Research paper · Experiments
A/B Tests on Casino Landing and Sign-Up Pages
A good test on a casino or sportsbook page answers one question, keeps Google seeing the same page as users, checks that traffic split the way it should, and never tests a variant that weakens a required check.
Keep search engines out of the experiment
Google's guidance on testing sets four rules that matter for organic landing pages:
- No cloaking. "Don't show one set of URLs to Googlebot, and a different set to humans." Showing Googlebot something different from users breaks the spam policies, even during a test.
- Canonical links. Google recommends "using rel="canonical" rather than a noindex meta tag" on variant URLs, so search engines group them with the original.
- Temporary redirects. If a test redirects users to a variant URL, "use a 302 (temporary) redirect, not a 301 (permanent) redirect."
- Short tests. When the test ends, "remove all elements of the test as soon as possible."
Google Optimize, once a common free testing tool, has been "no longer available as of September 30, 2023," so most teams now use a third-party tool or server-side testing. The rules above apply whichever tool you use.
Check the split before reading results
A sample ratio mismatch is, in the words of Microsoft's experimentation team, "a statistically significant difference between the ratio of users counted in variants A and B" compared with the split you set. Their analysis found "about 6% of A/B tests have an SRM," and they advise: "Don't trust the results of A/B tests with an SRM until you diagnose the root cause."
On casino and sportsbook sites, common causes include a consent banner that loads differently in one variant, a redirect that drops some users, bots that reach only one URL and market routing that sends some visitors elsewhere. Run a chi-square test on the user counts in each variant before you look at any outcome.
Decide the plan before the test starts
| Item | Example for a registration page test |
|---|---|
| Question | Does showing accepted ID documents before the form change completed registrations? |
| Primary metric | Registrations per landing visit, counted on the server |
| Guardrail metrics | Verification pass rate, complaints, help page visits |
| Market and pages | One licensed market, one page group |
| Sample and end date | Calculated in advance from the smallest change worth detecting |
| Stopping rule | Run to the end date unless a guardrail breaks |
Stopping as soon as a result crosses a significance line inflates false positives, because random variation crosses the line from time to time. Fix the end date, or use a sequential method designed for repeated looks.
What not to test
Some variants should never reach a test, because the "winning" result would be a breach:
- Removing or delaying age, identity or market checks.
- Showing offers in markets where public inducement advertising is prohibited.
- Countdown timers, false scarcity or wording that suggests gambling is without risk.
- Pre-ticked marketing consent boxes.
Treat every variant as live marketing. It needs the same compliance review as the page it replaces, using the rules summarized in matching landing pages to search intent.
Reporting results honestly
Report the effect with its confidence interval, the sample, the dates and any SRM check. Say what was not measured, such as long-term retention. If the result is flat, publish that too: a flat result on a well-run test is useful knowledge. The guide to publishing funnel numbers sets out what each published figure should carry.
Limits of this paper
This paper gives design principles, not a statistics course. It uses no operator test results, and the prevalence figure comes from Microsoft's own experiments.
Questions
Can a variant page be noindexed during a test?
Google recommends rel="canonical" pointing to the original rather than noindex, so the variant is grouped with the original page.
How long should a test run?
Long enough to reach the sample you calculated in advance, and at least one full weekly cycle, since casino and sportsbook traffic changes by day of week.
What should we do when an SRM appears?
Stop reading the outcome, find the cause, fix it and restart the test. Results from a broken split are not reliable.
Sources
Each source was read on October 27, 2026. How we choose sources is set out on the sources and reading list page.
- Google Search Central, Minimize A/B testing impact in Google Search, last updated December 10, 2025.
- Microsoft Research, Experimentation Platform, Diagnosing Sample Ratio Mismatch in A/B Testing, September 14, 2020.
- Google Analytics Help, [Sunset September 2023] Google Optimize, no date shown.