Validating a video campaign before producing it

TL;DR. In the last 18 months, two AI tools have become accessible to a mid-sized agency: multi-agent simulation for testing concepts, and predictive neural analysis for iterating on score, edit, and mix. Applied to the ECLAG project: 3 videos · 18 neuro analyses · 3 weeks. Seven operational lessons we now apply by default.

Opening

Nine out of ten video campaigns get signed off on a judgment that sounds something like: “we like it.” The second opinion arrives too late — when the campaign is already live and the numbers are set. For years this has been the norm even for campaigns with the highest stakes: NGOs, public health, advocacy. Not out of bad faith, but because of cost. A traditional neuromarketing test costs tens of thousands of euros and weeks of setup. Inaccessible for most projects.

In the last eighteen months that balance has shifted. Two tools have emerged that, when integrated into the creative process, give a mid-sized agency the ability to bring the client a second quantitative opinion on creative choices — before the video is rendered, before asking for the go-ahead.

The first is multi-agent simulation (in English: swarm intelligence): a panel of “synthetic viewers” generated by frontier language models, each with a defined demographic and psychographic profile, evaluating creative materials before production. It’s a system we built ourselves and that today is part of our toolkit. The second is our predictive neural analysis stack, which uses a research foundation model to predict how the cortex of an average subject responds to video, audio, and text stimuli. It’s not a black box that says “this campaign will win”: it’s a crafting tool that lets you iterate with structured feedback.

We applied both to the ECLAG Missing Children project, a European video campaign on children’s online rights. Three weeks of work, eighteen consolidated analyses, three videos delivered. This is the story of how we did it, what we learned, and why we believe this way of working will soon be the norm even for projects with ordinary budgets.

What the neural analysis stack actually measures

Our predictive neural analysis stack is based on a research foundation model trained on roughly 5,000 hours of naturalistic stimuli paired with real fMRI scans. Given a video or audio as input, it returns a predicted activation vector across 20,484 cortical points — an average subject, under attentive viewing conditions. Six groups of brain regions are particularly informative for creative judgment on an emotional campaign:

• OFC (orbitofrontal cortex) — affective response, value evaluation
• STS (superior temporal sulcus) — recognition of human voice and biological motion
• Speech — speech processing
• Attention (frontoparietal network) — attentional engagement
• TPOJ (temporo-parieto-occipital junction) — audio-visual integration
• DMN (default mode network) — mind-wandering, distraction (we want this low: it means the brain is following the stimulus rather than drifting)

A single stack run in audio-only mode takes 2–3 minutes on a standard machine. The full video+audio pipeline takes a few hours of CPU, or a few minutes on cloud GPU. Cost per iteration: zero.

Method note. Our stack is a predictive model, not a measurement. The predictions are reliable in relative terms (A is higher than B on OFC) and directionally (this prompt change raises or lowers activation). They aren’t reliable as absolute values, and they don’t predict behavioral metrics like watch-through rate, share, or conversion. They guide creative crafting, not certify final success.

Three weeks, eighteen iterations

3 videos delivered · 18 neuro analyses · 3 weeks of work

The iterative cycle was structured this way: generate a stimulus (a music track via Suno or ElevenLabs Music, a video edit, a mix), test it with our stack, identify the single ROI with the largest deficit, make a targeted change to the prompt or edit, re-test. Five to ten minutes per audio-only cycle. Twenty to thirty iterations a day, in practice.

The clearest example of what “targeted change” means came on the musical composition, when the affective response (OFC) was consistently below the global activation average. The literature indicates that the OFC responds to the resolution of expected harmonic tension: meaning we have to give the brain “release points” of affect that build and then resolve. Two specific edits to the composer’s prompt:

Edit 1. “Extended major chord moments at second 50 and 58, each held 3-4 seconds, before returning to minor”

Edit 2. “Final modulation to C major, sustained 4 seconds at fortissimo before the cut”

Result in four iterations: OFC = −0.003 (essentially neutral, from a starting point of −0.065). Translation: the brain now produces a normalized affective response where there was none before.

The same outcome in a standard process would require 2–3 rounds of “blind” client feedback — each round costs days, and each round is a moment in which the client-agency relationship can deteriorate through mutual frustration.

