Why You Cannot Tell Which Outreach Actually Worked
Most businesses running several channels cannot say which one produced a customer. The tracking broke in 2021 and the answer moved into systems most companies never connected.
A business runs ads, sends cold email, posts on LinkedIn and gets referrals. A deal closes.
Ask which channel produced it and you get a guess. Usually the last thing the client mentioned, which is the least reliable evidence available and also the one everybody uses.
Why the platforms cannot tell you
They used to be able to, more or less. That changed with Apple’s App Tracking Transparency, released with iOS 14.5 on 26 April 2021, which required apps to ask permission before tracking a user across other companies apps and websites. Flurry, tracking the response daily from launch, found the overwhelming majority of US users declined.
So the platforms lost sight of much of what happens after a click and moved to modelling instead. The conversions in an ad dashboard are increasingly estimates. Useful for optimisation, unreliable as a ledger, and completely blind to anything that happened on LinkedIn or in an email thread.
Meanwhile the channels that never had tracking, outbound and referral, still do not.
The multi touch reality
Even with perfect tracking the question would be hard, because the premise is wrong.
A typical deal involves several contacts across weeks. They saw an ad, ignored it. Got an email, did not reply. Saw a post, recognised the name. Got a second email and replied to that one.
Which channel produced the customer? All of them, in sequence. Attributing it to the email is technically accurate and strategically misleading, because cutting the ads would reduce email replies and nothing in the reporting would explain why.
Where the answer actually lives
Not in any platform. In your own records, if they are built to hold it.
One record per person, carrying every touch from every channel, with its source attached, surviving all the way through to the closed deal. That is the only place the full sequence exists, and in most businesses it does not exist anywhere because each channel writes to its own system.
The practical failure is usually simpler than the theory. The same person appears as four records in four tools and nobody knows they are the same human until somebody calls them and looks foolish.
Where AI closes it
Two specific jobs, both previously impractical.
Recognising the same person across channels. A LinkedIn profile, a work email, a form fill with a personal address and a phone number are one contact. Matching those reliably needs judgement about partial and inconsistent information, which is exactly what rules based matching could never do well.
Reading what actually happened. The useful detail in a sales process sits in unstructured material: the email thread, the meeting notes, the DM exchange. AI reads it and writes the structured outcome to the record, so the history is complete without a salesperson doing data entry.
Once every touch lands on one record with its source intact, the question becomes answerable. Not perfectly, because multi touch reality does not permit perfect, but well enough to decide where next quarter’s budget goes.
What to measure instead of attribution
Ask on the form. How did you hear about us, in the prospect’s own words. Unfashionable, imperfect, and better than a model.
First touch and last touch, both recorded. Not one or the other. The gap between them is where your nurture lives.
Channels present in won deals, rather than credit split between them. If four in five closed deals had an ad impression somewhere in the sequence, that tells you what happens if you switch the ads off.
Cost per closed deal at the whole business level. Blunt, hard to argue with, and immune to attribution disputes.
Duplicate records are the practical blocker
Before any of the theory matters, most businesses fail at something more basic. The same person exists several times over.
A LinkedIn connection, a form fill from a personal address, a work email on a cold campaign and a phone enquiry taken by somebody who spelled the surname differently. Four records, one human, and every report built on top of that counts them as four.
The damage is not only in reporting. It shows up in the room, when a prospect who has been in an outbound sequence for six weeks is contacted by somebody who believes they are a fresh inbound lead.
Deduplication is unglamorous and it is the precondition for everything else here. Until one person is one record, attribution is arithmetic on the wrong numbers.
The question to ask on every form
The most useful attribution data available to a business of your size costs nothing and is deeply unfashionable.
Ask people how they heard about you, in a free text box, and read the answers. Not a dropdown, because a dropdown constrains them to the channels you already thought of and produces a tidy distribution that means nothing.
Free text is messy and it is honest. It surfaces referral sources you did not know existed, it tells you which specific piece of content is doing work, and it occasionally reveals that a channel you have been funding for a year is invisible to the people actually buying.
AI makes the messiness manageable. Reading and categorising a few hundred free text answers used to be the reason nobody did this. It is now a job that takes seconds and turns the best attribution signal you have into a chart.
What to do with the answer
The purpose of attribution is not to settle an internal debate. It is to decide where next quarter’s money goes, and that decision tolerates a fair amount of imprecision.
If four in five closed deals had an ad impression somewhere in the sequence, that tells you what happens if you switch the ads off, regardless of which touch a model wanted to credit. If half your revenue traces to referrals, the highest return action is probably to build a referral process rather than to optimise a campaign.
Directionally right and acted on beats precisely modelled and argued about, every time.
Sources
Apple, App Tracking Transparency framework. Introduced with iOS 14.5, released 26 April 2021.
Flurry Analytics, iOS 14.5 opt-in rate tracker, daily measurement from launch.
AI Optimize builds the single record, matches contacts across channels and keeps the source attached through to the closed deal. That work sits under Custom CRM and Reporting & Data.
Related reading
What Apple's Privacy Change Did to Your Ad Targeting
The single largest shift in paid acquisition this decade was not a platform update. It was a permission prompt, and most advertisers still run accounts built for the world before it.
Why Three Systems Give You Three Different Numbers
When nobody can agree what a closed deal is, every report becomes an argument. Fixing that is a definitions exercise, and it costs an afternoon rather than a licence fee.
WHAT WE BUILD