Synthetic control and difference-in-differences both solve the same underlying problem: isolating the effect of an SEO change from everything else that’s simultaneously moving rankings, most importantly algorithm updates that affect the whole site or the whole market regardless of what any individual page did. Synthetic control builds a weighted composite of untreated “donor” pages or sites to estimate what a treated page’s trajectory would have looked like without the change. Difference-in-differences compares the change in outcome between a treatment group and a control group across a before/after period, which cancels out trends that both groups share. Both are established econometric methods, not SEO-specific inventions, and both are a meaningful upgrade over a simple before/after comparison, which cannot distinguish an SEO change’s effect from a concurrent algorithm update, seasonality, or market-wide movement.
The mechanism: two distinct methods solving the same confounding problem
A basic before/after comparison of rankings or traffic for a page that received an SEO change is vulnerable to a specific and common failure: attributing a change to the intervention when it was actually caused by something happening to everyone at the same time. If Google rolls out a core update in the same month a site restructures its internal linking, a simple before/after read will conflate the two, crediting or blaming the SEO change for movement that the algorithm update actually caused. Both synthetic control and difference-in-differences exist specifically to strip out that kind of shared, concurrent trend, but they do it in different ways and are worth treating as genuinely separate techniques.
Difference-in-differences works by comparing two groups: a treatment group that received the intervention and a control group that didn’t, both measured before and after the intervention date. The logic is that whatever changed for the control group over that same period reflects the general trend, including algorithm updates, seasonality, or market shifts, that would have also affected the treatment group even without the intervention. Subtracting the control group’s change from the treatment group’s change nets out that shared trend, leaving an estimate that’s attributable to the intervention itself. The key assumption difference-in-differences depends on is that the treatment and control groups would have followed parallel trends in the absence of the intervention, meaning they were moving similarly before treatment and had no reason to diverge afterward except for the treatment. This is a foundational technique taught broadly in applied econometrics, not attributable to a single origin paper the way synthetic control is, because it developed as a standard tool for policy evaluation over an extended period rather than being introduced in one landmark study.
Synthetic control addresses a related but distinct problem: what happens when there isn’t a single good comparison group, but instead many possible untreated units, none of which alone is a convincing match for the treated unit. Instead of picking one control group, synthetic control constructs a weighted composite, a “synthetic” version of the treated unit, built from a donor pool of untreated units, where the weights are chosen so the synthetic composite’s pre-treatment trajectory closely tracks the actual treated unit’s pre-treatment trajectory. Once that weighting is validated against the pre-period, the same weights are applied to the donor pool’s post-treatment outcomes to estimate what the treated unit’s post-treatment trajectory would have been without the intervention. The gap between the treated unit’s actual post-treatment outcome and this synthetic counterfactual is the estimated treatment effect. The method was introduced by Alberto Abadie and Javier Gardeazabal in their 2003 study examining the economic effects of terrorism in the Basque Country, where they built a synthetic version of the Basque Country from a weighted combination of other Spanish regions to estimate what its GDP per capita trajectory would have looked like absent the campaign of terrorism that began in the mid-1970s. That paper is the standard citation for the method’s origin, and Abadie has continued to develop and formalize the technique in subsequent work.
Google’s own data science team published a related, closely connected tool: the CausalImpact R package, released publicly by Google in 2014 and documented in an accompanying paper. CausalImpact uses Bayesian structural time-series models to construct a counterfactual prediction for what a time series would have looked like without an intervention, then compares that counterfactual against the observed post-intervention series to estimate the causal effect. Conceptually it sits in the same family as synthetic control, in that it builds a projected “what would have happened anyway” baseline from control time series and covariates rather than relying on a naive before/after comparison, but it does so through a Bayesian state-space modeling approach rather than the linear weighting scheme used in classic synthetic control. It’s real, it’s published under Google’s name, and it’s a legitimate and citable example of a major search-adjacent company using exactly this class of method to measure the impact of interventions on time-series data, including the kind of traffic and conversion data that SEO measurement deals with.
How an SEO team would actually apply these methods
The practical workflow looks similar across both methods, and the differences are mostly in how the comparison baseline gets constructed.
Select comparable non-treated pages or sites as donors or controls. For difference-in-differences, this means identifying a control group of pages that are similar in nature to the treated pages (similar topic, similar traffic tier, similar historical trend) but that will not receive the SEO change being tested. For synthetic control, this means assembling a donor pool of untreated pages or sites that can be combined, with appropriate weights, into a composite that mimics the treated page’s historical behavior. In both cases, the quality of the eventual estimate depends heavily on how well the comparison group or donor pool actually resembles the treated unit before the intervention. A poorly chosen control group undermines the whole exercise regardless of which method is used.
Establish a pre-period baseline. Before applying the treatment, both methods require a sufficiently long pre-treatment window where the treatment and control (or donor pool) can be compared. For difference-in-differences, this window is used to sanity-check the parallel-trends assumption, that the two groups were moving together before the intervention. For synthetic control, this window is used to actually fit the weights that make the synthetic composite track the treated unit’s real pre-treatment path as closely as possible.
Apply the treatment to only a subset of pages. This is the step that makes the whole exercise possible: rather than rolling an SEO change out to every relevant page or the entire site simultaneously, a team applies it to a defined subset, holding the rest back as the control or donor pool. This is essentially the SEO equivalent of running a proper experiment rather than a natural before/after observation, and it’s the single most important practical requirement for either method to produce a credible estimate. Without a genuine held-back control group, there’s no way to separate the treatment effect from concurrent, sitewide, or market-wide changes.
Estimate the counterfactual and compare it to the observed outcome. Rather than comparing the treated pages’ post-change performance to their own pre-change performance (the naive before/after approach), the team compares the treated pages’ actual post-change performance to what the control group or synthetic composite predicts would have happened without the treatment. The gap between actual and counterfactual is the estimated effect, and because the counterfactual is built from data that experienced the same algorithm updates and market conditions as the treated pages, that estimate is far less likely to be contaminated by confounding events that hit everyone at once.
The main practical caution, consistent across both methods and consistent with how they’re used in economics and policy evaluation generally, is that they estimate a plausible counterfactual, not a guaranteed one. The credibility of the result rests on how well the pre-treatment period was matched and how comparable the donor pool or control group genuinely is. Applied carelessly, with a poorly matched control group or too short a pre-period, either method can produce a confident-looking number that isn’t actually reliable. Applied carefully, they represent a substantially more rigorous way to attribute a ranking or traffic change to a specific SEO intervention than the simple before/after comparisons that dominate typical SEO reporting.
A hypothetical illustration
Consider a hypothetical example: a national appliance retailer, hypothetically called Cascade Home Appliances, rewrites the product descriptions and adds FAQ schema to 400 of its 3,000 product pages, holding the remaining 2,600 back as an untouched control group with similar traffic tiers and product categories. Two weeks after the change, Google rolls out a core update that affects the entire site. A naive before/after comparison on just the 400 treated pages would conflate the FAQ-schema effect with whatever the core update did site-wide. Using difference-in-differences instead, hypothetically, the team compares the change in organic traffic for the treated 400 pages against the change for the untouched 2,600 pages over the same before/after window; if both groups dipped by a similar percentage right after the core update, that shared movement nets out, and whatever gap remains between the two groups’ trajectories is a more defensible estimate of the FAQ-schema change’s actual effect. If, hypothetically, the treated pages recovered faster and ended up 8 percent above the control group’s trend, that gap, not the raw before/after number, is the more credible estimate of what the change actually did.
