Can you afford to finish this Google Ads experiment?

Cascader

A Google Ads experiment can run out of money before it runs out of uncertainty. The campaign keeps buying clicks. The comparison keeps saying “maybe.” Eventually, “let it run another week” becomes a subscription nobody remembers approving.

Before committing Search traffic to a test, price the answer you want. Name the change that would justify acting, estimate what each arm can observe, and set two endings: when you stop exposing traffic to the test, and when its delayed outcomes will be ready to judge. An inconclusive result needs a funded exit, too.

The three worked plans make different spending decisions: decline an unaffordable small-effect question; conditionally fund an on-site page experiment for identifiable registered business buyers; decline a sparse-sales bid test and pay for bounded preparation. The middle route needs existing business-ID assignment and outcome logs, fed by an unchanged Search campaign. It is not a native experiment set up from the campaign interface or a design for anonymous visitors. All three plans use invented planning assumptions, not account benchmarks.

Buy an answer to a decision

“Does the new page convert better?” leaves too much room for a technically successful disappointment.

Try: Would we replace the current page if it produced at least 10% more qualified leads per click, without worsening lead value or pushing qualified-lead CPA above our limit?

Suppose the current page produces a qualified lead from 5% of clicks. A 10% relative improvement means 5.5%, a gain of 0.5 percentage points. At 10,000 clicks a month, that is 50 additional qualified leads. It does not mean a ten-percentage-point jump to 15%.

If an additional qualified lead is worth $200 in expected contribution before advertising cost, the potential gain is $10,000 a month at unchanged traffic and cost. That contribution estimate must already allow for close rate and delivery costs. It is not the face value of every enquiry.

Now ask whether the extra contribution, over the period you expect the change to remain useful, justifies implementation, evaluation and exposure to a worse treatment. A benefit worth pursuing can still be too small for this campaign to distinguish affordably.

Keep the outcome tied to that decision. A form submission is not a qualified lead because it arrives sooner. If changing bids could change click prices and traffic mix, a conversion-rate calculation alone cannot answer the CPA or profit question. Plan to evaluate those costs and outcomes together.

A calculator can reject a plan faster than it can approve one

Here is a deliberately simplified screening calculation for the 5% baseline. It treats each exposure as an independent opportunity with one binary outcome: qualified lead or no qualified lead. It uses an approximate two-proportion calculation, a two-sided 5% significance level and 80% power to distinguish the stated improvement from no difference. Penn State’s sample-size lesson explains this approximation and its assumptions.

These are illustrations of how quickly a small question becomes expensive. They are not Google Ads sample requirements. Eighty percent power also leaves a chance of missing the assumed effect; it is not an 80% probability that your particular result will be correct.

Real clicks can include repeat visitors. Conversion credit can be fractional. Lead quality can drift. CPA and revenue have variability this binary calculation does not represent. Google’s experiment statistics use bucketed data and jackknife resampling, rather than this little calculator. Google’s methodology describes the difference.

There is a second trap: detecting an improvement from zero does not establish that it exceeds your business threshold. A convincing 3% lift is still too small if you need 10%. Before accepting a viable-looking plan, check the analysis against the actual assignment unit, repeat exposure, outcome definition and decision threshold. If you cannot support that analysis, the table is a warning light, not permission to launch.

Plan 1: the affordable campaign, the unaffordable question

An operator proposes a small page edit. The business would act on a 10% relative lift in qualified-lead rate. The campaign has 10,000 clicks per 30 days, $3 average CPC and a mature 5% qualified-lead rate. Assume those conditions persist and, for this first screen, treat clicks as independent exposures.

The team can commit 45 days of traffic and $45,000 of total campaign media, then wait 14 days for the qualified-lead cohorts to settle. It reserves another $2,300 for implementation and operator time. A sale begins after that window, so six months of steady conditions is not an available option.

The simple screen asks for about 62,400 clicks to detect that 10% lift from zero: $187,200 of total media and 187 collection days, before the conversion wait. Even that does not promise to establish a lift above 10%.

