Content experimentation in financial products

In an email marketing campaign developed by Parcelex in partnership with Portocred to promote personal loans, we sought to identify which messaging approaches generated the most interest.

I was responsible for the content strategy and email copy, including writing the four CTA variants used in the experiment. The email's visual design was developed by Product Design.

We tested four variations of the call-to-action button text, randomly distributed through HubSpot. The rest of the email remained unchanged, allowing us to compare the alternatives' performance.

The hypothesis was that different language choices, even within a single component, could influence people's willingness to click and learn about the offer.

The four alternatives explored distinct approaches:

A — Exclusivity: highlight the possibility of special terms with a low-commitment invitation. CTA: “Check special terms”.

B — Urgency: use a direct invitation associated with immediately applying for the product. CTA: “Apply for a loan now”.

C — Personal identification: bring the message closer to everyday language, expressing the need for credit in the first person. CTA: “I need a loan”.

D — Rationality: present a loan estimate as a step before deciding to take out a loan. CTA: “Get a personal loan estimate”.

Variation B, with a direct invitation to apply for a loan, recorded the highest observed click rate: 38.71%. It was followed by variation D, focused on getting an estimate, at 35.34%.

The exclusivity and personal identification approaches had lower rates in this experiment.

Calls to action that explicitly described a concrete action — applying or getting an estimate — recorded more interaction than those centered on generic benefits or personal identification.

The experiment provided evidence to guide future communication decisions, showing how small language changes can be associated with meaningful differences in behavior.

The percentages represent results observed in the campaign, without documented proof of statistical significance or higher loan conversion.

With the introduction of the Special Return Mechanism (Mecanismo Especial de Devolução, or MED), established by Brazil's Central Bank, Pix disputes became part of a journey that was still unfamiliar to many people.

There was uncertainty about the situations covered by the mechanism. Transfers sent by mistake to the wrong account, for example, also prompted dispute attempts, although they were not covered by the MED.

This contributed to a substantial increase in requests and created operational challenges.

At the same time, there was a fundamental constraint: the journey could not prevent people from opening a dispute.

The challenge was to make the rules easier to understand and help people assess their requests more carefully, without creating undue obstacles for those who needed to use the feature.

I was responsible for designing the four interface variants and developing the content tested, exploring different ways to present the MED rules and request confirmation during the Pix dispute journey.

The experiment was conducted in partnership with Itaú's Experimentation team, which was responsible for methodological rigor and statistical validation.

The alternatives explored different ways to present information and request confirmation, including:

• →Presence or absence of a modal.
• →Single or double confirmation.
• →Confirmation on the first or second screen.
• →Inclusion of police report information.
• →Changes to the presentation of guidance and the conditions for using the mechanism.

We sought to understand how these decisions might influence continuation through the journey and contribute to more responsible use of disputes.

The objective was not to encourage more disputes or prevent requests, but to balance clarity, responsibility and access.

The experiment involved thousands of clients, was overseen by Itaú's Experimentation team and underwent statistical validation.

During the 13 days analyzed, we compared the four variants using indicators of behavior and the appropriateness of requests.

Variant D was selected for implementation after recording the highest abandonment rate at the stage analyzed: 12%, compared with 4% in the control group.

Although variant C had the highest observed percentage of valid disputes opened (45%), the decision prioritized increased abandonment at that stage, in alignment with the experiment's operational objective.

Selecting variant D was part of a strategy to reduce the continuation of potentially inappropriate requests while preserving the ability to file a dispute.

The experiment showed that a metric traditionally interpreted as negative can take on a different meaning depending on the context.

In a dispute journey, abandonment was not necessarily a problem to eliminate. The challenge was to provide enough information for people to understand the rules and reconsider requests that did not match the MED's purpose.

At the same time, higher abandonment alone does not prove that only inappropriate requests stopped progressing. The assessment therefore needed to consider the operational objective, appropriateness indicators and regulatory constraints.

Selecting variant D reinforced the importance of evaluating content decisions within the full context of the experience, rather than through a single metric.