before comparing anyone, we cleaned the journeys:
Removed noisy events that don’t show intent
Removed duplicate actions
Eliminated loops where users repeat the same step endlessly
Focused only on meaningful actions
What remained was the true behavioral path of each user.
Step 3: Compare Sequences, Not Just Events
Most tools look at counts:
How many users viewed a page
How many added to cart
How many dropped off
We looked at sequences.
For every non-converted user, we asked:
How close is their action sequence to a converted user’s sequence?
Do they perform similar actions in a similar order?
Do they share a large subset of important events?
Do they spend a similar amount of time progressing?
If a converted user did:
A → B → C → D → E → FA non-converted user who did:
A → B → C → DScores higher than someone who did:
A → C → DEven if the order is not perfect, subset similarity matters.
Step 4: Rank Users by Conversion Likelihood
Once we compared:
Sequence similarity
Duration spent
Shared meaningful events
Behavioral structure
We ranked non-converted users by how closely they resemble converted users.
The result surprised us.
The Result: Hidden Converters
Out of nearly 120,000 non-converted users, the algorithm identified:
130 users with a very high likelihood of converting
That’s nearly a 76% increase over the number of users who actually converted, identified purely from behavior.
These users:
Behaved almost exactly like buyers
Reached deep checkout steps
Spent comparable time
Showed strong intent
They did everything right.
They just didn’t complete the final purchase.
Most likely reasons:
Payment failure
Shipping cost shock
Distraction
Technical issues
These are missed opportunities, not uninterested users.
Why This Matters
Traditional analytics treats all non-converters the same.
Our approach proves they’re not.
Some users are noise.
Some are browsers.
And a small, valuable group is already “mentally converted.”
They just need the right nudge.
What Zinzu Does Differently
Zinzu prioritizes:
Behavioral sequences, not isolated events
Duration, not just clicks
Subset similarity, not rigid funnels
Intent patterns, not averages
Every user journey is treated as a story over time.
And when you compare stories, patterns emerge.
Final Thought
Conversion doesn’t happen at the purchase button.It happens much earlier, in behavior.If you can identify users who already behave like buyers,you stop guessing and start acting with precision.
That’s what we built.
And this is just the beginning.
Comments