How to Close a Google Ads Experiment When Nobody Won

Friday arrives. The landing-page experiment reaches its scheduled end. The team opens the results expecting a winner, as if the end date were a finish-line camera.

The new page has more qualified leads. It also has a result uncertain enough to include a useful gain and a loss you would not have agreed to take. The calendar did its job; the evidence has not done the job the team hoped for. Someone will still have to decide what happens to Monday's traffic.

Here is one experiment notebook, from the sentence that started the test to the note that closes it. The campaign, people and numbers are illustrative. The distinctions are the real work.

Entry 1: What decision are we buying?

Question. Should we replace the current Search landing page with a shorter form?

Reason to try. The current form asks for a company size and a phone number before showing the demo calendar. The team suspects those fields cost qualified leads. Sales suspects removing them will admit more people they cannot help.

Decision written before launch. Move the new page into the campaign only if the observed relative lift in qualified-lead rate reaches 10% and the lower end of its 95% uncertainty range is better than a 3% loss. If either test fails, keep the current page while we decide whether another experiment is worth its traffic.

Primary measure. Sales-qualified leads divided by landing-page visits from the experiment. Check that the qualification rule and reporting lag are the same for both arms.

The 10% and 3% are this team's trade-off, not universal testing thresholds. The new page would cost design, QA and sales time to adopt, so they want a sizable observed gain and a reported range that no longer includes the loss they cannot stomach. They accept that the real gain might be smaller than 10%; the observed lift is a decision hurdle, not proof of a 10% effect. Another team with a costly incumbent page or a cheap, reversible change could set a different bar.

Writing the bar first prevents a familiar Friday ritual: choosing whichever metric wears the nicest green arrow, then announcing that it was the objective all along. Google recommends choosing success metrics before a test and changing one variable at a time for a reason: a result only answers the comparison you actually made. Google Ads' experiment guidance makes both points.

Entry 2: The date arrived; the verdict did not

Four weeks later, each page has received about 14,000 visits. The current page has 480 qualified leads; the shorter page has 512. That is 3.43% versus 3.66%, an observed lift of about 6.7%.

An impatient slide would put +6.7% in large type and the word winner in even larger type. The notebook puts the uncertainty next to it. For these illustrative counts, a simple approximate 95% interval for the relative difference runs from about 6% worse to 21% better. It is an illustration of the spread, not a Google Ads result or a promise about a future campaign. Google's own experiment reports use their own method to calculate intervals and significance; use the actual report for an actual test. Google explains that method here.

The result fails both parts of the prewritten rule: the observed 6.7% lift is below 10%, and the range's lower end, near 6% worse, crosses the 3% loss limit. Its upper end reaching 21% does not clear either hurdle. A worthwhile gain remains possible, as does a loss the team would not take. The observed lead is a description of this run. It is not permission to replace the page.

Nor is the result a tie. “We didn't establish a difference” and “the pages perform the same” are different claims. The first describes what this test can support. The second asks the evidence to rule out differences, including ones the team cares about. These counts do not do that.

Entry 3: Could we trust the comparison at all?

Before calling the result inconclusive, the operator checks the comparison. The team had planned an even traffic split. The operator inspected the saved assignment setting and arm-level delivery by week and device: the setting stayed in place, with no unexplained shift, outage or reroute. Roughly equal total visits alone would not establish that. Both final URLs loaded, both arms used the same definition of a qualified lead, the offline-qualification lag had caught up, and the change log shows no consequential change unique to one arm. The shorter form was the planned difference. There is no identified break that makes these numbers unusable.

If assignment had shifted mid-test, one page's conversion import had failed, or a base-campaign change had confounded the split, the closing label would be comparison broken or compromised, with the fault and affected period named. Waiting longer would not repair it. Google warns that base-campaign changes can obscure the cause; unequal impressions or spend alone call for inspection, not an automatic failed-test verdict. Google's setup guidance and experiment guidance describe these limits.

An inconclusive performance result might still be good enough for another decision: if the possible downside is tolerable, an inexpensive, easily reversed change could be worth making for accessibility or maintenance while the team says plainly that the test found no performance winner. That is not this costly page replacement. Friday's comparison survived its checks, but its result cannot clear the team's bar.

Entry 4: Close the work, not the uncertainty

Could another fortnight change the call? The next landing-page decision is due at a refresh in two weeks. At the test's mature pace of about 250 qualified leads a week across both pages, two more weeks might ultimately yield roughly 500 additional qualified leads; some could still await sales qualification at the refresh. If the pages kept performing as they have, the observed lift would remain 6.7%, below the team's 10% hurdle, even as the range narrowed. Different outcomes could move both the lift and the range.

At the same split and visit pace, extending would also send about 7,000 more visits to the shorter page while its material downside remains in range, using the last pre-refresh testing window on this question. The team chooses to stop now.

The closing note goes to the people who expected the verdict. It is short enough to read in a meeting and complete enough to prevent the same argument next month:

Decision: Keep the current landing page and end the split Friday. Do not apply the shorter page to the base campaign.

What we observed: The shorter page produced 512 qualified leads versus 480 from roughly equal visits: 6.7% observed lift, with an illustrative range from about 6% worse to 21% better. The lift missed 10%; the lower end crossed our 3% loss limit. Inconclusive, not equal.

What we checked: The planned split stayed configured, with no unexplained arm-level delivery shift or reroute. Both pages and qualification paths worked; the lag matured; no one-sided change was found. Save the report, dates, change log and page versions.

Why stop now / traffic: At the mature pace, two more weeks might ultimately yield roughly 500 qualified leads; fewer could be confirmed by the refresh. At unchanged rates, 6.7% still misses the bar; extending also sends about 7,000 more visits to the uncertain shorter page. Return traffic to the current page, archive the shorter page as a candidate, not a “loser,” and check auto-apply settings.

Next question: Was the phone field, company-size field, or extra step the costly part? Test one change only if available traffic could resolve that decision; otherwise use the slot elsewhere.

The incumbent keeps the traffic because the proposed replacement did not clear the team's bar, not because the experiment proved the incumbent superior. Another team with more time, more traffic or a different decision cost might pay for another run. This one has closed its traffic decision without closing the question.

The notebook ends with a period after keep the current page and a question mark after which form change, if any, deserves another test. That is a finished experiment. It is also an honest piece of unfinished knowledge.