A paid-ads experiment needs a stop rule before it needs a budget: how to pre-register the economics of a B2B paid test
Write down your stop rule, economics, and evidence bar before spending a dollar.

A B2B paid-ads experiment needs a rule for when to stop before it needs a budget. For technical founders and growth engineers, the task is to pre-register the economics, the required evidence, and the conditions for ending the test before spending begins. That document turns an open-ended bet on clicks into an experiment that can answer whether the acquisition model deserves more capital.
Why B2B paid tests fail before the budget runs out
A founder stares at two weeks of campaign data and still cannot make a clear call. One number looks promising, another looks weak, and neither settles whether to continue. Before the data arrived, nobody wrote down what evidence would be enough to scale, stop, or wait.
Without that agreement, each daily check is a chance to change the experiment. A promising result invites an early declaration of victory, while a discouraging result invites a premature shutdown. Declaring a winner from repeated peeks inflates the false-positive rate beyond the tolerance planned for a fixed-sample test. It is not the same experiment as waiting for the committed sample.
Teams readily describe success, but what would show that the campaign failed, and when would they accept that verdict? When that question stays open, hope and fear count as much as evidence. More budget cannot repair a test whose decision criteria remain negotiable.
What rising B2B CPCs make an undisciplined test cost in practice
B2B SaaS keyword prices have risen sharply while paid search’s reach has narrowed. An extra week of ambiguous data collection therefore costs more than it did two years ago. Delay has a direct economic price. Yet ending the test too soon can be just as misleading as extending it without a limit.
Take paid search with a six-month conversion cycle. Paid search may generate opportunities within weeks, while revenue from those opportunities takes quarters. Leaders who expect closed deals from a six-week test can cut the channel before those opportunities turn into deals. But even with a long sales cycle, spending cannot go on forever without evidence that it is getting closer to bringing in contribution-positive customers.
What is the minimum credible experiment, given the smallest worthwhile improvement, the required sample, and the attribution delay? Assessing whether the acquisition model is worth putting money into starts with a document defining the target contribution.
What a pre-registration document for a paid test must contain
A pre-registration is a one-page specification for decisions that become easy to bend after results arrive. To draft it, begin with one business question. For a B2B SaaS campaign, that might be whether the message should lead with speed or risk reduction. Record the current approved experience as the control, then define the eligible population and exclusions by geography, segment, and prior-customer status.
Next, name one primary metric that is tied to qualified-pipeline rate or sales-qualified lead rate, or SQL rate. Click-through rate and form fills are not enough. Specify the guardrails that can block rollout, including lead rejection, refund rate, and tracking corruption. These fields separate a promising response from an acceptable business outcome.
The statistical fields turn that question into a scoped test. Record the historical baseline rate, the minimum detectable effect, or MDE, the false-positive tolerance called alpha, and the desired statistical power. The MDE is the smallest absolute difference worth detecting. Calculate the required eligible units in each arm, commit the allocation, and state the attribution delay.
Finally, include the stopping rule and economic limits. Version-control and timestamp the document before launch. Any later departure needs transparent disclosure, because silently revising the plan after seeing data defeats its purpose.
How to write the stopping rule — the field most teams leave blank
The stopping rule must require sample, duration, and attribution delay together. For a fixed-sample test, end exposure only after every arm reaches its planned eligible sample and the declared minimum business cycles have elapsed. The sample comes from the MDE calculation. It is not a round number chosen to fit a budget meeting.
For B2B SaaS, declare a minimum duration of four to six weeks, even if traffic arrives sooner, because early arrivals need time to convert. Then wait the full conversion-delay window after final exposure before reading the outcome. State that delay in days in the document. A result that looks ready is not a substitute for the agreed window.
Safety exceptions need equally firm boundaries. Stop immediately for policy violations, broken tracking, runaway spend, or demonstrable customer harm. If board pressure or quarterly reviews require continuous decisions, use a sequential stopping rule with boundaries declared before launch. An improvised dashboard threshold does not qualify.
Sandeep Dhuri’s pre-registered LLM evaluation committed hypotheses, refuters, and a one-shot generation rule before collecting results, a useful analogue for prohibiting variant swaps or restarts prompted by disappointing campaign data. The discipline is simple, even when following it is uncomfortable.
How to set the economics before the first dollar — minimum deal size, maximum allowable CPC, and required sample
A winner that is statistically valid can still wipe out margin. To prevent that outcome, derive the economics from the business model rather than borrowing a universal lifetime-value-to-acquisition-cost ratio. Subscription businesses, professional services, and high-value B2B businesses differ in gross margins, cash timing, service costs, and retention uncertainty. Those differences belong in the acquisition decision.
Take a dev-tool SaaS that sells annual contracts. Its first input is the minimum annual contract value that justifies acquisition at the planned CPC, after allowing for its costs and required contribution. This is a floor. The CRM must make it possible to distinguish contracts that clear that floor from deals that merely raise conversion volume.
