A $2M Indie Release, Turned Top-10 Debut.

Client

Meridian Films

Share

smartphone with Google Ads on the screen and display notebook ba

Category

Client

Meridian Films

Meridian came to us eleven weeks out from release with a finished film, a modest P&A budget, and a distribution deal that paid on view-through. Every dollar had to find an audience that would actually finish the movie — not just click on it.

The previous release had spent 68% of its media budget on broad interest targeting and generic trailer cutdowns. It opened, it charted for a weekend, and then it fell off. The studio’s read was that the film underperformed. Our read was that the film never met its audience — the distribution was the problem, not the story.

Where the funnel actually broke

We pulled three years of first-party viewing data, festival sign-ups, and platform-side completion curves into a single model. Within a week, the pattern was obvious: the audience that finished the film was not the audience the media plan was buying.

Three findings that changed the plan

  • Completion rate correlated with cast affinity, not genre affinity — a signal no one was bidding on.
  • 62% of high-value viewers discovered the title on a second-screen surface, then watched on TV within 48 hours.
  • The 30-second trailer outperformed the 90-second cut everywhere except the audience that actually converted.

Rebuilding the release as a system

We rebuilt the plan around four cohorts and let the model reallocate spend daily. Creative was produced against each cohort rather than against the film as a whole — nine cutdowns instead of two, each one testing a single hypothesis.

  1. Model the audience. Three years of first-party data into a single completion-weighted cohort map.
  2. Produce against cohorts. Nine creative variants, each with one testable claim.
  3. Reallocate daily. Spend followed completion rate, not click-through.
  4. Hand the system over. The dashboard shipped with the film and still runs the catalog.

Instrumentation

Every surface reported into one event schema, so the model could compare a TV completion against a second-screen start without reconciliation.