A multi-touch attribution model credits only the touches it can identify. Four kinds of touchpoint never reach it, and the channel that loses their credit is the one that gets cut.
A multi-touch attribution model is easy to name and hard to audit. Most of the pages that explain one are published by a company selling it, so the diagram you are reading was drawn by the vendor whose model sits on it. That is not a scandal, and it is a reason to know exactly where the diagram ends: four kinds of touchpoint never appear on it at all.
Here is what that costs. A channel report prints a percentage for every channel, and no line anywhere for what it excludes. That omission does not announce itself. A channel that earned nothing this quarter and a channel whose work was never counted land on the same line of the same report with the same value, and the arithmetic behind both is correct.
So the weakest-looking line is often the one whose work happened somewhere the model does not reach. The touches leave through four specific doors, and each one has a name, a mechanism and a cost. Once you know which of them touch your worst channel, the budget conversation stops being about the dashboard.
What Multi-Touch Attribution Actually Assigns
Multi-touch attribution splits credit for one conversion across every touch it can observe, so a journey that runs through a search ad, a webinar, a retargeting impression and a demo request gets four fractional claims instead of one winner-takes-all click. The fractions sum to one, and the report prints a channel total that reads like an answer.
Underneath that arithmetic sits one assumption worth saying out loud: every contributing touch was observed, tied to one identifier, and joined to the conversion inside a window that had not yet expired. The model does not test the assumption. It assumes it and divides.
If you are choosing a product instead of questioning an output, our best attribution software roundup covers that decision. What follows holds whichever tool you install.
The Four Touchpoints Your Model Cannot See
Each mechanism below shrinks the set of touches the model can identify, and none of them raises an error. Each returns a clean number with a piece of the journey folded into somebody else's column.
Consent loss and the shrinking identifiable denominator
Start with the denominator, because that is the part people get wrong. When a visitor declines tracking, or is never asked in a way the law recognizes, the session still happens and the click still happens. What disappears is the identifier that would have tied that session to a conversion three weeks later.
The model does not respond by printing a lower total with a hole in it. It prints the same conversion, attached to the touches that carried an identifier. Credit is redistributed among the survivors, not deleted.
The published acceptance numbers are more fragile than they look. Didomi's 2026 State of Data Privacy Benchmark, published in May 2026, reports a 94.2 percent US cookie acceptance rate with 5.8 percent opting out. Read the method and the figure moves: most state privacy laws treat inaction as implied consent, so it combines active opt-ins with visits where nobody chose anything. A 94.2 percent acceptance rate and a 94.2 percent identifiable share are different claims, and the benchmark measured the first.
That is arithmetic, not an observation, and it runs like this: non-consenting conversions do not surface as unattributed. They surface attributed to the touches that could be seen, which skews toward whatever carries the strongest identifier, usually the logged-in and the lower-funnel.
Cross-device and the household problem
Relay42's Match Rates Guide, reported through Landbase in a vendor guide published in 2026, counts 6.58 connected devices and roughly 12 identifiers per person, and concludes that perfect matching is mathematically impossible. Treat the source as what it is, a vendor document, and notice that the direction survives without the precision.
You will go looking for a verified cross-device match rate. I did, and I did not find one. The sources that discuss the problem describe the mechanism and stop. The figure a vendor quotes you is most likely their own benchmark on their own traffic, and it should not be carrying a budget line.
The household version is worse because it is not a device gap. Forrester's 2026 research, reported in B2B attribution guidance, puts a typical buying decision at 13 internal stakeholders and nine external influencers, each running a separate journey. Your model tracks individuals. The decision belongs to a group, on devices you have never seen.
A conversion that took three people and five devices lands as one row, attached to whichever identifier arrived last on the device that happened to convert. The other four devices were not underweighted. They were absent, and their share moved to the channel that owned the final session.
Walled gardens
This one is structural. That is why it has stopped registering as a problem at all.
