That is not a reason to skip it. It is a reason to use it for the decisions it can actually support — roughly ranking channels, killing obvious losers, spotting sources that never convert — and to reach for experiments when the decision is expensive.
In short
- Every attribution model is an assumption made visible, not a measurement.
- The same six deals can hand a channel 70,000 or 0 depending on the model. The worked example below shows it.
- Platforms have narrowed the menu: GA4 now supports three models after removing four in 2023 (Google).
- Data-driven models need volume. Google recommends at least 200 conversions and 2,000 ad interactions in 30 days (Google Ads).
- MitHub's ladder: record the source honestly, then test, then model. Most teams do it backwards.
- Before choosing a model, name the decision it will change. If no decision changes, do not build the report.
Start where the money is
MitHub's revenue engineering method begins at the payment and walks backwards: payment → decision → conversations → first contact → source (follow the money). Attribution is the last link of that chain formalized into a repeatable system.
Which means the first attribution question is not "which model?" It is: can you take one payment from last month and reconstruct where it came from? If the answer is no, no model will rescue you, and the fix belongs in CRM architecture, not analytics.
The models, and what they assume
| Model | Rule | Implicit assumption | Bias |
|---|---|---|---|
| First touch | 100% to the first interaction | Discovery is what matters | Over-credits awareness channels |
| Last touch | 100% to the last interaction before conversion | Closing is what matters | Over-credits whatever is nearest the sale |
| Linear | Equal split across all touches | All touches matter equally | Rewards channels that appear often and cheaply |
| Time decay | More credit to recent touches | Influence fades | Penalizes long nurture |
| Position-based | Most credit to first and last | Discovery and closing dominate | Arbitrary percentages |
| Data-driven | Model learns credit from converting and non-converting paths | You have enough data to learn from | Opaque; needs volume |
Vendors implement these differently, and the differences are not cosmetic. HubSpot's attribution reporting includes linear, first interaction, last interaction, U-shaped (40% first, 40% lead creation, 20% spread across the middle), W-shaped (30/30/30 with 10% across the middle), time decay with a seven-day half-life, and full path (22.5% to each of four key interactions, 10% across the middle), on Marketing Hub Professional and Enterprise (HubSpot).
Google went the other way and simplified. GA4 now supports data-driven attribution, paid and organic last click, and Google paid channels last click; first click, linear, time decay and position-based were retired in November 2023. Direct visits are also excluded from credit unless the entire path is direct (Google).
Read that last detail twice, because it is the kind of rule that quietly reshapes a report.
A worked example: the same quarter, three answers
This is a hypothetical example with invented numbers, built to show the arithmetic.
Six deals closed for a total of $180,000. Their touch paths:
| Deal | Value | Path |
|---|---|---|
| D1 | $30,000 | Referral → Webinar → Sales call |
| D2 | $20,000 | Google Ads → Email → Sales call |
| D3 | $40,000 | Referral → Sales call |
| D4 | $15,000 | LinkedIn → Webinar → Email |
| D5 | $50,000 | Outbound AI call → Email → Sales call |
| D6 | $25,000 | Google Ads → Webinar → Sales call |
Now apply three models. Linear splits each deal's value equally across its touches.
| Channel | First touch | Last touch | Linear |
|---|---|---|---|
| Referral | $70,000 | $0 | $30,000 |
| Google Ads | $45,000 | $0 | $15,000 |
| $15,000 | $0 | $5,000 | |
| Outbound AI call | $50,000 | $0 | $16,667 |
| Webinar | $0 | $0 | $23,333 |
| $0 | $15,000 | $28,333 | |
| Sales call | $0 | $165,000 | $61,667 |
| Total | $180,000 | $180,000 | $180,000 |
Three observations that survive into real life:
- Last touch is useless here. The final touch before a closed deal is nearly always the sales call, so the model reports that sales closes deals. True, and worth nothing.
- Email looks invisible under first touch and important under linear. Whether you keep funding it depends entirely on a modelling choice.
- Webinars earn nothing under either single-touch model and $23,333 under linear, because they always sit in the middle.