The most interesting finding

Video 3 in the series (Restore) brings on screen, with a progressive blur effect, the silhouettes of children reemerging in the background — it’s the hopeful closing of the narrative arc. When we tested the complete video and then the isolated soundtrack, we ran into an unexpected data point:

“The soundtrack alone produces a neuro score of −0.284. The complete video, with the same audio, climbs to −0.094. The visual carries 67% of the neuro value.”

Translation: the visual concept of video 3 — the human figures reassembling — activates face-recognition and biological-motion areas much more powerfully than the music does. The visual saves the audio. It’s a compliment to our team’s editing work. And it’s also a missed opportunity: had the soundtrack been neuro-optimized to the level of the other tracks in the series, the overall pattern would have been meaningfully higher.

Seven operational lessons

From the eighteen runs we distilled seven operational principles we now apply by default on any project with an emotional audio-visual component:

1. The human voice is the most powerful neuro lever. An isolated narrating voice produces stronger social activation than any music, than any video. Mix design: voice in the foreground, always.

2. Minor → major harmonic modulation is the single most effective edit for activating the affective response.

3. Wordless female vocalise is the most reliable lever for activating the social cortex, even in purely instrumental scores.

4. Tension + earned catharsis beats both pure drama and sober ambient.

5. A strong visual can compensate for weak audio. We verified it. But it’s a waste: with both optimized, the gain is cumulative.

6. Duration dilutes the average. A/B comparisons should be made at equal duration.

7. Musical mental imagery is a real effect. Instrumental tracks, even without video, activate visual areas.

What we DON’T promise

To be clear, the method isn’t an oracle. Three stated limits:

• Our stack predicts the neural response of an average subject in attentive viewing. It doesn’t simulate the platform algorithm, the 2-second scroll, or the fragmentation of attention on mobile.

• Predictions are correlated, not identical, to behavioral metrics. A good neuro pattern is a necessary but not sufficient condition for success. Success also depends on algorithm, targeting, sponsorship, and timing.

• The model covers the cortex, not the subcortical structures (amygdala, nucleus accumbens). For messages that play primarily on deep fear/reward, our stack is blind.

Why it matters that this is accessible

Until twenty-four months ago, two services like these were available only to multinationals with dedicated R&D budgets and relationships with academic neuromarketing labs. Costs: tens of thousands of euros per study. Timelines: months.

Today both technologies are accessible to a mid-sized agency that knows how to integrate them into its process. What’s needed is the operational know-how: knowing which ROIs to watch, how to translate numbers into brief edits, how to manage iteration without being overwhelmed by data.

That know-how is what Pianeta.Studio built on the ECLAG project, and what we carry into all subsequent projects as part of the standard creative process, with no extra fee.

FAQ

Who needs AI validation of a video campaign? Organizations running campaigns with high stakes: NGOs, public health, advocacy, sustainable brands, institutional communication. When a mistake costs a lot, a second quantitative opinion before rendering pays for itself.

How much time does it add to the creative process? Zero, because it’s integrated into the iterative cycle. A single audio-only run takes 5–10 minutes, and the number of iterations depends on the project. For ECLAG, it was 18 analyses over 3 weeks.

Does it replace human creative judgment? No. It’s a second quantitative opinion that guides crafting, not an automated decision. Human creative judgment remains the final decision-maker.

How much does it cost? It’s included in Pianeta.Studio’s standard creative process, with no separate line item. The cost of the creative project follows the standard rate card.

Does it also work for projects smaller than a European campaign? Yes. It works on single spots, educational content, fundraising assets, brand videos. The threshold is the emotional relevance of the message, not the budget.

See also

→ Choose to See Them — the ECLAG case → Service · Creativity and neuromarketing → The live campaign at choosetoseethem.childsafetyineurope.com

Talk to Alba

If you have an upcoming project with a significant video or audio component, let’s figure out together whether it makes sense to apply the method.

info@pianeta.studio

---

Technical stacks used: Pianeta.Studio’s proprietary predictive neural analysis stack (based on a multimodal research foundation model, Glasser HCP cortical atlas, fsaverage5 mesh with 20,484 vertices) · Pianeta.Studio’s proprietary multi-agent engine based on frontier LLM models.