What’s wrong with multi-touch attribution in B2B, and what to do anyway

Multi-touch attribution struggles in B2B: long buying windows, large buying groups, few deals and channels that are hard to measure. Here is what to do instead.

By Mehul Nagrani ·· Updated

What is multi-touch attribution?

Multi-touch attribution (MTA) is a method that splits credit for a deal across several of the touches a buyer had before it closed. Common models are linear, time decay, U-shaped, W-shaped and algorithmic. First-touch and last-touch give all the credit to one touch, so they are single-touch models, not MTA. None of them test what would have happened without a touch.

That gap is easy to live with when you can test the model against an experiment. In B2B you usually can’t, which is why the problems below matter.

Why does multi-touch attribution fail in B2B?

Multi-touch attribution fails in B2B because B2B breaks the conditions it assumes. Buying windows outlast the lookback, buying groups are large and mostly unseen, and deal counts are small. Experiments are impractical, so the credit can’t be checked. And channels with few deals but high impact lose to channels that are easy to capture.

Why do long buying windows break MTA?

Because the model only looks back as far as its window. B2B journeys run for months, and enterprise deals can run for years, yet many teams still report on a 90-day window. The report sees the end of the journey and credits it with the whole result.

The earliest touches, often the ones that created the demand, fall outside the window and score zero. On the three-year deal above, a 90-day window would see less than a tenth of the journey.

Why do large buying groups break MTA?

Because attribution tracks people, and a deal is bought by a group. The enterprise deal above had 22 people in its buying group. Most people in a buying group never fill out a form or click a tracked link, so the model credits the few who did.

The Pedowitz Group reports that half of salespeople still won’t attach a contact to an opportunity. The buying group is incomplete in the CRM before any model runs.

How much of the journey happens where tracking can’t see?

A lot of it. Buyers research in communities, podcasts, peer conversations and AI assistants, none of which leave a tracked touch. Refine Labs compared software to customer answers across 620 conversions. Software said web search sourced 78% of customers. Customers said 12%.

In the same study, customers credited podcasts with 53% of closed-won revenue, about $11.4 million. Attribution software credited podcasts with 0%.

Why do low deal volumes break MTA?

Because models need volume to separate signal from noise, and B2B doesn’t have it. Most B2B companies close hundreds or low thousands of deals a year. Spread across dozens of touch types and journeys that run for months, few patterns repeat often enough to trust.

With that few deals, adding or removing a handful of large ones can change the credit split.

Why do low-volume, high-impact channels lose?

Because the model rewards volume and clean records. Events, dinners, partner programs and syndication produce few touches per deal, often high-value ones, and their records arrive late or thin. The model ranks how easy a channel is to capture, not how well it works.

  • The bias compounds. Measurable channels get funded, gain volume, measure better and get funded again. The rest get starved.

Example: 32 regional dinners with 18 senior buyers each add up to 576 conversations. In most reports they show up as 32 separate rounding errors.

Why do B2B records break before analysis starts?

Because the records that matter most are entered by hand. Badge scans, partner spreadsheets and late uploads arrive thin, late or never, so the model scores the channel on its paperwork, not its effect.

  • An event lead typed in as “Tradeshow” carries no meaning.
  • A badge scan with a misread domain never joins its account.
  • A buyer downloads from three publishers. Dedupe keeps one, and two publishers show zero.
  • A partner spreadsheet never becomes a row. A missing row reads as a channel that didn’t contribute.
  • A late upload can land near conversion and make an opener look like a closer.
  • Demand is generated in one system and measured in another, so teams learn what worked months after the budget was spent.

What happens when a deal doesn’t end at closed-won?

Attribution treats each opportunity as a fresh start, but B2B buying runs in loops. Closed deals expand, lost deals come back, and stalled buying groups re-form. When a lost deal returns as a new opportunity, the touches from the first cycle are either dropped or credited to the second as if they were new.

