KatvTech started as a conversation between a group of friends trying to improve search rankings on several websites we were building. We read the same advice everyone reads. We applied it. Sometimes it worked. Sometimes it didn't. And we could never tell why, because nobody was testing anything properly. With backgrounds in engineering and data analysis, we decided to approach SEO the way we'd approach any other testable system: form a hypothesis, control the variables, measure the outcome.

Most SEO advice is based on what appears to have worked, for someone, at some point. That is not evidence. It is pattern-matching dressed up as expertise. At KatvTech, we run structured experiments because we want to know not just what happened on a particular site, but why it happened and whether the same approach would work on a different site, in a different niche, at a different point in Google's development.

This page documents our methodology. Every experiment article on this site links back here because understanding how we test is as important as understanding what we found.

---

Why does SEO need controlled experiments?

SEO advice has a credibility problem. When a site improves its ranking after publishing more content, adding schema markup, improving page speed, and changing its internal linking structure all in the same month, there is no way to know which change drove the improvement. Controlled experiments isolate variables so causes and effects become separable.

The majority of SEO guidance circulating in 2026 is based on correlation studies, case studies, or practitioner intuition. None of these are worthless. Pattern recognition across thousands of sites reveals real signals. But correlation has a ceiling. A correlation study can tell you that faster sites tend to rank higher. It cannot tell you whether speed caused the ranking gain, or whether well-resourced sites that invest in speed also happen to invest in content quality, link building, and author credentials at the same time.

Google's algorithm evaluates hundreds of signals in combination. When you change five things at once and your rankings improve, you have learned very little about SEO. You have learned that you, on your specific site, in your specific niche, at this particular point in Google's development, saw an improvement after making five changes.

So we started changing one variable at a time. We spent long nights designing each test, arguing over methodology, and waiting out 30 and 60-day measurement windows before drawing any conclusions. We have not found all the answers. We are still running tests. But after hundreds of hours of structured experimentation, we decided to publish what we have found so far rather than keep it internal — because the SEO community publishes opinions freely and data rarely.

Controlled experiments replace that ambiguity with specificity. They do not eliminate uncertainty. SEO is not a physics lab. But they narrow the range of plausible explanations considerably.

---

How does KatvTech select experiment variables?

We select variables based on three criteria. The variable must be isolable, meaning we can change it on some pages without touching other ranking factors. It must be measurable through tools available to any site owner, primarily Google Search Console. And it must be practically actionable, something a content site builder can implement without specialist resources or significant budget.

Our experiment pipeline starts with a shortlist of variables that meet all three criteria. Current active areas include:

• Schema markup types (FAQPage, Article, HowTo, BreadcrumbList)
• Heading structure and format (question-based H2s versus descriptive H2s)
• Content freshness and update patterns
• Internal linking architecture (pillar and cluster versus flat structure)
• Core Web Vitals metrics (LCP, INP, and CLS measured individually and as a composite score)
• Author bio completeness and credential specificity

We deprioritise variables that require significant off-site activity. Link building campaigns, for example, are difficult to control for during a 60-day experiment window. Our experiments focus on on-page and technical factors where we have direct control over the change being made.

Variables are added to the experiment queue based on two inputs: current Google documentation and confirmed update behaviour. When a Google core update demonstrably changes ranking patterns in a specific direction, the factors associated with winning pages become candidate experiment variables.

---

What does our control group setup look like?

For each experiment, we identify matched groups of pages or matched domains. The test group receives the change being evaluated. The control group receives no change. Before any experiment begins, we document baseline metrics for both groups across at least 30 days to establish a stable starting point before anything is touched.

Matching criteria vary depending on whether the experiment runs at the page level or the domain level.

For page-level experiments (testing heading structure, schema markup, content updates), we look for pages that share:

• Similar word counts, within 20% of each other
• Similar backlink profiles, between 0 and 3 referring domains in both groups
• Similar average position in the 30 days before the experiment begins, within 5 positions
• Published within a similar timeframe, both groups within the same 60-day window
• Same category or topic cluster on the same domain

For domain-level experiments (testing pillar architecture versus flat structure, or comparing schema-enabled domains against non-schema domains), we use dedicated test domains registered specifically for the experiment. These are not monetisation assets. They exist solely to provide isolated, controlled environments where we can introduce a single change and measure the outcome without interference from other variables.

