A new channel can spend a small budget without answering a useful question. It can also produce a handful of promising leads that tell you very little about what would happen at a larger scale.
We would start a first conversational-ad campaign with one decision: what evidence would make us run a second test? That decision should shape the audience, offer, measurement and spending limit together.
Pick a question you can actually answer
“Does AI advertising work?” is too broad for a pilot. “Will people researching shared inboxes visit our comparison page and start a qualified trial?” is something a campaign can begin to investigate.
Keep the first offer and destination stable long enough to understand them. If the audience, copy, price and landing page all change mid-test, the final total will conceal several different experiments. Change a broken experience immediately, but mark the change in the report.
Before launch, write a short agreement with the people evaluating the results:
- Audience: the specific task or context in which the offer belongs.
- Useful action: an event that has value to the business, with an exact definition.
- Limit: the most you are willing to spend to investigate this question.
- Review point: a time or delivery threshold for inspecting the evidence.
- Stop conditions: a broken destination, unsuitable placements, measurement failure or an exhausted limit.
A stop condition is not pessimism. It stops a small experiment from turning into unattended spend.
Work backwards from the action
Suppose, purely as an example, a software company sets a $600 learning budget and defines a qualified trial as a new account that completes setup and matches its intended customer profile. If the pilot produces six such trials, the observed cost is $100 per qualified trial.
That arithmetic does not tell us the cost of a customer. We still need to know how many trials become customers and how much those customers are worth. Nor does six trials establish a stable channel average. Report the count alongside the cost so the uncertainty stays visible.
A budget should come from what the business can afford to learn, not from a generic “recommended minimum.” Jolts' pricing page describes the available buying models; the relevant volume and terms need to be agreed for the particular campaign.
Read the failures before the average
No delivery suggests an inventory, eligibility or setup question. Delivery without visits suggests the offer or placement may not be compelling—or that click measurement needs checking. Visits without qualified trials push the investigation toward the destination, product fit and qualification rule.
These are hypotheses to investigate, not diagnoses a dashboard can make automatically. Look at a small set of permitted placement examples, validate the event path and ask sales what the resulting enquiries were actually about.
A campaign can meet a cheap cost-per-click target while producing no useful outcomes. It can also have expensive clicks and a promising downstream result. Choosing the target before the test keeps the evaluation from drifting toward whichever metric looks best afterward.
Decide the second test before scaling
At the review, choose among three actions: stop, repair one identified problem and repeat, or expand a specific working segment. Record why.
For a promising trial campaign, expansion might mean another audience context with the same offer, not an immediate tenfold budget increase. For a failed handoff, keep the placement constant and fix the landing page. The first campaign earns its place by improving the next decision. A large impression count alone does not do that.