Expansion deals have the same problem from the other side. They look like they came from nowhere, or from the last few touches before the upsell. I covered this with Nadia Davis and Eric Westerkamp on Hitchhiker’s Guide to Marketing Analytics: the job is staying in front of the right people the whole way around, not just to the first close.

Why can’t B2B teams fix it with experiments?

Because B2B can’t afford the tests or reach the sample size. Holdouts cost real revenue, small deal counts leave results too noisy to read, and long cycles mean the answer arrives after the budget decision.

  • Constantine Yurevich of SegmentStream said on the Humans of Martech podcast that for some businesses a geo holdout test costs $1 million to $2 million in lost revenue while it runs. The episode notes add that these tests only register effects from channels driving at least 5% of revenue, which leaves out the low-volume channels B2B teams most need to test.
  • Marketing mix models need years of spend variation, and long sales cycles smear the signal.

Can anyone check whether the credit is right?

Rarely, and that is the real danger. In consumer advertising, teams can test a model against an experiment. In B2B they usually can’t, so a wrong credit split never gets caught. The report looks precise and goes unchallenged.

Consumer research shows how wrong it can be. Blake, Nosko and Tadelis ran eBay’s paid search both ways. Standard regression put the return above +4,000%. The experiment measured -63% (Econometrica, 2015). Lewis and Rao ran 25 field experiments on advertising returns and found the median confidence interval on ROI was more than 100 percentage points wide (Quarterly Journal of Economics, 2015). Those are consumer results with far more volume than B2B has, so B2B has less room to test, not more.

Why measure at all if it’s never perfect?

Because the alternative is worse, not better. A blended view built on your own deals, clean records and buyer answers is the best first-party evidence you will have of what works. It won’t be exact. It will be yours, it will show where it is uncertain, and it beats single-touch models and platform-reported numbers on almost every decision.

Why not just use first-touch or last-touch attribution?

Because single-touch models aren’t neutral, they’re biased in a fixed direction. First touch credits whatever caught the lead first. Last touch credits whatever happened right before the form or the deal, usually a demo request or branded search. In a deal with 22 people and 2,274 touches, both credit one.

Other reasons the extra work pays off:

  • It runs on your own CRM. Ad platforms each report their own conversions, so the same deal can be claimed more than once. Your deal records count it once.
  • It settles operational choices well, like publisher A against publisher B, or which of two events to repeat.
  • It shows sequence: which touches tend to come before an opportunity moves forward.
  • It protects high-impact, low-volume channels from being cut by a model that can’t see them.
  • Clean capture helps sales too. Records that carry source, account and context improve routing and follow-up.
  • Finance trusts a method with written rules and stated ranges more than one precise-looking number.

What can B2B teams do instead?

Start with the records, then ask the buyer, widen the window, read by account and split at opportunity creation. Nothing here is perfect. Each step shrinks the error or shows you how big it is, and every method downstream inherits whatever the records got wrong, so fix the inputs before you add a model.

How do you fix capture at the source?

Make every record carry its meaning when it is created. A record with no source, campaign or account can’t be repaired later by a better model.

  • Capture source, campaign, asset, partner, event and account at creation.
  • Date every touch when the influence happened, not when the file arrived.
  • Roll small events into one named program, so 32 dinners get one number.
  • Keep every touch in the history. Dedupe the person, not the touches.
  • Match records to accounts, not just leads.

How do you ask the buyer how they found you?

Self-reported attribution means asking buyers directly what influenced them. Add an open-text question to every form and ask it again on the first sales call. Free text beats a dropdown because buyers name things you didn’t list.

  • Use “How did you hear about us?” as an open field. Sales logs the second answer in its own field.
  • Code the answers to channels every month. Add a code for AI assistants, and note which one.
  • Put software, form answers and sales answers side by side. The gap between them is the size of your dark funnel, the part of buying that leaves no trackable record.
  • After its study, Refine Labs published a hybrid attribution framework that combines software data with self-reported answers.