The team can afford to run the campaign. It cannot afford this experiment’s intended answer within its decision window.

Do not launch the test and quietly replace the desired answer with “we’ll see what happens.” Also do not rewrite the business minimum as 35% merely because that row fits the calendar. A calculator cannot make an unambitious page edit ambitious.

The stopping decision happens before launch: retain the current page, leave the $2,300 unspent, and use the preparation time to find a more consequential, plausible change. If the edit has a separate reason to ship—say, it corrects broken information—make that decision on that reason. Do not present it as an experimentally demonstrated lift.

Plan 2: conditionally fund a page test for identifiable businesses

This is the specialized route signposted at the entrance: an on-site page experiment among already registered business buyers. The team already has stable business-ID assignment and outcome logs. One unchanged Search campaign supplies the visits; ads, bids and audience stay fixed.

The proposed page reduces documented mobile form friction and makes qualification requirements clear before submission. Usability evidence makes a larger improvement plausible. It does not forecast one. The team will fund the test only if it can count the businesses, judge the worthwhile lift and afford the resulting bill.

Count each business once

One business gets one page assignment, one outcome and one analysis row. At its first eligible paid arrival, randomly assign the business 50/50 and record its ID, page and enrollment date. Every colleague, browser and returning visit from that business gets the same page.

The outcome is whether that business makes at least one request meeting the unchanged qualification standard within 14 days of enrollment. Two requests still count as one success. A return by direct or another unpaid route can supply the outcome. A business that never returns, or never qualifies, still stays in the denominator.

For this plan, the qualified-business rate is:

Businesses with at least one qualifying request within 14 days ÷ all enrolled businesses.

The business is choosing on that fourteen-day request outcome, not on lifetime sales. Assume distinct businesses act independently, assignment is preserved and all outcomes are captured. Shared purchasing decisions between businesses, lost identity or missing follow-up break this construction; more clicks cannot repair them. Microsoft’s experiment-design guidance explains why stable assignment units and effects between units matter.

This observable business row is also the limit of the example. It does not reconstruct a native Google Ads cookie experiment: Google says advertisers cannot recreate its user buckets and jackknife analysis from available platform data. The page test can answer a question for these identified buyers, not which bid strategy wins or what anonymous visitors would do.

The information target will become a spending decision. Keep its three prices separate: media across both pages, additional implementation and evaluation cost, and exposure to a potentially worse page.

Decide how much lift earns a yes

Now keep that business denominator through the decision. The baseline is 5% of enrolled businesses making a qualified request, not 5% of clicks. The 10% relative minimum takes that rate to 5.5%: a gain of 0.5 percentage points. The stronger planning hypothesis is 35%, taking it to 6.75%. A bigger hypothesis has to come from the proposed page, not the available budget.

At review, compare the two qualified-business rates:

Treatment’s qualified-business rate ÷ control’s qualified-business rate.

A ratio of 1.10 is the worthwhile boundary. The rate gate passes only if the lower limit of the planned two-sided approximate 95% confidence interval exceeds 1.10. The team selects the independent-binomial log-ratio interval, with business rows as its two samples. Outcome counts too sparse for that approximation cannot support adoption under this rule. The method’s documentation states the independence requirement and distinguishes ratios from differences.

That rule determines the information to buy. The team requires at least an 80% modeled chance of clearing this rate gate at its stronger hypothesis. At 7,000 independent businesses per arm, evaluating the rule over possible outcome counts gives about an 85% chance of clearing the gate if the true rates are 5% and 6.75%. At true rates of 5% and 5.5%—exactly the minimum—it clears the gate in about 2.5% of modeled runs.

Those are author-calculated properties of this constructed design, not observed results or Google’s power. Nor is 85% the chance of adopting the page: cost, value, data quality and safety can still prevent adoption. Smaller positive lifts or fewer independent business rows make this rate gate less likely to clear. The earlier zero-difference screen cannot authorize this plan; it did not evaluate this unit and rule.

Turn the business rows into a bill

The information target is roughly 14,000 businesses, with 7,000 in each arm. The campaign has to buy their arrival, including the repeats along the way.

