When attribution models change the answer

Eligible workspaces	Attributable signups	Attribution models	Lookback window
94	545,795	7	90 days

The result is not a claim that one model is correct and another is wrong. It is a warning that a channel ranking can depend on the accounting rule used to produce it.

The median looks small. The tail does not.

The median net difference in channel credit allocation was 3.0 percentage points when the model changed from first touch to last touch. At the 75th percentile the shift was 11.2 points, and at the 90th it reached 19.0 points.

The workspace-level distribution shows variation that a pooled result can hide. Larger workspaces carry more weight in pooled figures, and opposing channel shifts can cancel out.

The benchmark in one table

Comparison	25th	Median	75th	Pooled
First touch vs last touch	0.5 pp	3.0 pp	11.2 pp	1.7 pp
Linear vs last touch	0.4 pp	1.5 pp	4.3 pp	0.7 pp
Linear vs first touch	0.4 pp	1.5 pp	6.9 pp	0.9 pp
Last touch vs last non-direct	0.9 pp	3.7 pp	13.9 pp	2.1 pp

Decision note. If the attribution view supports a material budget change, compare at least two models before acting.

Direct traffic is the hinge of the model debate

Direct carried a median 50.4% of signup credit under linear attribution. The same median fell to 40.8% under non-direct models and rose to 54.3% under last touch. No other channel showed a comparable model-driven range.

Organic search moved in the opposite direction because non-direct models reassign credit that would otherwise stop at direct. Its median share ranged from 14.6% under last touch to 20.8% under last non-direct attribution.

What changes. Agree on how direct traffic should be treated before comparing teams, quarters or campaign portfolios.

What does not. The reassignment describes credit placement. It does not establish causal lift.

A different winner is common enough to plan for

First-touch and last-touch attribution produced a different top-credited channel in 16 workspaces, or 17.0% of the eligible sample. Comparing all seven models raised the count to 18 workspaces, or 19.1%.

The quarterly rate moved between 8.3% and 17.7% for first touch versus last touch. Because the eligible sample changes by quarter, the full-period figure is the stronger benchmark; quarterly movement is a stability check, not a trend claim.

Reporting implication. A single top-channel label should be presented as model-specific whenever the ranking changes under a reasonable alternative.

Sensitivity is concentrated, not universal

Direct had the widest median model spread at 3.7 percentage points, followed by organic search at 2.0 points and referral at 0.3 points. Paid search, paid social and most smaller channels had negligible median spread.

Pooled paid-channel credit stayed between 10.1% and 10.5% across all seven models. Paid search alone ranged from 6.3% to 6.6%. Paid-channel credit was comparatively stable in this dataset. This does not establish stability for every workspace, efficiency or incrementality.

Practical read. Prioritize sensitivity checks based on your own channel mix; do not assume every channel needs the same level of scrutiny.

Publish the model, then publish the sensitivity

Step	Practice	Instruction
1	Name the model	State the exact attribution model and 90-day lookback beside the result.
2	Show an alternative	Pair the primary view with one reasonable sensitivity model, especially when direct is large.
3	Separate credit from causality	Do not translate attribution stability into efficiency, lift or causal impact.
4	Flag ranking changes	If the top channel changes, report the result as model-dependent rather than definitive.

Methodology

Source and period. Anonymized, aggregated Usermaven product-analytics data from 1 October 2025 through 30 September 2026, extracted read-only on 5 November 2026. Quarters use UTC.

Sample. Ninety-four eligible company workspaces and 545,795 signups attributable to a tracked session within 90 days.

Eligibility. A workspace-quarter required at least 10 verified signups, continuous pageview tracking in the quarter and previous three months, verified signups in every month, at least 50% identity linkage and at least 10 attributable signups. Duplicate, internal and demo workspaces were excluded.

Journey and exclusions. Sessions containing pageviews from browser IDs linked to the user were included up to signup. Bots, deleted events, tag-preview sessions and browser IDs shared by two or more accounts were excluded.

Models. Linear, first touch, last touch, first non-direct, last non-direct, position based 40/20/40 and time decay with a seven-day half-life. Every signup contributes exactly one unit of credit under each model.

Disagreement metric. For one workspace and two models, half the sum of absolute differences in channel shares measures the net difference in channel credit allocation. It does not measure the percentage of individual signups reassigned.

Aggregation. Metrics are calculated per workspace and summarized with medians and percentiles. Pooled values combine all attributable signups and therefore weight larger workspaces more heavily.

Limits and independence

Cookie-based identity can miss cross-device activity. Direct can include bookmarks, private messaging and sessions with lost referrers. The conversion is a verified signup; pipeline and revenue can produce different patterns. No verified ARR or company-size segmentation was available.

Research independence. This report uses anonymized and aggregated Usermaven data. It was not sponsored, and no company names, campaign names or personal data were used.