How long should the attribution window be?

Longer than your sales cycle. If the window is 90 days and your deals take nine months, the report sees only the last third of the journey. Rerun your report at 180 and 365 days and see which channels move.

Then look back from closed-won instead of forward from the lead. Starting at the deal shows you the whole path that actually led there.

Why measure accounts instead of leads?

Because a deal is bought by a group, and with few deals you need every signal each one gives you. Count engaged contacts and touches per account across the full cycle, and judge channels by what they did to the account.

  • Compare won accounts to lost and no-decision accounts.
  • Include accounts with no touch at all. Without them you only learn what your current targeting already favors.

Should you split attribution at opportunity creation?

Yes. Before and after an opportunity exists, you are asking different questions, and the data breaks in different ways. Measure them separately: what created the opportunity, and what moved it once it existed. Use different metrics on each side, and agree on when an opportunity starts before you show anyone the numbers.

  • Pre-opportunity asks what created the opportunity. Measure opportunities and pipeline created, by account. Long windows, the dark funnel and missing buying-group members hit hardest here, so lean on self-reported answers and account-level reads.
  • Post-opportunity asks what moved the deal. Measure win rate, cycle length and deal size. It has a defined start date, a shorter window and named people on the opportunity, so it is far more measurable.

In the enterprise deal above, almost four of every five touches came after the opportunity was created. Attribution that stops at opportunity creation would have missed most of the work.

Where the split breaks:

  • Opportunity creation is a sales habit, not a fact. One rep opens an opportunity at the first call, another after a demo. Without standard creation criteria the line moves by rep and region, and channels gain or lose credit with it.
  • Post-opportunity credit leans toward deals that were already winning. Healthy deals get more meetings, dinner invites and executive time. Compare won and lost opportunities with the same exposure, or you will credit the touches winning deals attract.
  • Contact roles are thin. If half of reps don’t attach contacts, contact-level attribution after creation is mostly empty. Read it at account level.
  • It invites a turf war. Marketing claims pre, sales claims post, and the fight moves to the boundary. Agree on the criteria first.

Decide up front how a returning lost deal is handled: a new opportunity that inherits the earlier touches, or a clean start. Either can work. Switching between them quarter to quarter can’t.

Why should you write down your attribution rules?

Because every credit split is a choice, and choices nobody can see get disputed. Write the rules down, share them with sales and finance, and log every change with a date. Then a shift in the numbers can be traced to the market or to a rule change, instead of argued about.

What the rules should cover:

  • What counts as a touch, and what doesn’t.
  • The channel names and what goes in each.
  • The model, its weights and the lookback window.
  • When an opportunity is created.
  • How reopened deals, returning lost deals and expansions inherit earlier touches.
  • How duplicates are merged, and how late records are dated.
  • How self-reported answers are coded.
  • A change log with dates.

Can you run experiments without a big holdout budget?

Yes, small ones. Pause or ramp one channel in one region or segment, compare against similar accounts, and measure pipeline at the right lag instead of next week. It is noisy and it is not proof. It still beats guessing.

What is MTA still good for?

  • Sequencing touches inside a journey.
  • Comparing sources inside one channel, like publisher A against publisher B.
  • Bottom-funnel click channels like paid search.

How do the methods compare?

No single method answers every question. MTA is best at order and source comparison, self-reported data at what buyers remember, holdouts at cause, and mix models at the market level. Each has a weakness the others partly cover.

Method What it answers What it needs Main weakness
Single-touch (first or last) Which touch started or finished the journey One timestamped touch per deal Ignores everything in between, and is biased in a fixed direction
Multi-touch attribution Order of touches, and source comparison inside a channel Clean person-level touch records Misses most of the buying group, and its credit can’t be validated at B2B volumes
Self-reported attribution What buyers say influenced them An open-text form field and monthly coding Favors the last memorable touch
Holdout experiments Whether a channel caused lift Budget and volume to hold out and detect lift Can cost real revenue and be underpowered
Marketing mix model Channel contribution at the market level Years of spend variation Long sales cycles smear the signal

How should you spend when the data disagrees?

