Jev makes ad research
feel like cheating.

01 / What does what

Collect. Classify. Find examples. Make something.

The app groups exact matching evidence, shows the age of active ads, and lets you inspect the original examples. A separate, optional writing model turns a reference pattern and your verified product facts into a brief. Jev reviews clarity, brand fit and claim support.

02 / Start without spending anything

Get the simulation running first.

1. Install Node.js, version 22.9 or newer.
2. Download the ZIP above and extract it. Open a terminal inside the folder containing package.json.
3. Run these commands, then open http://127.0.0.1:5194.

npm ci
npm run setup
npm run doctor
npm start

For the light-mode demo above, open the recording link below. It needs no keys. The thumbnail wall, map and counters advance together: each tile crossing the scan edge counts once. Choose 20, 40 or 80 seconds in the controls. C restores controls, R restarts and Space pauses. The reference layout is wide; the other three variants also offer portrait framing. For actual imports and analysis, open http://127.0.0.1:5194.

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

The timer is replay time. The starter uses generated artwork. The recording above draws from a local pool of 2,136 public creative images; its processed count follows the visible tiles, rather than the full pool size. Both use synthetic labels, ages and scores. It shows how the interface works, not how quickly a paid workflow will finish.

Full recording guide: controls, image packs and what the counters mean. The included MP4 is a 20-second preview. Screen-record the app to capture your own image pool or a different duration.

npm run demo:download

npm run demo:images -- /path/to/your/images

Set up https://github.com/artemnovitckii/ad-pattern-lab locally. Read README.md, docs/SETUP.md and docs/PRIVACY.md. Run the no-key simulation at /variations.html?variant=reference, doctor, syntax checks and tests. Do not read or print secret values. Ask me to add keys to my local .env only when needed. Start with a five-ad import and tell me what you actually verified. Do not launch paid collection without my instruction.

03 / Your first real analysis

Start with ads you can check yourself.

Choose Import ads. Paste ad copy, upload a JSON array, or upload images and videos. The repo includes a small fictional sample JSON so you can see the fields. For a real pilot, use five ads you can open and inspect.

Jev

Add a key from TypeSafe to TYPESAFE_API_KEY. It classifies the words and supplied observations.

Local OCR

Tesseract.js runs on your computer with no per-image API fee. First use downloads the English language model. Check its output on stylized or tiny text.

Speech, if needed

Choose Fireworks or Groq for video audio. Existing transcripts can be imported instead.

TYPESAFE_API_KEY=your_typesafe_key
TRANSCRIPTION_PROVIDER=fireworks
FIREWORKS_API_KEY=your_fireworks_key
PORT=5194

For Groq, set TRANSCRIPTION_PROVIDER=groq and fill GROQ_API_KEY. Leave unused keys empty. Save the file and restart the server, then press Analyze / retry unfinished.

brew install ffmpeg

winget install --id Gyan.FFmpeg --exact

sudo apt update
sudo apt install ffmpeg

JSON and pasted copy replace the current local batch. Uploads append. Export your work before importing another batch. Click a creative to inspect its source text, labels, errors and original link when available.

04 / Optional collection

Bring the Meta Ad Library into the same workflow.

Add your APIFY_TOKEN to .env, restart, and choose Meta collection. The connector uses this Apify Actor. It is implemented but has not had a paid live validation run.

1. Paste public Facebook page or Ad Library URLs for a focused category.
2. Start with 20 ads and a small Apify cap, such as $1. Check current Actor pricing in your account.
3. Click Start paid collection and inspect the run in Apify Console.
4. After collection, click Analyze collected ads separately. That is when the AI stages run.

The Apify cap does not cover Jev, speech or writing costs. Collection is a sample, not a promise to retrieve every ad. If a launch response is lost, check Apify Console and attach the existing run ID instead of starting a duplicate.

05 / Turn research into a test

Leave with a brief you can actually shoot.

1. Filter a hook, format, angle or offer. Check the number of ads, distinct families and brands behind it.
2. Open examples and study the mechanism: the opening problem, the evidence and the offer. Check weaker or different examples too.
3. Open Your product → original briefs. Add your audience, real offer and facts you can support.
4. Create up to six briefs. With no writer configured, these are editable outlines. Add a compatible text model for original drafts.
5. Review the words and product claims, shoot an original version and measure results in your own ad account.

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

Creative families use exact fingerprints, not visual similarity. Signed URLs and copy edits can split near-duplicates. Several ad IDs can also be placements or variants rather than independent tests. Use the map to find examples, then judge the evidence.

06 / Keep going when something fails

Check the stage that needs attention.