Measurement
Lead attribution in the agentic AI era: which model to use and why
By John Utley
Published

Buyers still leave some tracks. A form fill, a booked meeting, a proposal review. More of the real first impression now happens in a place your analytics never see. Someone asks an AI assistant to compare vendors, then arrives as direct traffic, or fills out a form that credits the last ad they clicked.
This article explains the standard models, what agentic AI changes, and the credit rule Sophizo uses with B2B clients. A shorter version also lives on the Fractional CMO page.
The short answer
For most B2B companies we work with, we use a W-shaped model. Thirty percent of the credit goes to the first touch, including an AI assistant that named you. Thirty percent goes to the booked conversation. Thirty percent goes to the proposal. The last ten percent is shared across the touches in between.
We check first touch against a simple question: how did you first hear about us? When the buyer's answer and the tracking disagree, the buyer wins. This is our reporting rule, not a claim that buyer recall is perfect. We retain both sources. The model shows where results come from. It does not promise that every dollar can be traced to revenue.
The seven standard models
Each model answers a different question. None of them is wrong. The risk is using one that answers the wrong question for your sales cycle.
First touch
All of the credit goes to the first thing that brought the lead in. It is useful for discovery, and blind to what happened after.
Last touch
All of the credit goes to the final touch before the conversion. Many CRMs still use this by default. It names what closed the loop, and it over-credits retargeting, branded search and the last email in a sequence. Last touch is not dead. It is incomplete.
Linear
Credit is split evenly across every touch. That feels fair when every step mattered about the same. It gets noisy when a long cycle includes many low-value touches.
Time decay
Touches closer to the sale get more credit. Helpful when late-stage work moves the deal, and weak at explaining how the buyer found you.
U-shaped, also called position-based
Forty percent goes to the first touch, forty percent to the lead-creation or signup touch, and the remaining twenty percent is split across the middle. This fits short-cycle, self-serve motions, where those two moments are the decision.
W-shaped
About thirty percent each goes to first touch, to the meeting or opportunity, and to the proposal or decision stage, with the rest spread across the middle. This fits B2B companies with a real sales conversation and a proposal.
Data-driven, or algorithmic
Machine learning assigns credit from your own conversion data. It can be useful once you have enough conversions for the model to learn. It is a poor starting point when volume is still thin.
Two methods that do not need a click
Two other methods help when the click trail is incomplete.
Incrementality tests, also called pause tests or holdouts. You pause a channel, a campaign or an agent for a set group, and you measure what changes in meetings, pipeline and revenue. On the Fractional CMO page we already call these holdouts and proof tests. Where volume and a suitable comparison allow, they help distinguish cause from correlation.
Marketing mix modeling. A statistical model of spend against revenue over time, without needing every click. It is worth considering once you have enough suitable history. It does not replace the W-shaped ledger. It sits beside it as a check on whether the big budget shifts are paying back.
Four ways agentic AI changes attribution
1. The buyer's research goes dark.
Buyers ask AI assistants to compare vendors and often never click a link. Models then credit direct or organic traffic and miss the real introduction. That is why the self-reported question matters more now, and why AI-assistant mentions deserve their own channel. Read more in Your next buyer may never visit your website.
2. Buyers send their own agents.
An agent may visit your pricing page, fill out a form or book a demo on the buyer's behalf. Those sessions can look like people in your analytics. Tag identifiable agent visits separately so you do not treat them as human demand. Not every automated visit can be reliably identified.
3. Your own bots create touches at scale.
Lifecycle email agents, research briefs and outbound drafts can fill a record with touches. Multi-touch models then over-credit whatever the bots did last. Log every agent touch with its cost, including review time, and judge agents on cost per result, not touch count. Your own agents can earn middle-touch credit. They do not get first-touch credit in Sophizo's model.
4. Correlation gets weaker.
More automated touches and fewer visible clicks make rule-based models noisier. Before you scale a channel or an agent, use a pause test where volume allows. That is the same holdout idea already on the Fractional CMO page: a share of matched accounts gets no agent activity, and the difference helps estimate the contribution.
The model we recommend, and why
W-shaped fits a B2B cycle with three real milestones: the first time the buyer becomes aware of you, the first real conversation, and the proposal that asks for a decision. Last touch would over-credit the late ads and the last follow-up. U-shaped would skip the proposal. Data-driven can come later, once monthly conversions are high enough for the model to be stable.
Four rules sit on top of the split.
What the buyer says counts. Every new lead is asked how they first heard about you, with "AI assistant" as an option. When that answer and the tracking disagree on first touch, we go with the buyer.
AI gets its own line. Mentions in AI assistants and AI-powered search answers are a channel, not a footnote inside organic search. Identifiable visits from a buyer's agent are tagged apart from visits by a person.
Your agents earn credit, not first-touch credit. Every agent action is logged against the account, with its cost and review time. Agent-sourced and agent-influenced pipeline still matter as operating metrics. They describe work your agents did in the CRM. They are not the same thing as first-touch credit.
Proof over correlation. Use pause tests before scaling where volume allows. When there is enough suitable history, consider marketing mix modeling as a second check.
For a short-cycle, self-serve product, where the second milestone is a signup rather than a proposal, we use U-shaped instead: forty percent to first touch, forty percent to signup, and twenty percent across the middle. The same four rules still apply.
This model shows where results come from. It does not promise that every dollar can be traced to revenue. We already say the same thing on the marketing ledger: we account for every dollar you spend, and we tell you which results we can prove.
A practical next step
If you want this model applied to your pipeline, start with a short marketing survey or book a 30-minute call. You can also read how we give credit, and how it connects to the monthly ledger, on the Fractional CMO page.