Next, work backward from that minimum contract. The contribution available for acquisition, combined with the expected SQL-to-close rate, sets the maximum affordable cost per SQL. Connecting that limit to the expected click-to-SQL rate gives the maximum allowable CPC. Record the result as a guardrail rather than a spending target, so a conversion-rate win cannot excuse uneconomic clicks.
Then calculate the required sample from the primary metric’s baseline, the MDE, and the declared statistical design. The expected cost of acquiring that sample establishes a budget floor. If the business cannot fund it, redesign the experiment or do not run it.
A fixed share of the total ad budget for experiments, with a minimum credible funding floor, can support this discipline across tests. Measurement must also reach the platform. Import SQL or opportunity events through offline conversions so the campaign can be evaluated against its economic inputs.
What the pre-registration must say about the “we lose” condition, not just the “we win” condition
A complete document spells out both what counts as success and what counts as failure. Could a neutral third party apply it to the results without inventing a judgment call? If not, the decision is still discretionary. For each hoped-for outcome, it must also say when to stop putting money in.
For the metric, pair the required qualified-pipeline improvement with failure to reach that floor after the committed sample, duration, and delay. For economics, pair a cost per SQL the business can afford with one that is above the most it can pay. Deal size must clear its floor too. For strategy, specify that no credible, bounded path to contribution-positive customers remains once the main controllable variables have been tested.
The response must match the scope of the failure. A campaign destroying value while another produces it calls for pausing that campaign. The document must also say when to leave the channel altogether. Exclude vague “needs more time” verdicts by name, so an adverse result cannot be converted into an indefinite extension.
How to pick the primary metric when the sales cycle is longer than the test window
Revenue may be the final outcome even if it cannot be measured within the test window. Waiting for annual recurring revenue can mean waiting a year for each lesson, while optimizing for marketing-qualified leads can reward poor-quality demand. The practical answer lies between those extremes. Pre-commit to the lowest-funnel metric that can conclude within the window, then document why it predicts revenue.
Take a SaaS company that is testing decision-maker targeting. Broad targeting may bring in cheaper leads, while showing ads only to director-level buyers may raise cost per lead but improve the SQL-to-opportunity rate. A plan that uses cost per lead as its primary metric could therefore turn down the stronger result.
The corrected plan names a decision metric such as SQL-to-opportunity rate, provided it can conclude within the declared window. It states the funnel stage, the historical SQL-to-closed-won conversion range from actual CRM data, and why that relationship is stable enough for this test. The proxy needs a documented connection to revenue. An assumed close rate does not supply one.
Microsoft Research’s metric interpretation framework separates data-quality metrics, an overall evaluation criterion, guardrails, and diagnostic metrics. With that structure, impressions and click-through rate help explain what happened, while the chosen pipeline metric sets the basis for the decision. Guardrails can still block rollout when lead quality deteriorates. Neither a click-through win nor an early result at the platform’s default eighty percent power removes the need to wait through the attribution window.
The pre-registration document as a team alignment tool, not just a statistical safeguard
Stopping too soon often comes from a failure to agree on the rules. Leadership sees two weeks of unremarkable data and wants to cut spend. The channel owner wants more time. Without an agreed document, neither side has a common rule to apply. The argument happens when money is already at risk.
Pre-registration moves that negotiation before launch. Sign-off covers permission to spend, the stopping rule, minimum duration, and loss condition. The people likely to demand an early stop have already accepted the circumstances that justify one. People may still disagree, but that agreement gives them something besides the latest dashboard to base the discussion on.
The document also protects a defensible exit. If the test reaches its stopping criteria and produces a negative result, the channel owner can distinguish a planned decision from panic after a bad week. A broader experimentation framework sets how ideas are framed, prioritized, and judged. The pre-registration applies it to one campaign. It is finished and committed before the campaign URL is generated.
What a completed pre-registration makes visible about AI-channel opportunity cost
An honest pre-registration may show that a paid campaign cannot bring in contribution-positive customers at current CPCs. That is useful evidence. It releases capital from an acquisition model that falls short of its declared terms and shows what other investment is being put off while money keeps going into paid ads.
A fast-growing share of B2B software discovery begins in AI answer engines, where buyers can form a first shortlist before clicking an ad. Over-investment in paid search can therefore coincide with under-investment in earning AI citations. But a failed paid test does not, by itself, prove that AI visibility will produce revenue. The alternative deserves the same measurement discipline.
For AI visibility, that means pre-committing the prompts, running them on a fixed cadence, and logging mention rate and citation rate separately for ChatGPT, Claude, Gemini, and Perplexity. A low citation share becomes a signal to act on. Engine-level measurement preserves distinctions that an overall visibility number could obscure.
Getting cited in an AI answer and ranking on Google are different jobs, despite overlap in their underlying signals. A Google Search campaign that works may not work for AI-driven discovery. The paid-test stop rule makes that trade-off visible while there is still capital to redirect. A pre-registration earns its place by making it possible to stop buying clicks when the evidence no longer justifies them, and to evaluate the work of earning citations with equal care.


