A customer sees a reel in March, searches your name in May, clicks a paid ad in June because it appeared above the organic result, and enquires.
Which channel produced that sale?
Every plausible answer is defensible and none is complete. The paid platform will claim it. Analytics will credit the search. The reel, which did the actual work of making the name worth searching, gets nothing. This is the attribution problem, and most businesses either ignore it entirely or trust a dashboard that quietly resolves it in one channel's favour.
Why the platforms disagree
If you add up the conversions each advertising platform reports, the total will exceed the number of sales you made. This is not fraud. It is that each platform is answering a narrower question than you asked.
A platform reports conversions where it saw an interaction within its attribution window. Two platforms that both touched the same buyer will both count it. Neither is aware of the other, and neither is aware of the channels that are not platforms at all — a recommendation, a hoarding, a conversation.
The practical consequence: platform-reported numbers are useful for comparing a channel to itself over time, and misleading when summed or compared across channels.
The models, and what each one distorts
Attribution models are rules for assigning credit. Each has a bias worth knowing.
Last click. All credit to the final interaction. Simple, and it systematically over-credits whatever sits closest to the purchase — branded search, retargeting — while under-crediting everything that created the demand.
First click. All credit to the first interaction. The mirror image: it over-credits discovery and ignores what closed.
Linear. Credit split evenly. Fair-sounding, and it treats a passing impression as equal to the conversation that decided it.
Time decay. More credit nearer the sale. Reasonable for short cycles, poor for long ones, where the early work matters most.
There is no correct model. What is useful is knowing which one your reporting uses, because that choice determines which channel appears to be working — and budget follows appearance.
The honest answer: ask
For most businesses below enterprise scale, the highest-quality attribution data is not technical. It is the question "how did you hear about us", asked by a person and written down.
It is imperfect. People misremember, they name the last thing rather than the first, and some do not answer. It is also the only method that can see the channels no analytics platform can — the referral, the sign, the colleague who mentioned you.
The requirement is that the answer is recorded somewhere permanent rather than remembered. A field on the enquiry record, filled in every time, produces something after ninety days that no dashboard can: a picture that includes offline reality. That is one of the practical arguments for keeping a proper customer record, covered in do you need a CRM, or just a better customer record.
What is genuinely unmeasurable
Some influence leaves no trace, and pretending otherwise causes worse decisions than admitting it.
Someone screenshots your post and sends it to a colleague. Someone sees your work at a client's premises. Someone hears your name three times over a year and searches it directly. Direct traffic is not a channel — it is the visible residue of everything that happened where you could not observe it.
The correct handling is not to build increasingly elaborate tracking. It is to accept a category of activity that is real, unattributable, and worth doing anyway — while making sure the channels that are measurable are measured properly.
The test that beats every model
When the numbers are ambiguous, stop attributing and start testing.
Turn a channel off for a defined period, hold everything else steady, and watch total enquiries — not that channel's reported conversions. If the total drops, the channel was contributing. If nothing changes, it was claiming credit for demand that existed anyway.
This is the only method that measures incremental effect rather than correlation, and it repeatedly overturns dashboard conclusions — most often for retargeting and branded search, both of which excel at being present at the moment of a decision they did not cause.
It requires patience: a fortnight is noise, and long sales cycles need longer. But one clean test is worth more than a year of model-switching.
What to build, in order
A single record of every enquiry, with source recorded at the point of entry. Nothing else works without this.
A known value per enquiry, so channels can be compared on cost rather than volume. The arithmetic is in what to fix before you spend on ads.
One agreed model, applied consistently, with everyone aware of its bias.
A stated tolerance for the unknown. Deciding in advance that some proportion is unattributable prevents the reflex of cutting the channels that are hardest to measure — which are frequently the ones building the demand everything else converts.
The wider set of numbers this sits inside is in which marketing numbers are worth tracking.
Frequently asked questions
Which attribution model should we use?
For a short cycle, last click is workable if you know its bias. For anything longer, combine a simple model with asked-source data and periodic hold-out tests. Consistency matters more than the choice.
Why do platform numbers exceed our actual sales?
Because each platform counts interactions it saw, independently, and buyers touch more than one. Compare each platform to its own history rather than summing them.
Is "how did you hear about us" reliable?
Individually no, in aggregate yes. Over a few hundred enquiries the pattern is informative, and it includes offline sources nothing else can see.
Should we invest in attribution software?
Only when spend is large enough that a few percent of misallocation exceeds the cost. Below that, a source field filled in consistently gives most of the value.
How do we justify channels we cannot attribute?
With hold-out tests rather than argument. Turn it off, watch the total, turn it back on. That evidence persuades where model debates do not.
Where to start
Add one mandatory field to your enquiry record: how did they hear about us. Fill it in for ninety days without changing anything else.
The resulting picture will differ from your dashboard — usually by revealing that something unglamorous and unmeasured is doing more work than the channel receiving the budget.
If you want distribution measured rather than assumed, get in touch.