Choose the next test from the evidence.

Confirm the action, reporting period and relevant scope before comparing totals. Was the goal changed? Did an import arrive late? Are the newest days less mature than the comparison period? Google documents conversion lag because actions can happen after the interaction and make recent cost or return metrics look worse before the data fills in.

Keep clicks, accepted inquiries, suitable inquiries and paid repairs distinguishable. Compare the same definitions and investigate measurement failures before interpreting a rate change as customer behavior. An aggregate difference may also reflect a different mix of queries, devices or services. It is a finding to understand, not a diagnosis already completed.

A reproducible form failure has a direct acceptance condition: correct it and verify that valid requests can be accepted without creating false success signals. You do not need to leave a known broken path in place merely to call the repair an experiment. Record the fix and retest the affected task.

A question such as ‘Will a clearer estimate-request label improve the task?’ needs a hypothesis and evaluation. Avoid treating every low-rate keyword as irrelevant, every expensive click as bad or every account recommendation as a command. Google’s optimization score can change when recommendations are applied or dismissed. Movement in that score is not a record of additional customers.

Tie the question to the job: helping suitable customers submit a clear estimate request. Define the exact change, the outcome you will examine and any guardrail needed to avoid improving a superficial action while harming the actual result. Preserve the existing version and record the assumptions.

Google’s experiment guidance recommends a clear hypothesis, one variable at a time, chosen success measures before the test and retained records. Have the authorized owner choose a supported method appropriate to the question and account. Keep offer, targeting, bidding, budget and measurement stable where the design requires it. Changing all of them at once makes it difficult to tell which change explains a difference.

Use the planned review method, sufficient relevant observations and appropriate conversion-maturity context. Do not stop only because an early percentage looks favorable. Record any unexpected changes, missing data and uncertainty. A controlled comparison needs its actual design and analysis; a before/after screenshot does not become an experiment merely because only one edit was intended.

Decide whether to keep, revise, investigate further or stop for a supported reason. An inconclusive result can still reveal a measurement gap or a poorly framed question. Save the old/new versions, setup, observations and decision so the next review builds on the evidence. There is no universal number of days, clicks or conversions that guarantees every test has a trustworthy answer.