What marketing attribution cannot tell you
Attribution can tell you which recorded touchpoints preceded a purchase. It cannot tell you which of them caused it, and it cannot see the conversation, the forwarded link, the podcast or the search someone did on a phone that was never connected to their work laptop. Those invisible paths are not an edge case, and a model that assigns full credit to the last click is mostly measuring which channel happened to be closest to a decision already made.
What is attribution actually measuring?
Correlation between recorded events and an outcome, weighted by a rule someone chose. That is the whole mechanism. First touch, last touch, linear and time decay are all the same data with different arithmetic applied, and switching between them changes which channel looks successful without changing anything about what happened.
This matters because the model is usually chosen for organisational reasons rather than analytical ones. Whichever team is being evaluated tends to prefer the model that credits its part of the journey, and both can defend their choice, because there is no correct answer inside the data itself.
The honest description is that attribution ranks channels by how often they appear near conversions. That is genuinely useful for spotting a channel doing nothing at all. It is not a measurement of causation, and treating it as one leads to cutting the channels that create demand in favour of the ones that harvest it.
Which paths are invisible to it?
More than most dashboards imply. A recommendation from a colleague leaves no trace. A link shared in a private message arrives with no campaign parameters. Someone who hears you mentioned on a podcast and searches your name later is recorded as branded search, which credits the search engine for work the podcast did.
Device and browser boundaries remove more. Research on a personal phone followed by a purchase on a work machine breaks the identity link entirely unless the person logs in on both. Browsers now limit how long client-side cookies survive, so a journey longer than a few weeks may lose its origin even on the same device.
Consent choices remove the rest. Where a visitor declines tracking, the touchpoints simply do not exist in the data, and the conversions that follow appear as direct. Direct traffic is not a channel, it is the folder where everything unmeasurable is filed.
| Model | What it credits | Systematically favours | Reasonable use |
|---|---|---|---|
| First touch | The earliest recorded touchpoint | Awareness channels and anything that ranks for a first search | Understanding what starts journeys, on short cycles |
| Last touch | The final touchpoint before conversion | Branded search, retargeting, email | Fast optimisation of harvesting channels only |
| Linear | Every recorded touchpoint equally | Whichever channel produces the most touchpoints | A rough map of what is involved, not a budget rule |
| Time decay | Recent touchpoints more heavily | Late-stage channels, less severely than last touch | Long cycles where late activity genuinely matters |
| Position based | First and last most, middle shared | Neither extreme, by construction | A defensible compromise when a rule is required |
| Self-reported on the form | Whatever the buyer says | Memorable channels over frequent ones | The only method that sees offline and word of mouth |
Why does the same conversion get counted twice?
Because each platform reports on its own view using its own rules and its own window. An advertising platform counts a conversion if its click or impression appeared within its lookback period, your analytics tool applies its own model, and your marketing automation platform credits whichever campaign is stamped on the contact record. Sum them and the total exceeds the number of deals you actually closed.
The lookback window is the biggest single cause and the least discussed. A platform crediting itself for conversions up to a month after an impression will claim a great deal of activity it merely coincided with.
The practical rule is to pick one system as the arbiter for reporting, usually the CRM because it holds the revenue, and treat every other platform's conversion count as an operational signal for optimising that platform rather than as a business number.
Is self-reported attribution better?
It is worse at precision and much better at coverage, which usually makes it the more useful of the two. A single open text field asking how someone heard about you captures word of mouth, offline conversations, podcasts and events that no tracking can see, and those are frequently the channels that matter most for considered purchases.
It has real biases. People name what they remember rather than what influenced them, they under-report advertising, and they over-report the most recent prompt. Nobody should treat the percentages as precise.
Used together the two methods answer different questions. Tracked attribution tells you what happened on your website. Self-reported tells you what happened in the person's life. Where they disagree strongly, the disagreement itself is the finding, and it usually points at an unmeasured channel doing real work.
What should you use instead of asking who gets credit?
Ask what changes when you stop. Holdout tests, geographic splits and simply pausing a channel for a defined period produce causal evidence that no attribution model can, because they compare an outcome against a world where the spend did not happen.
This is uncomfortable because it requires deliberately turning something off, and the results take as long as your sales cycle to appear. It is also the only method that answers the question everyone actually has, which is whether the money is doing anything.
For channels too small to test this way, accept that they are unmeasurable individually and judge them as a portfolio. Trying to attribute a small channel precisely produces confident numbers with no information in them.
What should you check in your own setup?
Three things this week. First, what proportion of your closed deals have no recorded first touch at all. If it is substantial, your model is describing a minority of your revenue. Second, whether campaign parameters on your contact records are overwritten on each visit or preserved from the first, because the two produce completely different reports from the same traffic.
Third, add or read a self-reported source field on your main enquiry form and compare a quarter of answers against what your analytics claims. The channels named by buyers but absent from the dashboard are the ones your budget is currently underfunding.
None of this requires a tool. It requires accepting that the number in the dashboard is a description of the measurable subset, and deciding what to do about the rest rather than pretending it is not there.
Common questions
- What can marketing attribution not measure?
- Anything without a recorded touchpoint: personal recommendations, links shared in private messages, podcasts, offline conversations and events. It also loses journeys that cross devices without a login, journeys longer than the browser's cookie lifetime, and every visitor who declined tracking. Those conversions usually appear as direct traffic, which is less a channel than the folder where unmeasurable activity is filed.
- Which attribution model is most accurate?
- None of them, because they are the same recorded data with different arithmetic applied. First touch favours channels that start journeys, last touch favours branded search and retargeting, and both are defensible. Since no model is correct within the data itself, the choice usually reflects which team is being evaluated. Attribution ranks channels by proximity to conversions rather than measuring cause.
- Why do platform conversion numbers add up to more than actual sales?
- Each platform reports on its own view with its own attribution rules and its own lookback window, so several will claim the same conversion. A platform that credits itself for conversions weeks after an impression will claim activity it merely coincided with. Nominate one system, usually the CRM because it holds the revenue, as the arbiter, and treat other platforms' counts as tuning signals for those platforms.
- Is self-reported attribution reliable?
- It is imprecise and much better at coverage than tracking, which usually makes it more useful. A single open text field on your enquiry form captures word of mouth, offline conversations and events that no tracking can see. People name what they remember rather than what influenced them, so treat the answers as directional. Where self-reported answers and analytics disagree strongly, that gap is the finding.
- How do you prove a marketing channel actually works?
- By changing it and observing what happens: a holdout group, a geographic split, or pausing the channel for a defined period. These produce causal evidence because they compare against a world where the spend did not occur, which no attribution model can do. The cost is discomfort and time, since results take at least one sales cycle to appear. Channels too small to test should be judged as a portfolio.
