Spillover & Contamination
Consensus definition
The Stable Unit Treatment Value Assumption (SUTVA) requires no interference — one unit's outcome is unaffected by which treatment others receive — and no hidden treatment versions1. When interference holds, standard difference-in-means estimators are biased because comparison-group outcomes are no longer independent of the treatment2. That failure is spillover or contamination: the treatment leaks into the comparison group and shrinks, or even reverses, the measured effect. In SEO/GEO at least four mechanisms generate it: internal-link changes redistribute authority across the link graph3; keyword cannibalisation lets a treated page suppress a comparison page on the same query; sitewide signals (crawl budget, domain authority) are shared by every page; and topic-cluster cohesion ties a page's ranking to the whole cluster, so improving one pillar lifts untreated cluster pages34.
rhinegold operator refinement
Rhinegold's reframe: a comparison page that lives in the same topic cluster or internal-link neighbourhood as a treated page is not a clean control — it is a partially treated unit. A naive within-site comparison then overstates the baseline, suppresses the measured lift, or inverts its sign when cannibalisation dominates. Pool- or peer-based comparison groups — drawn from clusters or domains with minimal link-graph and topical overlap — restore the independence SUTVA requires.
Operational use
Before assigning experiment buckets, audit whether comparison pages share internal-link sources, keyword targets or cluster membership with treated pages. For internal-link tests, measure three cohorts at once — source pages, destination pages, and collateral pages outside the treated neighbourhood — and flag any comparison page receiving indirect link equity or topical association as potentially contaminated.
Measurement boundary
Spillover magnitude is rarely directly observable — its presence is inferred from implausibly small or negative effects, or from comparison pages moving in lockstep with treated ones. Cluster-level assignment reduces interference but cannot remove sitewide signals, and general-equilibrium effects (a sitewide authority lift from a large push) can persist even with peer-domain controls.
Distinct from
From the control group: that is the intended solution — untreated comparison units; spillover is the condition under which the control is no longer truly untreated. From difference-in-differences: DiD assumes parallel trends; spillover breaks that by making the comparison group move for treatment-related reasons. From the counterfactual: spillover makes the observed comparison outcome a poor proxy for it, because the comparison group was itself affected.
Common mistakes
- Assigning comparison pages by traffic or template similarity without checking link-graph or cluster proximity — the most common route to contaminated buckets.
- Treating cannibalisation as a content-quality issue rather than an interference problem that biases effect estimates downward.
- Running internal-link tests while measuring only the source pages, missing the authority redistribution on destinations and dilution of collateral pages.
- Confusing symmetric algorithm-update noise (which DiD usually absorbs) with treatment-generated, asymmetric spillover (directional co-movement of treated and adjacent pages).