U-shaped, W-shaped, time-decay: picking a multi-touch attribution model that fits your funnel

If you've read our multi-touch attribution primer, you know that the model is the easier half of attribution. The hard part is the data. But "easier" doesn't mean "obvious" — pick the wrong model and your entire channel ranking inverts. Budget decisions flip. The content team argues with the paid team for a quarter.

This post is a deep-dive on five multi-touch attribution models, when each one matches your funnel, and the shorthand I use to pick one fast.

The five models, visually

Imagine a journey of 5 touches: [Blog post] → [Retargeting ad] → [Webinar] → [Demo] → [Proposal]. Here's how each model distributes 100% of credit:

| Model | Touch 1 | Touch 2 | Touch 3 | Touch 4 | Touch 5 | |-------|---------|---------|---------|---------|---------| | Linear | 20% | 20% | 20% | 20% | 20% | | U-shaped | 40% | ~7% | ~7% | ~7% | 40% | | W-shaped (if Touch 3 is opp-creation) | 30% | 5% | 30% | 5% | 30% | | Time-decay (7-day half-life, 14-day journey) | ~6% | ~12% | ~18% | ~28% | ~36% | | Algorithmic | Whatever the data says |

Five models, five very different stories about who deserves credit for a deal. Same journey, same conversion.

Model 1: Linear

Every touchpoint gets equal credit, regardless of position, timing, or anything else.

The intuition: if you can't prove any one touch mattered more, don't claim it did.

When to use it:

Short sales cycles (under two weeks)
High-volume conversion events (e-commerce, transactional SaaS)
Small number of touchpoints per journey (1–3)
When you want a simple story that no one can argue with

When to avoid it:

Long B2B cycles with 10+ touches where you know the first and last matter more
When late-funnel sales activity is dominating the journey (SDR emails, sales calls) and you don't want to over-credit sales ops

The risk: linear will always credit whichever channel fires most frequently, regardless of whether those touches moved anyone. A spammy retargeting campaign that fires 50 impressions per journey will look like your best channel under linear.

Model 2: U-shaped (position-based)

First and last touches each get 40%. The middle 20% is split among all middle touches.

The intuition: the introduction (first touch) and the close (last touch) are the two moments of genuine commitment in the journey. Everything in between is reinforcement.

When to use it:

B2B with cycles of 30–180 days
Funnels where the first touch is usually meaningful content (blog, referral, demo request) rather than tangential (social impression, random ad)
When you want to balance demand-gen credit (first touch) with sales/conversion credit (last touch)

When to avoid it:

If first touches are systematically weak (retargeting ads, auto-tracked LinkedIn impressions) — U-shaped will overcredit them
If your last touches are always the same channel (e.g., SDR email) — every deal will look like SDRs are closing, which will be politically fraught

Default for most B2B: this is where I start if the team doesn't have strong priors.

Model 3: W-shaped

First touch, opportunity-creation touch, and last touch each get 30%. The remaining 10% is split among all other middle touches.

The intuition: U-shaped, but with a third "major milestone" added: the moment the prospect was qualified into an opportunity.

When to use it:

You have a genuine, human-gated opportunity-creation process. An SDR qualifies the lead on a call. A human decides "this is a real opportunity."
Mid-to-long B2B cycles where MQL→SQL→opp is a meaningful funnel progression, not automated workflow clicks
When your marketing and sales teams both want credit for the "handoff moment" that tips a prospect from curious to committed

When to avoid it:

Your opportunity creation is automated (a workflow triggered by lead score hitting 80). In that case, you're adding credit to an arbitrary data point.
Short cycles where there's no distinct opp-creation moment — the prospect booked a demo and closed in the same week

The subtle bug: W-shaped is only as good as your opp-creation data. If "opportunity created" in your CRM is accurate to the hour, great. If it's a batch-updated field that trails real events by 5 days, the model will miscredit.