Comscore's 2026 State of Programmatic Report, released in January 2026, found that 87 percent of media buyers rate cross-channel performance metrics inside programmatic platforms as critical or valuable for decision-making. IAB's 2026 Outlook Study puts advertiser focus on cross-platform measurement at 72 percent in 2026, up from 64 percent a year earlier. Both sets of buyers want the one number that would let them compare platforms, and both are asking the platform itself to produce it. The platform is therefore a participant in the auction and the scorekeeper of its own result.
Each platform counts conversions inside its own window, against its own identity graph, under rules that it writes. None of them shares an identity layer with the others that would recognize the same sale when it appears in three of them.
Which means, arithmetically and not observationally, that the sum across platforms can exceed the real total. Nobody is inflating anything. Each platform correctly claims one sale under its own definition of claiming, and three incompatible definitions produce three claims for one sale. That is why a vendor's own figure cannot be audited from outside. Your report does not print an overcount. It prints four totals with a footnote, and then somebody adds them up on a slide.
Offline and long-lag conversion
The quietest of the four. It looks like a clean attribution instead of a missing one.
Optifai's 2026 research, reported in B2B SaaS attribution guidance, puts sales cycles 22 percent longer than in 2022, with a median of 84 days and enterprise deals above 100,000 dollars in annual contract value often running 180 days or more. Over six months the touchpoints multiply, cookie windows expire, people switch devices, and the signal degrades.
Here is the sting. A conversion that closes after the window shuts is not filed as unattributed, which would at least be honest and would show up in a gap report. It is filed as direct or organic, because no paid touch survived to claim it. A team that spent five months building demand with events, content and an expensive sales motion gets a report saying the deal arrived on its own.
I have watched a channel get defunded on exactly that reading. The work was real. The report was clean. The two had stopped describing each other.
The blind spot | What the model loses | What the report shows instead | What it costs you to miss it |
|---|---|---|---|
Consent loss | The identifier that would join a session to a later conversion | The same conversion, credited to the identifiable touches, usually lower-funnel | The top-of-funnel work that created the demand looks like it did less than it did |
Cross-device and household | Every touch on the other devices, and every stakeholder but the one who converted | One row, owned by the last identifiable session | Channels that start journeys get systematically underweighted against channels that end them |
Walled gardens | Any shared identity layer that would deduplicate across platforms | Platform totals that can legitimately sum to more than the real total | You cannot audit a vendor's own figure, and the sum overstates what you bought |
Offline and long lag | The touches that fell outside the window before the deal closed | A conversion filed as direct or organic | Long, expensive, high-value cycles get read as self-sourced and become the first cut |
What Fractional Credit Does to a Budget Decision
Let me turn the arithmetic into money, with round numbers. This is an illustrative example, not a benchmark from your account or anyone else's.
Suppose you spend 100,000 dollars a quarter across four channels: 40,000 on paid search, 30,000 on paid social, 15,000 on webinars and content, and 15,000 on events. Your attribution tool reports revenue, and you have been asked to find 20,000 dollars for next quarter.
Paid search: 400,000 dollars attributed. Paid social: 250,000. Webinars and content: 150,000. Events: 80,000.
Rank those four numbers and events is last by a wide margin. On the page there is no ambiguity about it. The obvious move is to take the 20,000 from events.
Overlay two blind spots on that ranking. The event program is where the enterprise relationships start, and deals above 100,000 dollars in annual contract value run longer than 180 days, which is past the end of the window. When those conversions close, they arrive as direct or organic. Meanwhile the event conversations happen across several people and several devices, and one of the attendees never accepted a tracking prompt, so the sessions that did carry an identifier went to whichever channel they touched closer to the purchase.
Events did produce the revenue. It was counted, just not in the events column.
Run the same overlay on paid search. It sits closest to the conversion, so it inherits everything the other channels lost. Its 400,000 includes conversions that began in a webinar and closed through a branded search the buyer ran afterwards. That does not make search a bad channel. It means search's number and events' number measure different quantities and are being ranked as though they measured the same one.
