A rigorous diagnostic compares ranking and traffic trajectories of a subset of pages restructured into a hub-and-spoke cluster against a matched control subset deliberately left flat, rather than relying on a simple before-and-after comparison of the whole site. The reason a simple before-and-after isn’t good enough is that Google runs core updates and smaller ranking-system changes continuously, and any of those concurrent changes could produce a shift that gets mistakenly attributed to the architecture change if there’s no control group to isolate it from.
Why naive before-and-after measurement fails here
The intuitive approach, restructure the whole cluster, watch aggregate rankings and traffic before versus after, and attribute the difference to the architecture change, has a fundamental confound: the “after” period isn’t just “the same site with a different architecture.” It’s the same site, with a different architecture, during whatever else was also happening in that time window: seasonal demand shifts, competitor changes, and any of Google’s core or smaller ranking updates that occurred in that same window. Since core updates can move rankings for reasons entirely unrelated to your architecture, a coincidental update landing near your restructuring can easily be mistaken for the effect of the restructuring itself, in either direction. A cluster that happened to get restructured right before a core update that independently favored the site could look like a huge architecture win when it wasn’t one; the reverse is equally possible.
The matched-control approach
The fix is to treat this like a genuine experiment rather than a single-arm rollout. Select a subset of the topic cluster to restructure into hub-and-spoke (treatment group), and hold out a comparable subset of pages, similar in topic breadth, current ranking position, age, and traffic level, that remains in its existing flat structure for the duration of the measurement window (control group). Track both groups’ ranking and traffic trajectories over the same period. If the treatment group moves meaningfully relative to the control group, that difference is a much stronger signal of the architecture’s actual effect, because both groups experienced the same concurrent algorithm changes, seasonal effects, and external factors; only the architecture differed between them.
This is the same logic behind causal-inference methods used more broadly in SEO measurement, such as synthetic control or time-series methods that model what a page’s performance would likely have been absent the change, using a broader unaffected reference set to estimate the counterfactual. Applied here, the “unaffected reference set” is the flat control subset, and the “treatment” is the hub-and-spoke restructuring.
A worked example of how this plays out
Suppose a site has forty articles covering a broad topic area, all currently flat: no pillar page, minimal cross-linking, each page competing on its own for whatever queries it happens to be relevant to. The plan is to restructure twenty of them into a hub-and-spoke cluster: build one pillar page targeting the broad head term, link it out to the twenty spokes, and link the spokes back to the pillar and to each other where topically adjacent. The remaining twenty articles, chosen because they cover a similarly broad sub-area of the same general topic with comparable current rankings and traffic, are left untouched as the control group.
Four weeks after restructuring, suppose the treatment group’s average ranking position improves by some amount and the control group’s average position also improves, though by less. The naive before-and-after read on the treatment group alone would credit the entire improvement to the architecture change. The matched-control read is different: since the control group also improved, at least part of the movement is attributable to something both groups experienced, a core update in that window, a seasonal demand increase, a broader site-wide trust signal maturing, and only the gap between the two groups’ improvement is plausibly attributable to the architecture itself. If the treatment group moved measurably further than the control group over the same window, that differential is the actual estimated effect of the restructuring, not the raw before-and-after number for the treatment group alone. This is the entire point of running both arms simultaneously: it converts an ambiguous single number into a comparison that has a denominator to judge it against.
Edge cases that complicate the comparison
Cross-contamination between treatment and control. If the two groups are topically close enough that they compete for some of the same queries, restructuring the treatment group could indirectly affect the control group’s rankings too, for instance by displacing a control-group page from a SERP position the newly-strengthened pillar now occupies instead. This is a real risk with matched-control designs generally, not unique to SEO, and it means the control group isn’t perfectly “unaffected” if there’s query overlap between the two topic areas. Choosing control pages from a genuinely separate sub-topic, close enough in competitiveness and site-wide context to be comparable, but not so close that they’re direct SERP competitors with the treatment pages, reduces this risk without eliminating it entirely.
Uneven starting momentum. If one group happened to be trending upward or downward independent of anything being tested, perhaps because of an external factor like a recent backlink acquisition or a negative event affecting one sub-topic specifically, that pre-existing trajectory can persist through the measurement window and get misread as an effect of the architecture. This is why baselining both groups over several weeks before the change, not just a single snapshot, matters: a multi-week baseline reveals trend, not just position, and a treatment group already trending upward before restructuring needs that trend accounted for when judging post-change movement.
Partial rollouts within the treatment group. If internal links from pillar to spokes are added gradually rather than all at once, or if some spokes get relinked before others due to implementation sequencing, the treatment group itself isn’t a clean single event, and averaging across it can mask which specific pages actually experienced the architecture change at which point in the measurement window. Logging the exact date each page’s linking changed, not just the date the “project” started, keeps the analysis honest when the rollout wasn’t instantaneous.
Practical construction of the test
Choosing the treatment and control groups. Pick pages from within the same broader site and ideally the same general topic area, so both groups are subject to the same site-wide quality signals and the same competitive landscape. The control group shouldn’t be an unrelated part of the site with a different competitive dynamic, since that would reintroduce a confound of its own.
Matching on starting conditions. Before restructuring anything, baseline both groups’ rankings, impressions, clicks, and average position over a period long enough to smooth out normal day-to-day noise, typically several weeks at minimum. Groups that started from meaningfully different positions or trajectories are harder to compare cleanly after the change.
Isolating the variable. During the measurement window, avoid making other substantive changes (content rewrites, major technical fixes, new backlinks) to either group beyond the architecture change itself in the treatment group. If other changes are unavoidable, apply them symmetrically to both groups where possible, so they don’t become a second confounding variable specific to only one arm.
Measurement window length. Because a restructuring changes crawl patterns and internal link discovery, allow enough time for Google to recrawl and reprocess the new structure before drawing conclusions, typically several weeks to a couple of months depending on site crawl frequency, rather than judging the result from the first few days.
What to avoid
Don’t quantify the benefit using a single site-wide before-and-after number with no control group; that number can’t distinguish the architecture’s effect from concurrent algorithm changes or seasonal shifts. Don’t declare a result from a measurement window too short to outlast normal recrawl and re-ranking noise following a structural change. And don’t assume a positive result generalizes automatically to a different topic cluster on the same site; topical breadth, competitiveness, and existing authority all vary, so the same architecture that helped one cluster isn’t guaranteed to produce an identical effect elsewhere without its own matched test.
Practical implication
Treat hub-and-spoke restructuring as a testable hypothesis, not an assumed best practice to roll out uniformly. A matched-control comparison, ideally analyzed with a counterfactual or time-series method rather than raw before-and-after deltas, is the only way to separate the architecture’s real effect from the noise of Google’s continuous algorithm evolution, and it’s worth the additional setup time before committing to restructuring an entire site’s information architecture based on a single uncontrolled observation.
