Data science and creativity / Robert Urban
The numbers can sharpen the idea. They cannot have it for you.
How to turn marketing evidence into a better creative question, then give the answer a fair test.
I have always been interested in the space between science and creativity.
Since opening PaperBoat in 2011, that has been part of how I describe my work. The useful connection is not that a spreadsheet can replace an idea. It is that evidence can make the idea more specific, and a thoughtful experiment can help us learn whether it does what we hoped.
A report can show that visitors leave a page. It cannot, by itself, tell the writer which explanation would make the page worth reading. A designer can make the offer easier to see. That does not establish that the offer is the right one.
The best conversation lets analytical and creative work challenge each other. The analyst asks what the evidence actually supports. The writer asks what a person still needs to understand. The business owner asks which decision would change if the work succeeds.
A useful finding is not yet a creative brief.
“The landing page converts poorly” is a complaint. It leaves the team guessing about the audience, desired action, traffic source and reason for the result. “People who request commercial installation details cannot find the site requirements before contacting us” is a much more useful starting point, if the evidence supports it.
Before commissioning creative work, write down the observed behavior, the interpretation and the proposed change separately. That separation gives someone else a fair chance to disagree.
What was counted or seen? Include the period, source and relevant audience. Preserve the original records where appropriate.
What might account for the behavior? Name plausible alternatives. A drop in inquiries could reflect traffic mix, availability, tracking changes or demand as well as page quality.
What must the reader understand or feel able to do? Give the writer and designer room to solve that problem without inventing a conclusion.
Fictional example / A museum membership page
More clicks might be the wrong improvement.
Imagine a museum team reviewing a membership campaign. Its report shows 1,000 campaign visits and 20 membership purchases. For this simplified example, purchases divided by visits give a 2% purchase rate. That number is a description, not a diagnosis.
The team initially wants a louder button. A review of visitor questions suggests another possibility: people do not understand which membership level fits their household. The existing page emphasizes discounts but leaves guest eligibility difficult to compare.
The creative brief becomes specific: help prospective members choose the appropriate level before asking them to purchase. The team develops a plain-language comparison, explains the actual guest rules and puts the next step beside that explanation.
Every policy in the new copy must be confirmed by the museum. A persuasive explanation of an invented benefit would create a customer-service problem, not a successful campaign.
Give the idea a fair chance to be wrong.
Causation unproven
Suppose the next period records 30 purchases from 1,000 visits. That is 3%, an increase of one percentage point from the original 2%, or 50% relative to that original rate. Those calculations are correct. “The new design caused a 50% improvement” is a stronger claim than the before-and-after numbers establish.
The campaign audience may have changed. A school holiday may have increased interest. A special offer may have started. The membership mix may differ. Tracking may have changed. Look for these explanations before crediting every additional purchase to a headline.
NIST’s experimental-design guidance distinguishes association from causation. Variables can move together without one causing the other. Designed experiments help investigate causal relationships under appropriate conditions.
When the decision and volume justify it, agree on a suitable test with the person responsible for measurement. Define the primary outcome, assignment approach, comparison conditions and stopping rule before examining results. Do not keep changing the test until a pleasing number appears.
For a low-volume organization, observation, interviews and usability work may produce valuable learning without a statistically decisive conversion experiment. State what that evidence can tell you. You can make a reasoned improvement while remaining honest about what has not been established.
Keep a measure of quality beside the count.
A campaign can generate more inquiries while creating more unsuitable work for the team. A membership page can increase purchases while confusing buyers about benefits. A dramatic headline can increase clicks while disappointing nearly everyone who follows it.
Choose a measure that protects the actual purpose: qualified inquiries, completed suitable purchases, fewer avoidable clarification calls or another outcome the business can define and observe.
Check the timing. A campaign launched yesterday has not had the same opportunity to produce a sale as one running for three months. Compare outcomes at a similar stage and keep incomplete cases visible.
Big data is not a prerequisite for a better question.
Data analytics examines information to understand patterns and support decisions. Data science draws on analytical, computational and domain methods to investigate problems and build models. Big data describes challenges of scale, speed or variety that may require different processing approaches. These terms overlap, but buying a more elaborate platform does not make the business question clearer.
A manageable set of inquiries and outcomes may be enough to reveal a useful creative problem. Start with consistent definitions, trustworthy records and permission to use the information. If the task genuinely requires integrating large or complex sources, choose the infrastructure around those requirements.
I would be cautious about a tool shortlist assembled before anyone has named the decision, data sources, owner and expected output. The relevant question is what the team needs to learn or do, not which product makes the most impressive slide.
For measurement systems and reporting ownership, see my marketing analytics and reporting work. For estimating future demand or behavior, the predictive analytics guide addresses a different analytical task.
What to bring to the next creative review
Bring the page or campaign, the intended audience, the action it should support and the actual questions people ask. Add the relevant report with definitions and dates. Include examples that contradict your preferred explanation.
Ask the creative team to explain which uncertainty their proposed change addresses. Ask the analyst what evidence would change their interpretation. Ask the operating team whether the promise can be delivered.
That combination makes the work more interesting. The numbers identify a place to look. The creative work proposes a better experience. The evaluation helps decide what deserves another round.
Should data decide every creative choice?
No. Evidence can inform the assignment and test relevant outcomes, but judgment still matters. Brand character, clarity, accessibility and truthful communication cannot always be reduced to the largest short-term response count.
Does a correlation tell us why customers acted?
It suggests a relationship worth investigating. Other factors may explain it, and the records may not capture the person’s reasons. Combine appropriate analysis with a suitable evaluation and direct customer understanding.
What if there is very little data?
Use what is available carefully. Inspect individual cases, listen to recurring questions and test whether people can complete the task. Avoid presenting a handful of observations as a reliable market-wide estimate.