And one more choice that moves everything
Is a sales call a "touch"? If you decide it is an internal activity rather than a marketing touchpoint and remove it from the paths, the same linear split gives:
| Channel | Linear, sales call excluded |
|---|---|
| Referral | $55,000 |
| $40,000 | |
| Webinar | $32,500 |
| Outbound AI call | $25,000 |
| Google Ads | $22,500 |
| $5,000 | |
| Total | $180,000 |
Referral moved from $30,000 to $55,000 because of a definition, not because of anything that happened in the market. This is why MitHub insists that every attribution report carries its rules in writing next to the numbers.
MitHub's attribution honesty ladder
Rather than asking "which model is right?", ask "what level of honesty can our data support?"
| Level | What it is | What it costs | What it can answer |
|---|---|---|---|
| 0 | Nothing recorded | — | Nothing |
| 1 | Self-reported: "How did you hear about us?" captured at the sale | One question | Which channels buyers remember |
| 2 | One source field per record, written once at creation, never overwritten, plus cost per channel | Data discipline | Cost per deal by source, roughly |
| 3 | Multi-touch models over logged interactions | Clean tracking, consent, integration | Relative contribution, under stated assumptions |
| 4 | Experiments: holdouts, on/off tests, geo tests | Willingness to lose some volume | What actually changes when you stop |
The mistake MitHub sees most often is jumping to level 3 while level 2 is still broken — building a multi-touch dashboard on top of a source field that three systems overwrite. Level 2 plus one experiment a quarter beats an elaborate model that nobody believes.
Level 1 deserves more respect than it gets. Referrals, word of mouth, a branch manager's reputation and a podcast mention are invisible to every tracker and obvious to the buyer. Ask the question, store the answer in a picklist plus a free-text field, and compare it against your tracked source monthly. The gaps are where your real growth is happening unmeasured.
The limits, stated plainly
- Offline and phone. Walk-ins, inbound calls and branch visits exist outside web analytics. They must be logged as touches in the CRM or they do not exist.
- Consent and privacy. Rejected cookies, blocked trackers and cross-device journeys break path data. Modelled gaps are estimates.
- Long cycles. If your sales cycle is four months and your lookback window is 30 days, the model cannot see the beginning.
- Volume. Data-driven models need scale; Google's own recommendation is 200 conversions and 2,000 ad interactions per 30 days (Google Ads). A twelve-branch business closing forty deals a month will never reach it for closed revenue.
- Group buying. Attribution tracks individuals; companies buy in committees. The person who clicked the ad is often not the person who signed.
- Correlation. A channel that touches every winning deal may be a symptom of intent, not a cause of it.
What to do instead when the decision is expensive
Attribution ranks. Experiments decide. If you are about to double or kill a budget line, run a test:
- Holdout: withhold the activity from a random 10–20% of eligible leads and compare outcomes.
- Geo / branch on-off: run the channel in some locations and not others for a fixed period.
- Sequenced pause: turn it off for four weeks, watch the lagging metrics, turn it back on.
These are cruder than a dashboard and far more convincing, because they measure the difference the activity makes rather than the credit a rule assigns. It is the same logic as proving value fast: produce facts, not opinions.
An attribution setup you can defend
- One immutable source field on the record, written at creation by the system that created it. Never overwritten by later activity.
- A self-reported question at the moment of purchase, stored separately.
- Offline touches logged in the CRM: calls, visits, events, referrals.
- A single published definition of channels, touches and the lookback window, so two reports cannot disagree.
- One default model for routine reporting, with the rule printed on the report.
- A quarterly experiment on the largest line item.
- A named owner for the whole thing.
Notice that six of the seven are architecture and discipline, not analytics. That is the real finding of most attribution projects.
The decision test
Before building any attribution report, finish this sentence: "If this report says X, we will do Y."
If nobody can complete it, the report is entertainment. If they can — "if paid search costs more than $4,000 per funded deal, we move the budget to reactivation calling" — then you know exactly which numbers must be trustworthy, and you can stop arguing about the rest.
Attribution is one instrument in a revenue system, useful next to a leak map and a defensible forecast. The Faculty of Revenue Reverse Engineering teaches the whole set, starting from the only number that is never a model: the payment that actually arrived.