Attribution

Attribution is broken because it credits the last click, splits your data across siloed tools, and cannot see the dark funnel; the fix is one joined identity layer that ties every touch to revenue. Fix the data first, then let agents measure against money instead of clicks.

TL;DR in 60 seconds

• Last-click bias hands all the credit to the final touch and starves everything that created the demand.
• Siloed data means ads, site, email, and CRM each tell a partial story that never reconciles.
• The dark funnel (podcasts, communities, word of mouth) is invisible to any click-based model.
• The fix is a single identity layer joining every touch to revenue, not a smarter attribution model on broken data.
• Agents then measure each experiment against revenue and margin, which is the number that actually matters.

Three ways attribution fails

Last-click bias. Most reporting gives the closing touch full credit. So the branded search that a customer typed because of a podcast six weeks earlier looks like the hero, and the podcast looks like a waste. You optimise toward the visible touch and slowly defund demand creation.

Siloed data. Your ad platform counts conversions its way, your email tool counts its way, and your CRM counts a third way. None of them share an identity, so the same buyer is three different rows. Totals never add up, and every team defends its own inflated number.

Dark-funnel gaps. A lot of real influence leaves no click: a Slack recommendation, a conference chat, a founder's post reshared privately. Click-based attribution cannot see any of it, so it systematically undercredits the channels that build reputation.

The fix is a data problem, not a model problem

Teams keep reaching for a cleverer attribution model (first-touch, linear, time-decay, data-driven) and layering it on the same fragmented data. That just produces a more confident wrong answer.

The real fix is upstream: build one joined identity layer where every touch, from every channel, resolves to the same person and the same revenue record. Once ad, site, email, and CRM events share an identity, you can ask what actually influenced a closed deal instead of guessing. This is exactly what it means to wire agents into your data layers, and it starts with instrumenting the funnel for agents.

How agents change the measurement

With a joined layer in place, agents stop optimising toward clicks and start measuring every experiment against revenue and margin. An agent can run a test, wait for the revenue to land, and judge the variant on money rather than a mid-funnel proxy.

This also sidesteps the attribution wars. You are no longer arguing about which touch deserves credit; you are running controlled experiments and reading the revenue difference. That shift, from crediting touches to testing against money, is the heart of experiments that move margin.

FAQ

Is multi-touch attribution the answer?

Not on its own. Multi-touch spreads credit across touches, but if it runs on siloed data with an incomplete identity graph, it just distributes a wrong total more evenly. Fix the identity layer first; the choice of model matters far less than whether the underlying touches are joined to real revenue.

What is the dark funnel?

The dark funnel is every influence that leaves no trackable click: private shares, community recommendations, podcasts, in-person conversations. It often drives a large share of real demand, yet click-based attribution cannot see it, so those channels look worthless in reports even when they are doing the heavy lifting.

Can agents attribute revenue perfectly?

No, and they should not claim to. Perfect attribution is not achievable. What agents can do is run controlled experiments and measure the revenue difference, which answers the question that attribution was trying to answer all along: did this change make money?

Where do I start if my data is a mess?

Start with the identity layer, not the reporting. Get every channel's events resolving to one person and one revenue record. Until that exists, better dashboards only render the same broken data more convincingly. Instrumentation before insight, every time.

Cadence builds the joined identity layer first, then puts agents on top that judge every experiment by revenue, not by whoever grabbed the last click.