Conversion rate optimization, from evidence.

• Choose conversion rate optimization when visitors with intent do not act.

  Key events are reliable, the traffic arrives, and the open question is why intent turns into exits. That question is what this program answers.

• Choose analytics implementation first when the numbers are in doubt.

  Tests judged on key events that are missing, duplicated or unreconciled produce confident wrong answers, so measurement is fixed before any experiment.

• Choose UI/UX design when the whole journey needs rethinking.

  When research shows the problem is the structure of the product or site rather than specific points of friction, redesigning the experience is the honest answer.

A conversion hypothesis is testable when it names one change, the visitors it applies to, the effect expected and the single metric that will decide it, and when that metric is already measured reliably. A hypothesis that cannot fail, such as a new design will perform better, gives a test nothing to decide.

• The evidence: what analytics, recordings or usability review showed.
• The change: one specific difference between the control and the variant.
• The audience: which visitors see the test, such as mobile visitors from paid search.
• The expected effect: its direction and the smallest change worth detecting.
• The primary metric: one key event or step completion that decides the result.
• Guardrail metrics: what must not get worse, such as lead quality or refunds.
• The stopping rule: the sample size or duration agreed before launch.

An A/B test result is judged at the end agreed before launch, on the primary metric fixed in the hypothesis, after checking that visitors were split in the planned proportions and that guardrail metrics did not worsen. Stopping a test the moment one variant looks ahead, or searching segments for a win afterwards, turns random variation into false conclusions.

The split check matters more than it sounds. A sample ratio mismatch is a statistically significant difference between the ratio of users counted in each variant and the ratio configured before the experiment began, and the missing users are often the ones most affected by what was tested. A test with a mismatch is investigated, not read.

The sample size is calculated before launch from the current conversion rate and the smallest effect worth detecting, and the test runs in complete weekly cycles so weekday and weekend visitors are both included. When the planned end arrives without a clear difference, the honest result is that the change made no detectable difference.

A/B testing is compatible with Google Search when it is set up correctly. That means using rel=canonical on alternate test URLs, a 302 temporary redirect rather than a 301 for redirect tests, never showing Googlebot one set of URLs and people another, and removing test elements as soon as a test ends.

Showing one set of URLs to Googlebot and another to people is cloaking, which Google Search treats as spam, so tests never target or exclude search engine crawlers. Small changes, such as the size, color or placement of a button or image, or the text of a call to action, often have little or no impact on the search result snippet or ranking of a page.

Larger tests that change the content or structure of an indexable page are planned together with the SEO work, because a variant that removes the content a page ranks for can cost more than the test could gain.

Conversion changes are prioritized by how many visitors reach the page or step, how strong the evidence behind the hypothesis is, and how much effort the change takes to build and test. A small fix on a step every buyer passes through usually outranks an ambitious redesign of a page few visitors see.

Scoring frameworks help make that ranking consistent, but they are aids to judgment rather than measurements. Evidence from several independent sources, such as a funnel step that loses visitors and that recordings and a usability review both explain, ranks above a single observation or an opinion.

Some findings are not tests at all. A broken form field, a missing price or an error message nobody can read is fixed immediately, because testing whether a defect matters wastes traffic that a real test needs.

• Conversion rate optimization
• Hypothesis backlog and prioritization
• Experiment design
• Conversion analysis
• Result interpretation and rollout

Every change in the program passes through the same four stages. A change that cannot be traced back to evidence, or forward to a recorded decision, is not part of the program.

• Effort aimed at the real loss

  Research locates the step and the visitors where intent is lost, so changes go where they can matter rather than where opinion is loudest.

• Decisions that survive scrutiny

  A hypothesis, a fixed metric and a stopping rule written down before a test starts make the result hard to argue with afterwards.

• Knowledge that accumulates

  Losing and inconclusive tests are reported alongside winners, so the team learns what does not move its visitors and stops repeating it.

• Experiments that respect search

  Tests use rel=canonical on variant URLs and temporary redirects, so an experiment is not mistaken for cloaking or a permanent move.

• A lead form visitors start and abandon

  Form start and submission events show the gap, recordings show where visitors hesitate, and the form is simplified and measured against its baseline.

• A paid campaign landing page

  Message match between the advertisement and the page is tested first, because a visitor who does not recognize the promise leaves before reading the offer.

• A checkout that loses buyers

  Funnel steps from cart to purchase locate the loss, and changes to costs, fields or payment options are tested one at a time.

• A test starts only when its primary metric, guardrail metrics, audience and sample size are written down, and it is not stopped early because one variant looks ahead.
• The split between variants is checked before a result is read, because a split that differs significantly from the planned ratio is a common cause of incorrect conclusions.
• Experiments are set up for Google Search: rel=canonical on variant URLs, 302 redirects for redirect tests, no cloaking, and test elements removed as soon as the test ends.
• Testing scripts are chosen and loaded so they do not delay the main content or shift the layout, and pages are measured before and after a test.
• No dark patterns: manufactured urgency, disguised opt-ins and obstructed cancellation are excluded, whatever a test might appear to show.
• Results are reported against the baseline for the whole audience and period tested, including tests that lost or found no difference.