In this constructed case, 10,000 paid clicks per 30 days supply 8,000 newly enrolled business IDs. The other 2,000 clicks are repeats, including returns by businesses already enrolled. That is 1.25 paid clicks per new business across the planned collection. An ID appearing in another month is still the same enrolled business. Assume this flow, $3 CPC and the outcome rates remain applicable under the planned stopping protocol.

The bill follows the rows:

About 14,000 businesses × 1.25 paid clicks per business = 17,500 paid clicks. At $3 CPC, that is $52,500 of total media.

At the assumed traffic flow, collection takes about 53 days. Then let the last enrolled business finish its 14-day outcome window, plus an assumed two days to reconcile reporting: the expected decision is around day 69. Repeated clicks are on the invoice; they are not extra rows in the analysis.

The $75 CPA limit belongs to this example. For these economics, count one qualified-request credit per successful business, matching the rate numerator. At the baseline, 1.25 clicks at $3 per business divided by a 5% qualification rate is $75 per qualified request. At the planning hypothesis it would be about $56, assuming the same paid-visit cost per business. Assess actual CPA and qualification mix with their uncertainty at review; a point estimate of $74 is not proof that the $75 guard holds. The rate gate does not certify these economics. If they remain unresolved, retain control.

With 8,000 new eligible businesses per month, the minimum lift would add 40 qualified requests. At $200 in expected contribution per request before advertising cost, that is $8,000 a month. Six months of useful life after rollout would mean $48,000 of additional contribution, at unchanged traffic, cost and value. That payoff horizon is an assumption, not guaranteed earnings; the $200 estimate already allows for close rate and delivery costs.

Three prices, one ending

The $52,500 expected media is spend across both pages. If this campaign would otherwise run, it is not a $52,500 extra research invoice. The $2,300 implementation and evaluation cost is additional. Treatment exposure is the business risk: at the full $60,000 ceiling, roughly $30,000 could run through the changed page if spend divides evenly. Random assignment does not promise equal realized spend.

Stress that risk. Under the same $3 CPC and 1.25-click cost model, $30,000 represents about 8,000 treatment businesses. If their qualification rate fell from 5% to 4%, that would mean about 80 fewer qualified requests than baseline, or $16,000 of foregone contribution at $200 each. That is a hypothetical downside, not a measured loss or a worst-case bound. The team must tolerate that exposure and name harms it will stop for sooner. A broken form does not need a significance badge. The $90 CPA safety stop is a deliberately chosen business loss limit, not proof that the page caused the loss. It can end a potentially useful test; that is the accepted cost of protecting the business.

At the planned final review:

• Adopt only if the rate-ratio interval clears 1.10 and the CPA, expected value and data-quality guardrails hold.
• Keep control if the interval rules out a worthwhile improvement, shows harm, or the economics fail.
• Keep control and record inconclusive if the interval spans the 1.10 boundary or a required guard remains unresolved.

If a cap arrives before both arms have 7,000 businesses, use the actual counts at the planned review and expect less information; do not borrow the 85% calculation. Do not discard early or late businesses to make the result look better. A safety or measurement failure ends this protocol without an adoption finding from its planned rule.

Reaching the collection limit ends paid test traffic even when the report is undecided. Keep the assigned pages available for enrolled businesses’ unpaid returns and finish their outcome windows; preserve attribution and reporting. If the two-day reporting allowance does not produce complete records, keep control and resolve the missing outcomes without buying more test traffic. The 5% and 6.75% hypotheses refer to outcomes under this bounded traffic-and-follow-up protocol, not to an indefinitely running campaign. Waiting for outcomes already purchased is different from buying more clicks.

This plan can finish without a winner. It has priced a supported chance of answering a particular page question, plus the decision if that answer does not arrive. A native bidding experiment needs its own design and evidence; this one cannot lend it a power calculation.