The minimum test group size for most experiments is 10 pages. Smaller sample sizes produce results that are too easily explained by normal variation in Google's crawl behaviour.

---

How do we measure ranking outcomes?

Google Search Console is our primary measurement tool for all experiments. We track impressions, clicks, and average position for each test page across the experiment period. For AEO-specific experiments, we supplement GSC data with manual checks for AI Overview appearances, People Also Ask inclusions, and featured snippet captures, since Google Search Console does not separate these from standard organic traffic.

We use GSC as our primary source for three reasons. It is first-party data provided directly by Google, which means it reflects how Google sees our pages rather than how a third-party crawler estimates that Google sees our pages. It is free and available to any site owner, which means our methodology is fully reproducible without paid tools. And it provides the specific metrics that matter most for search visibility: impressions, clicks, and average position.

Supplementary tools we use across different experiment types:

• Ahrefs Free: For domain rating tracking across test domains
• PageSpeed Insights: For Core Web Vitals measurement in CWV-specific experiments
• Google Rich Results Test: For schema verification before and after schema experiments
• Manual search queries: For AI Overview and People Also Ask tracking

For each experiment, we define the measurement period before making any changes. Most experiments run for 30 days at minimum. Experiments testing structural changes to site architecture run for 90 days because architectural effects accumulate gradually rather than appearing overnight.

We establish a pre-experiment baseline by pulling the prior 30 days of GSC data for every page in both groups before any changes are introduced. This baseline makes percentage changes comparable across pages with different absolute traffic volumes and different starting positions.

---

What are the limitations of our experiments?

Every experiment on this site has limitations, and we document them explicitly in each article. SEO is not a controlled laboratory environment. Google's algorithm updates mid-experiment, our sample sizes are small relative to what would be required for formal statistical significance, and results in one niche may not translate directly to another.

The specific limitations that apply to most of our work:

Sample size. We typically run experiments on groups of 10 to 20 pages, or on 2 to 4 domains. This is enough to identify directional patterns. It is not enough to claim statistical significance in any formal sense. We describe our results as indicative rather than definitive.

Niche specificity. An experiment run on a home improvement site may produce different results on a finance site or a travel site. Where we know the niche of our test site, we state it clearly. Where we suspect niche characteristics may influence the results, we flag that as a variable we could not fully control.

Algorithm interference. Google releases core updates, spam updates, and product changes continuously. A major update midway through a 90-day experiment can alter ranking patterns in ways that affect our results. When a confirmed update occurs during an active experiment, we document it as a limitation and assess whether the experiment should be extended, restarted, or published with adjusted conclusions.

Correlation caveat. We are measuring the output of Google's algorithm, not its internal signals. We cannot confirm that our measured variable caused a ranking change. We can confirm that the ranking change followed the variable change, and that the control group did not experience the same change. That is meaningful evidence of a relationship. It is not proof of causation.

Our results reflect our specific test conditions. We document those conditions in detail so you can evaluate how closely they match your own site before deciding whether to act on a finding.

---

How often do we publish new results?

We publish 2 to 3 articles per month, a mix of new experiment results, updates to ongoing long-run experiments, and analysis of confirmed Google algorithm updates. Experiments requiring 60 or 90 days of data include a preliminary update at the 30-day mark so readers following the research in progress can track developments without waiting for the final publication.

The Experiments Index page maintains a live list of every experiment we have published, with the variable tested, the result direction (positive, negative, or inconclusive), and the date of publication. Long-run experiments currently in progress are listed with their start date and expected publication date.

Experiments are not invalidated by time. A schema experiment published six months ago remains relevant as long as Google's treatment of schema markup has not changed materially. When confirmed algorithm updates alter the relevance of a past experiment, we add a note to the original article flagging the change and explaining what has shifted.

If you have a specific variable you want us to test, use the contact page. We prioritise suggestions that come with a clear hypothesis and a practical application.

---