Act when your sources agree, and treat disagreement as information. A gap between software, self-reported and sales answers marks the part you can’t measure yet. That is a reason to move slowly, not a reason to cut.

  • When software, self-reported and sales answers agree, act on it.
  • Move budget in small steps, 10 to 20% at a time, and watch two quarters.
  • Keep a fixed bucket for brand, events, partner and syndication so measurable channels can’t absorb it.
  • Report ranges, not point estimates. Show the unattributed share on the same page as the numbers.

Where does this plan fall short?

The plan narrows the error. It does not prove cause. Every step here has a known bias, and you should know it before you present the numbers.

  • Self-reported answers favor the last memorable touch. People say “Google” or “a colleague” and skip the webinar they saw in March.
  • Sales will fill the field with whatever is easiest.
  • Looking back from closed-won still favors what you already did, unless lost accounts are in the comparison.
  • Natural experiments are noisy and easy to over-read.

What can you check this week?

Three checks, in this order. Each takes a few hours and tells you how much to trust the report you already have.

  1. Rerun last quarter’s report at 180 days and note which channels move.
  2. Pull 50 closed-won deals and compare the software source to what the buyer says.
  3. Measure touch-to-record time for a trade show and a paid ad. If it is weeks against minutes, the model is ranking latency.

Common questions about B2B attribution

Is first-touch or last-touch attribution good enough for B2B?

Rarely. Both give all the credit to one touch, so they are biased in a fixed direction. First touch favors whatever caught the lead, last touch favors whatever came right before the form or the deal. Over long journeys with large buying groups, both ignore almost everything in between.

Should B2B teams split attribution at opportunity creation?

Yes. Before the opportunity, measure what created it, using opportunities and pipeline created by account. After it, measure what moved it, using win rate, cycle length and deal size. Standardize when reps create opportunities first, or the line moves and credit moves with it.

Is multi-touch attribution dead?

No, but it is narrower than vendors say. It still works for sequencing touches inside a journey, comparing sources inside one channel, and bottom-funnel click channels like paid search. It should not be your only basis for moving budget between channels.

What is the difference between multi-touch attribution and marketing mix modeling?

Multi-touch attribution works from individual touches and splits credit among them. Marketing mix modeling works from aggregate spend and outcomes over time. MTA needs clean person-level records. A mix model needs years of spend variation, and long B2B sales cycles blur its signal.

How do I track buyers who found us through ChatGPT or other AI tools?

You usually can’t see them in tracked touches, because they arrive with no click path. Add an open-text “How did you hear about us?” field to forms and ask again on sales calls. Code the answers each month, with a separate code for AI assistants and a note on which one.

How long should a B2B attribution window be?

Longer than your sales cycle. If deals take nine months, a 90-day window sees only the last third of the journey. Rerun your report at 180 and 365 days and see which channels move.

Why doesn’t multi-touch attribution work with low deal volume?

Models need volume to separate signal from noise. Most B2B teams close hundreds or low thousands of deals a year. Spread across many touch types and long journeys, few patterns repeat often enough to trust, and a handful of large deals can change the credit split.

Why do events and partner programs look weak in attribution reports?

Their records arrive late, with little detail, or never. A badge scan, a partner spreadsheet or a hallway conversation leaves a thinner trail than an email click. The report ranks how well a channel is captured, not how well it works.

Does self-reported attribution replace software attribution?

No. Self-reported answers favor the most memorable touch. Put software, form answers and sales answers side by side. Where they agree, act on it. Where they disagree, treat the gap as the part you can’t measure yet.

Sources


About the author. Mehul Nagrani has spent 20 years working at the intersection of data, AI and go-to-market. He is CEO of Integrate and CaliberMind.