Model 4: Time-decay

Recent touches get more credit. Older touches get less. Usually implemented with exponential decay — a 7-day half-life means a touch 7 days before conversion gets half the credit of a touch on the day of conversion. Two weeks prior? Quarter credit. Three weeks? Eighth.

The intuition: the last thing someone saw before converting was probably the thing that moved them. Memory decays, urgency decays, influence decays.

When to use it:

Cycles under 30 days where the full journey is genuinely recent
Journeys that feel "impulsive" — someone saw a Twitter post, clicked through, demoed, bought, all in 10 days
When you want to de-emphasize the "random blog post from 6 months ago" touches that U-shaped will overcredit

When to avoid it:

Content-marketing-heavy businesses where the first touch (six months earlier) genuinely mattered
Any time you want to understand top-of-funnel ROI — time-decay will systematically undercredit the channels that seed interest months in advance

The tell: if your CMO loves time-decay, it's probably because it credits their team's recent work. If your head of content hates time-decay, same reason inverted.

Model 5: Algorithmic (data-driven)

A machine-learning model (usually Markov chains or Shapley values) infers credit distribution from the data. Touches that reliably appear before conversions get more credit. Touches that don't move the needle get less.

The intuition: let the data decide. No prior beliefs about position or time.

When to use it:

You have thousands of conversions per month. Not hundreds. Thousands.
You have clean, high-fidelity touchpoint data across all channels
You have the internal capacity to explain the model when the CMO asks why

When to avoid it:

Mid-market B2B. If you close 40 deals a quarter, algorithmic models will fit noise, not signal. You'll get a model that looks sophisticated and changes its rankings quarter over quarter for random reasons.
You need to explain the model to a board. "The ML said so" is not a defensible answer to a CFO.
Your touchpoint data has gaps. Algorithmic amplifies data quality problems.

The decision matrix

A shortcut I use when a team asks "which model should we use?":

| Question | Answer → Model | |----------|---------------| | Sales cycle under 2 weeks? | Linear or Time-decay | | Sales cycle 2 weeks – 3 months, no strong first-touch? | Time-decay | | Sales cycle 30 days – 6 months, strong first-touch? | U-shaped (default) | | Long cycle with genuine opp-creation milestone? | W-shaped | | >5K conversions/month, good data team? | Algorithmic as a comparison layer |

If you're reading this and you're a mid-market B2B company and none of the specific conditions apply: go U-shaped. Revisit in 6 months when your data is cleaner.

Running two models in parallel

Almost every mature RevOps team I've talked to ends up running two models in comparison, not one. The point isn't to pick a winner — it's to stress-test the story.

Common pairings:

U-shaped + time-decay: is the story the same whether you credit early or late? If yes, you trust the ranking more. If no, figure out which channels diverge and why.
U-shaped + linear: is any high-ranked channel a high-volume spam channel (heavily credited by linear)? Then linear is wrong and U-shaped is closer.
U-shaped + W-shaped: is there a big delta for channels that contribute around opp creation? If yes, marketing operations and SDRs are adding more value than U-shaped implies.

Reporting two models side-by-side in your Monday dashboard is overkill. Pick a primary for weekly reporting; run the comparison quarterly.

For more on the weekly reporting cadence, see our Monday morning revenue dashboard.

Where Elir fits

Elir runs all five models out of the box — U-shaped, W-shaped, linear, time-decay, and our probabilistic keyword-weighted model we wrote about in keyword-level attribution. You can switch between them without rebuilding reports, run comparisons, and see the channel ranking diverge (or not) in real time. If you're stuck picking a model because every vendor oversells their own, book a 20-minute walkthrough and we'll show you what the actual differences look like on your pipeline.

TL;DR

Five models, five stories. Linear for short cycles. U-shaped for most B2B. W-shaped if you have real opp-creation milestones. Time-decay for recent-touch-dominated journeys. Algorithmic only if you have the volume. Run two in comparison to stress-test the story. If in doubt, U-shaped.