Plan 3: six sales a month cannot settle a tiny sales-rate question on demand

A different campaign produces 3,000 clicks per 30 days at $8 CPC. Its mature sale rate is 0.2%: about six sales a month. Sales and their imported values take roughly 60 days to settle. The business wants to know whether a bid change raises sale rate by 25%, from 0.2% to 0.25%, while maintaining profit.

Even the independent binary screen asks for roughly 282,000 total clicks to detect that difference from zero. At this assumed rate of traffic, that is about 94 months and $2.25 million of media, before the sales wait. No claim about an eight-year-stable auction market is hiding in that arithmetic. Its purpose is to reject a 30-day proposal.

The proposed commitment was 50/50 traffic for 30 days, $24,000 total media and a final sales review around day 90. Each arm would expect roughly three baseline sales from that month’s clicks. A single large sale could dominate the value comparison. The simple rate calculation does not even price that value variability.

Decision: do not launch this experiment. Test allocation is zero; the ordinary campaign budget remains a separate business decision. No bid-test winner will be declared at day 30 or day 90.

Instead, spend a bounded preparation budget: eight operator hours at $100 plus $1,200 of technical help, $2,000 total over 14 days, to reconcile sale imports, repeat leads and which earlier qualification events predict profitable sales in already mature cohorts. Stop that piece of work after the agreed time and report what remains unresolved; do not turn it into another unlimited subscription.

That work may establish a more frequent outcome worth testing later. If it does, write a new plan with its own allocation, costs, maturity window and stopping decision. “More demos” can be a legitimate demo question. It cannot smuggle in an answer about profit while the sales are still on their way.

Make the split and the clock match the question

For a native Google Ads custom experiment, Google recommends a 50% split. Cookie-based assignment keeps a user in one version; search-based assignment can show the same person both versions across repeated searches. Choose with the return journey in mind, rather than choosing the option that seems likely to produce a result fastest. Google’s custom-experiment setup explains both options.

A 50/50 setting does not promise equal realized clicks, spend or conversions. Treat the arithmetic above as a starting estimate, then check both arms’ actual delivery. Moving to 80/20 to protect the control also makes the treatment slower to learn from. In the idealized equal-variance case, that imbalance needs about 56% more total observations for the same precision than 50/50. Price that delay before choosing it; do not halve the treatment and leave the deadline untouched.

A budget cap matters in another way: splitting an already constrained campaign can change what each arm gets to observe. Forecast from eligible traffic and plausible arm-level spend, not from the account’s total clicks. Plan for different CPCs, low treatment delivery and a less favorable baseline. More budget is useful only if it can buy relevant observations in time.

Finally, count reporting delay as part of the decision’s calendar. Google defines the conversion cycle to include the time needed for imported conversions to be reported, and advises evaluating mature periods. Its bidding guidance explains why recent clicks can still be awaiting outcomes. The 14- and 60-day waits in these plans are assumptions to replace with your own cohort evidence, not rules for Search experiments.

If a bidding change requires a settling period, pay for that period too. Specify whether you are deciding on performance from launch or on later steady operation; do not choose the exclusion after seeing which dates flatter the treatment. Include all paid days in the cost ceiling, even when some are excluded from the planned comparison.

Write the ending while you still like both versions

Before launch, put one short agreement beside the experiment:

We are deciding whether to [change], using [outcome and analysis]. The smallest improvement worth acting on is [amount], with [cost/value guardrails]. We will assign [split and unit], commit at most [total media, treatment exposure, build and operator time], and collect through [date or limit]. We expect outcomes to be judgeable by [date and maturity criterion]. We stop early for [specific harm or failure]. If the planned review is inconclusive, we [retain control or another agreed action].

Monitor delivery and harm while the test runs. Reserve winner selection for the planned review; repeated looks are not a reason to announce the first flattering ordinary confidence interval. If you need early statistical winner decisions, plan an analysis designed for that use before launch.

An extension is a new spending decision. It needs a fresh estimate of what the additional traffic can resolve and why that answer is still worth having. “We’ve already spent this much” supplies neither.

The most valuable preflight result may be a test you never start. You have kept the campaign’s money, the operator’s attention, and the right to ask a better question.