Case 04 · Marketing

I write the script after looking at what people ask and what they are looking at.

A dealership's content can be decided by taste or by evidence. I set up a loop that turns content metrics, customer questions and interest signals into concrete scripts, and that clearly separates a hypothesis, a test in progress and a proven result.

Project: Data-driven scripts and content
Context: Car dealership in Panama · organic content, salespeople and chat
By Antonio Duque: Reviewed against the code and documentation on November 14, 2026

Context and need

Videos and posts were decided without data, while WhatsApp conversations kept repeating the same questions and friction points.

Salespeople also record content, but had no guidance on which car to show or which hook to use.

My responsibility

As the marketing lead, I defined the process, the metrics to review, the hypotheses and the criteria for each piece, and directed the build of the reports that feed it.

Strategy and decisions

Look at behavior first: reach, engagement, saves and shares, and which models generate intent.
Listen to conversations: a month of anonymized chats showed financing dominates the questions, followed by requirements, availability, photos, payment and down payment.
Turn each finding into an explicit hypothesis and a concrete piece, not a general opinion.
Never call something an A/B test when it isn't: the available comparisons are observational.

What I built

A daily Instagram content report: flags a piece as high quality when it beats the median on both reach and engagement rate, compares reels with photos and groups by car model, using saves and shares as a sign of intent.
Video ideas for salespeople: it picks financeable cars with rising ad spend and clicks, delivers a ready description and a hook for the first three seconds, rotates salespeople and never repeats a car within 21 days.
An anonymized conversation analysis that changed the assistant's replies and detected ads still promoting sold models.
A priority-based rewrite of the WhatsApp assistant flow, with one rule: the assistant is never a dead end.
A production workflow with local transcription and per-network copy that never invents prices: it leaves a placeholder.

How it works

System
Team

System
Review metrics
Content, conversations and chat drop-off, in context.
Team
Spot patterns
Frequent questions, friction, models that generate intent.
Team
Form a hypothesis
For example: the biweekly payment in the first seconds holds attention better than the price.
Team
Script and visual proposal
Hook, car, message and call to action, formatted per network.
Team
Publish or test
Measured per piece, with UTMs when there is a link.
Team
Review and adjust
The hypothesis stays open until there is enough evidence to keep or drop it.

Demo

What is proven and what isn't

Hypothesis, test and result are not the same thing

Diagram

Reels perform better than photos for this catalog.Hypothesis
Saves signal purchase intent per model.Hypothesis
Opening a video with the biweekly payment holds attention better than the price.Hypothesis
Paying for portal highlighting helps move inventory (monthly observational comparison).Test in progress
Financing dominates customers' WhatsApp questions.Verified finding

The verified finding comes from a month of anonymized conversations. None of these items is a controlled A/B test.

Tools and what they do

Instagram API: Reach, engagement, saves and shares.
Laravel: Scheduled reports and car selection for videos.
Local Whisper: Video transcription without uploading footage to third parties.
Claude: Draft copy per network, reviewed and without invented prices.
Spreadsheets and UTMs: Per-piece tracking in per-vehicle creative campaigns.

Available evidence

Content report and video selection scheduled on the server.
A conversation-analysis document with the changes applied.
A per-vehicle creative campaign (September 2026) with UTMs per piece and the real quote formula.

Current status

Daily content report
Scheduled. Its recommendations are fixed rules, not a statistical model.
In production
Video ideas for salespeople
Scheduled with rotation. No documented measured result.
In production
Reply changes from conversation analysis
Applied to the assistant; impact not measured.
Built
Comparison of paid highlighting on a portal
Monthly observational design, not a controlled experiment. No results yet.
In progress
Effect of content on lead quality
Open hypothesis: no evidence proves it yet.
Proposal

“In production” is only used when the project’s documentation confirms it. “Built and tested” means automated tests exist, not that it is in use. “Proposal” is a documented idea, not something built.

Next steps

Match UTM-tagged pieces with lead quality in the CRM, not just engagement.
Design a controlled test once volume allows it.

Lessons learned

Saves say more about intent than likes, but they are still a signal, not a sale.

A conclusion written into code (“time of day makes no difference”) also needs its analysis; I mark it as pending review.

Customer questions are the best content brief there is.

Related services

Content audit

Content audit for your company

We analyze your current content to improve clarity, authority, conversion and consistency.

Video marketing

Video creation and editing

We write scripts and edit video to explain, sell, teach and feed your campaigns.

Marketing connected to data and operations

Marketing and Sales Process Automation

I connect inventory, campaigns, CRM and email so what you advertise exists, what you measure is true and what breaks raises an alert.

Connected cases

Inventory-driven dynamic marketing

Inventory as the single source for catalogs, content, conversion signals and ad decisions.

Monitoring and action dashboards

Dashboards that don't stop at the number: every out-of-range indicator suggests a concrete action.

Financing chatform

A credit form turned into a conversation: one question at a time, with guidance and context for the salesperson.

Discuss a project Hire me