OpenAI Decisions makes ad research feel like cheating.

01 / THE FASTEST WAY IN

Open the demo before connecting anything.

This is a separate OpenAI edition of my Jev marketing project. Install Node.js 22.9 or newer. Download and extract the code, or clone it below. Open a terminal inside the folder containing package.json.

git clone https://github.com/artemnovitckii/openai-decisions-marketing.git
cd openai-decisions-marketing
npm ci
npm run setup
npm run doctor
npm start

Open http://127.0.0.1:5194/. The app opens on a completed sample report: thumbnail wall, creative map, findings, an example inspector and 24 briefs. Filter by format and hook. Click a point or image to inspect it. Press Replay demo run to watch the wall and map build together.

http://127.0.0.1:5194/variations.html?variant=reference&clean=1&duration=20

The simulation needs no keys and makes no AI calls. The included starter artwork is fictional. The demo above uses a larger local public-image library. Sample labels, dates, grouping, scores and replay speed are invented for the interface.

npm run demo:download
# Or use your own folder:
npm run demo:images -- /path/to/images

Set up https://github.com/artemnovitckii/openai-decisions-marketing locally. Read README.md, docs/SETUP.md and docs/PRIVACY.md. Install dependencies, run setup, doctor, checks and tests, then start the no-key demo. Do not print secrets. For real analysis, ask me to add OPENAI_API_KEY to my local .env and start with five uploaded images. Do not launch paid collection without my instruction. Report exactly what was verified.

02 / YOUR FIRST REAL ANALYSIS

It sees the creative, as well as the words.

The real classifier uses OpenAI's dedicated Decisions API, currently in public beta with gpt-6-luna. The request includes the actual image pixels, available ad copy and optional speech. It is not a text-only description of the image.

1. Install FFmpeg and ffprobe using a command below.
2. Create an OpenAI API key and put it in your local .env. API access and billing are separate from a ChatGPT subscription.
3. Restart the server. Choose Import ads and upload five real static creative images you can check yourself.
4. Click Analyze / retry unfinished. Static uploads need no Apify or speech key.
5. Click a creative, then All labels & evidence. Inspect the exact submitted image, labels, confidence and provider information.

OPENAI_API_KEY=your_openai_key
ANALYSIS_PROVIDER=openai
PORT=5194

brew install ffmpeg

winget install --id Gyan.FFmpeg --exact

sudo apt update
sudo apt install ffmpeg

TRANSCRIPTION_PROVIDER=fireworks
FIREWORKS_API_KEY=your_fireworks_key

TRANSCRIPTION_PROVIDER=groq
GROQ_API_KEY=your_groq_key

Upload local files, paste copy or import an existing JSON export. JSON and pasted copy replace the saved batch; uploads append. Export your current work first. Keys stay server-side, and the repository does not contain mine.

03 / THE RESEARCH LENS

Seven labels that lead to a useful test.

Format

Creator demo, testimonial, founder, product, comparison or graphic. What is the creative doing visually?

Hook

Customer frustration, curiosity, result, contradiction, offer or direct introduction. How does it open?

Angle

Convenience, performance, price, trust, experience or identity. What reason to care does it give?

Offer

Discount, bundle, subscription, trial, guarantee or no explicit offer.

Proof

Demonstration, testimonial, comparison, numbers or a named source. This describes evidence offered, not whether a claim is true.

Objection

Price, trust, effort, effectiveness or suitability. Which hesitation is addressed?

CTA

Buy, learn, sign up, book, message or no explicit next step.

Decisions returns typed choices and confidence. The app joins collection dates and active status afterward. Unknown is a valid answer when the available evidence is insufficient. You can change the comparison dimension, filter the wall, inspect examples and export the analysis.

04 / OPTIONAL META COLLECTION

Bring the Ad Library into the same view.

Add APIFY_TOKEN to .env, restart, and choose Meta collection. The connector uses the Facebook Ads Library scraper. Start with a focused category and a small batch.

1. Paste public Ad Library or Facebook page URLs.
2. Set a small maximum, such as 20 ads, and an Apify cap. Check the Actor's current pricing in your account.
3. Click Start paid collection and inspect the run in Apify Console.
4. Click Analyze collected ads separately to start classification.

Collection does not promise every ad. The Apify cap excludes OpenAI, speech and writing. If a launch response is lost, inspect the existing run and attach its ID instead of paying for another run. You can skip this stage entirely with your own files or JSON.

05 / FROM RESEARCH TO A SHOOT

Make something original from the pattern.

1. Filter a promising format or hook. Check how many examples and families support the comparison.
2. Inspect the opening, proof and offer in several source creatives. Study different and weaker approaches too.
3. Open Your product → original briefs. Supply your audience, actual offer and facts you can support.
4. Create up to six briefs. Without a writer configured, these are editable outlines. A separate compatible text model can write the drafts.
5. Decisions reviews clarity, brand fit and claim support. Review the claims, shoot an original test, and measure it in your own account.

LLM_BASE_URL=https://api.fireworks.ai/inference/v1
LLM_API_KEY=your_writer_provider_key
LLM_MODEL=your_available_text_model_id

06 / COSTS, CHECKS & FIXES

Start small. Check what actually happened.

The demo has no provider charges. The documented Decisions base rate is $0.10 per million input tokens, with no output-token charge. Regional and long-context premiums can apply. The app estimates returned usage; it excludes Apify, speech, writing and unknown request costs. Official pricing and availability.