What comes out of this is an allocation, which is what you were actually being asked for: events keeps its budget and gains a stated caveat, search's figure is discounted by a named amount for what it inherited, and the 20,000 comes from whichever channel has neither a blind spot nor a demonstrated return. That is a different decision from the one the raw ranking produced, and it costs nothing to make once you have said out loud which touches each number cannot see.
One boundary matters here, because the neighboring arguments blur together. Our article on marketing measurement makes the case that channels the stack cannot see go unreported. This one shows which touches those are, where in the journey they drop out, and what the redistribution does to one channel's line. That piece works in numbers. This one works in touchpoints.
Reconciling against the CRM is a separate job, and it belongs to revenue attribution, which starts from the CRM record instead of the touchpoint. If you are assembling the deck that carries these numbers, the tooling question sits in marketing analytics tools, not here.
Fix the Data or Change the Method
Three honest responses exist, and they cost different things. None of them makes the blind spots disappear.
Fix the identity layer. Consent management that records a real choice, a customer data platform that stitches identifiers, server-side tagging, and a deliberate decision about what happens to unauthenticated traffic. This narrows the denominator problem and does nothing for the walled gardens or the long lag. It is the right first move when your identifiable share is small enough that the model is dividing a thin number into a large one.
Run an incrementality test. A holdout answers a different question, what would have happened without the spend, and incrementality testing covers the designs. Spend it on the one channel whose value you cannot settle by argument.
Change the method. Marketing mix modeling works at the aggregate level, not the user level, so it sidesteps identity entirely and brings its own limits in exchange. It is a legitimate option, and a different discipline.
Some teams will work through that list, fix what they can, and still hold a number they cannot defend in full. For them the honest move is to stop presenting the number as complete: attach the four blind spots to it, mark the channels they touch, and let the decision happen with the caveat visible. That is a worse-looking report and a better decision, and I would rather defend the first than repeat the second.
Conclusion
Go back to that channel report. The channel at the bottom is still at the bottom, and now you know what the ranking cannot see.
You will not get a clean answer on how much of the journey is invisible, and I am not going to invent one. Measuring what a model cannot see requires something other than the model. What you can do is name the four mechanisms, say which of them touch your worst-performing channel, and stop letting an incomplete number decide where the money goes.
A model that cannot see a touch does not report a gap. It reports somebody else's win.
Frequently asked questions
What is multi-touch attribution in marketing?
It is a way of splitting credit for one conversion across every marketing touch a system can observe before it, instead of assigning the whole conversion to the first or last click. Which tool does the splitting matters less than whether the touches reach it at all, so if you are past the diagnosis and into the product choice, our guide to [marketing attribution software](/blog/marketing-attribution-software/) takes that question from here.
Which attribution model should a small team use?
The simplest one your data can actually feed, which is rarely the one with the most parameters. A weighted model assumes a journey your identifier probably did not follow end to end, so on thin data it adds assumptions where you wanted insight. A last-click or linear model you can explain, shipped with a written list of what it misses, beats a sophisticated model nobody in the room can audit.
Is MTA the same as attribution modeling?
No. Attribution modeling is the general practice of assigning credit, and it includes aggregate, top-down approaches alongside user-level ones. Multi-touch attribution is the user-level branch, which is why it depends on identifiers and why the four blind spots hit it specifically.
What percentage of touchpoints can MTA actually see?
Nobody publishes that number. The explanation is structural, and it follows from what the model is: measuring the invisible share would require observing the touches the model cannot observe, so the figure has to come from something else, such as a holdout, a matched control, or a survey. What exists instead is a set of narrow figures for individual mechanisms, such as consent rates under a specific default or device counts from a vendor guide, and none of them composes into a total. If a vendor quotes you a coverage percentage for your account, ask what the denominator is.
See What Your Measurement Stack Cannot See
Allable connects your analytics, search and ad accounts and works through the same four questions on your own data: which touches are missing, which channels inherit them, and what that does to the line you are defending.
