Confounding Variable

The word has a precise history. VanderWeele and Shpitser write that the causal inference literature has a clear formal definition of confounding, expressed in terms of counterfactual independence, but no consensus on a formal definition of a confounder. They propose that a confounder be a covariate that exists before the exposure and is needed, with other covariates, to make the effect of the exposure on the outcome unconfounded^[1]. In plain terms, it is something that moves with the change and with the result, so that the comparison is not fair.

In search the usual suspects are well documented. Google’s guide to debugging traffic drops lists the main causes as algorithmic updates, technical issues, seasonality and changing interests, and site moves and migrations. It says core updates and other smaller updates may change how some pages perform, and points to its list of ranking updates^[2]. Any of these can move a page’s clicks in the weeks around a change of yours, and none is a result of that change.

The list is dated. Google’s Search Status Dashboard shows each ranking update with its start and its duration: the March 2026 core update began on 27 March 2026 and ran for 12 days and 4 hours, and the May 2026 core update began on 21 May 2026 and ran for 11 days and 21 hours^[3]. A change shipped inside either window shares its weeks with an update.

Randomising is how experiments get round the problem. Kohavi and colleagues write that controlled experiments neutralize confounding variables by distributing them equally over all values through random assignment. They also advise running an experiment for at least a week or two, then in whole weeks, and say the lesson applies to other time-related events such as holidays and seasons^[4].

A gap between two groups is not an effect when the groups differ in other ways. Of the sites QueryRev’s crawler asked for an llms.txt, 15.1% of the 5,168 that block no AI crawler by name had one, against 8.1% of the 862 that block one by name^[5]. That does not show that blocking keeps a site from publishing one. The same study shows that blocking varies with the place in the ranking, 17.6% of the sites read in places 1 to 1,000 against 13.9% of the rest read, and with the country-code ending, from 44.6% on .it to 3.7% on .cn^[5]. The two groups differ in many ways at once.

List what else happened. For the window of the comparison, write down every event that could move the metric: Google ranking updates, seasons and holidays, promotions, releases, migrations and outages. Google’s guide suggests a date range of 16 months, and a comparison with the same period a year earlier, to see whether a movement happens every year^[2].

Ask of each event which pages it reached. An event that reaches the changed pages and a comparison group alike, such as a core update, is cancelled out by the comparison. An event that reaches just the changed pages, such as a promotion on those products or a template release, is a confounder, and the comparison does not remove it.

Then adjust or narrow. Compare with pages the event did not touch, compare with the same period a year earlier, or leave out the pages that an event reached on one side. Report what was left unadjusted, and give the effect as a range.

Illustrative: an invented site, to show the method.

A coffee-equipment shop rewrites the title of /burr-grinder, and in the next two weeks its clicks are 18% above the two weeks before. In the same fortnight a Google core update rolls out, the shop starts a spring promotion on grinders, and a site-wide speed release goes live. The 18% has four possible causes and one comparison.

The shop cannot split the 18% four ways. It can compare the page with similar grinder pages whose titles it left alone: the update, the promotion and the release reached those pages too. What is left is the gap between the two, a modeled range. The shop reports it as that, and names the other three events as part of the same window.