We knew customers were coming in, we just didn't know where from

Where the customer's source kept getting lost

It wasn't one cause, it was three things happening at once

First: the campaign tag was getting lost on the way

Every ad carries a small tag that says where the visit came from. Its technical name is UTM, and it is supposed to travel with the visitor and settle in their record inside the CRM

Because of a database issue and a broken link between the website and the CRM, this tag never arrived, so the customer entered the system with no source

And once the source is lost at that moment it never comes back, because a month later nobody remembers where each name came from

Second: a lot gets lost when things are logged by hand

When the source doesn't arrive automatically, everything depends on manual entry, and manual entry on a busy day leaves gaps
A call happens and doesn't get logged, a follow up happens and the system knows nothing about it

And that isn't anyone's fault
Any system that asks people to log every step by hand on a busy day ends up the same way

Third: the person who tries the product isn't the person who pays

In B2B sales, the person who signs up for the trial is usually an employee or an operations manager, and when the company moves to a paid plan, the payment comes from the finance team's email

The system sees two different emails and treats them as two unrelated people, even though it's the same company
The only shared part is the company domain

What we had and what we didn't

Fixing the issue at the root needed time from the product team and a spot on a long priority list, and waiting meant more months of spending decisions based on guesses

And there was no budget for a ready made tool to connect the systems

So the question became: what can we do with what we have right now, in days, not months

The idea, simply put

I started from the end instead of the beginning

Instead of tracking every visitor and waiting until they buy, I took the list of people who had actually paid, then went backwards looking for a trace of each one in every available source: their mobile number, their email, their company domain, their company name

And instead of a yes or no answer, I gave it a score
This is conclusive evidence, this is strong evidence, this is possible evidence, and every match comes out with its confidence score

The technical name for this work is Attribution, which simply means knowing where a customer came from and what evidence proves it

The rest of this page is how it worked, and what we found

Where we looked for customers

Five places, and each one knows the customer in a different way
The ad platforms keep what the person typed into the signup form, the email base keeps company records with a business name and work email, and the subscription records know which company paid

How we know it's the same customer

There is no single ID shared across all these systems, so I ranked the evidence from strongest to weakest
We look for the strongest proof first, if it isn't there we move to the next one, and with every match we write down how confident we are

Why cleaning was half the work

The same number can be written five different ways, so nothing is compared as is
Numbers are stripped of spaces, symbols and country codes and cut down to the last nine digits so different formats of the same number match, emails are made lowercase with spaces removed, and company names are split into words with common generic words removed, like company, establishment and trading, because keeping them makes any two names look alike

And why confidence scores are different

A mobile number and an email are personal identifiers, and a match on either one means the same person, so I gave them 100%
A company email domain proves the company, not the person, and since a domain belongs to one company I gave it 95%, while excluding public domains such as gmail and yahoo because a domain used by millions of people proves nothing

Company name is the weakest evidence because similar doesn't mean identical, so I required three or more words in common and gave it 85%, and that number is my way of saying clearly that this kind of evidence can be wrong more often

We found where 47.7% of customers came from, with proof

Out of all paying customers, we traced 47.7% with evidence, and for 52.3% we didn't find enough

And for the customers we did trace, most rest on conclusive evidence, not on inference

And why we didn't make up an answer for the rest

52.3% of records came out with no source, and it's easy to read that number as failure, but it means much less than that

All it means is that we didn't find enough evidence to link this customer to a source
It doesn't mean they didn't come from marketing

The clearest possibility is that the customer signed up on one of the channels with a personal email and personal number, then signed up for the product with a work email and office number, so both records belong to the same person and nothing in the data links them
That is possible and logical, but it stays a possibility, and turning it into attribution means making up a result

So we decided to leave it as it is, with no source
We don't split it across channels and we don't credit it to whichever channel was spending that month, because that would make the numbers look better than they are

And this gap shrinks as data collection at the point of entry gets better, not as the guessing gets better

How many places a customer passed through before buying

The first thing that showed up after connecting the data is that one customer can exist in more than one place, and that never showed up in any earlier report, because every report picked one source per customer

Three things people mix up

This distinction is the most important thing I took from this work, and it's very simple once it's written down

First, that the customer showed up in a source, which only describes what's in the data: their name, number or email exists in that channel's records

Second, that we have evidence linking them to that source, and this is where match type and confidence come in
A mobile number match is one kind of evidence, a similar company name is weaker evidence, and the difference is written next to every customer instead of getting lost in one word

Third, that the source caused the subscription, and we don't prove or claim that
A customer showing up in an ad campaign and in email doesn't tell us which channel drove the decision or which came first

What we did answers the first and second, and the third needs the dates of every event in order, which wasn't available

What changed in the daily work

The amount of data didn't change and neither did the number of channels
What changed is that everyone now opens the same table and sees the same answer

• The budget now goes to channels where paying customers actually show up, not by guesswork
• Every channel is now measured by customers who paid, not by its own dashboard, and that made a big difference for email because its impact doesn't show up as a click
• Email can finally be judged, because we can see whether the people we email actually subscribe
• The customer journey got clearer, as 15% of customers went through more than one channel and used to be counted under just one
• The team now knows what it has evidence for and what it doesn't, instead of both getting mixed into one number

The limits of this work

So the numbers aren't read as bigger than they are, here are the honest limits of what we did

If this were my company, this is what I would ask to improve

The engine solved the urgent problem in days with no budget, and that was the goal
But anything built fast leaves things worth improving, and this is what I see when I go back to the code

And the most important fix stays completely outside the engine: getting the campaign tag into the customer record from the first moment, and logging the date of every touchpoint
Then talking about first touch and the order of the journey becomes possible, and fewer customers end up with no source in the first place

Tools and code

All the work happened inside Jupyter Notebook, with one run that reads the sources, does the matching and writes the final table

Processing with pandas for reading, cleaning and matching, NumPy for numeric work, plus the standard Python libraries glob, os and time for file handling and timing the run
The email base is read as Parquet because it's too big for a regular spreadsheet format, customer records are read from an Excel export, and the platform tables are pulled straight from their source

I used ChatGPT and Claude while writing the code, to speed up the writing and review edge cases
Defining the problem, designing the evidence ladder and confidence scores, and the exclusion rules were decisions I built from understanding the data itself

The repository only holds the engine logic, with no customer data, source files or access keys

Tags for this case study

The solution didn't come from a new tool or a budget, it came from defining the problem precisely and then building the smallest thing that answers it