Brand DNA extraction is the core of what makes Mani different from generic AI ad tools. Instead of starting from a blank slate, every generation starts from your actual brand identity. Here is what happens when you paste your URL.
The extraction pipeline
When you paste your URL, Mani follows a four-stage process:
- Scrapes your website: Fetches your homepage and up to 5 linked pages (About, Products, Pricing, Blog, Contact). Extracts text content, CSS stylesheets, images, meta tags, and structured data. The scraper respects robots.txt and uses a standard user agent.
- Runs 6 parallel extractors: Tone, audience, palette, products, keywords, and visual style extractors run simultaneously. Each uses a specialized AI prompt grounded in your scraped content. The tone extractor analyzes sentence structure, vocabulary, and formality. The palette extractor parses CSS custom properties and computed styles. The product extractor identifies items with names, descriptions, and prices.
- Applies a critic layer: A critic model reviews each extraction for specificity. Generic outputs ("modern teams," "high quality," "great products") get flagged and re-extracted with tighter constraints. This critic pass is what prevents Mani from producing the bland, one-size-fits-all descriptions that plague most AI tools.
- Assembles the Brand DNA card: All six extractions merge into a single Brand DNA profile stored on your account. This profile becomes the grounding context for every future generation.
Why 90 seconds?
The parallelization is key. Running six extractors sequentially would take 5+ minutes. By running them in parallel with cached scrape data, the full pipeline completes in about 90 seconds. The scrape itself takes 10-15 seconds. Each extractor takes 10-20 seconds. The critic pass adds 5-10 seconds for any flagged fields.
Accuracy
Extraction accuracy improves with richer websites. A site with clear product pages, an about page, pricing, and consistent styling produces excellent Brand DNA. A one-page landing page with minimal text produces a thinner profile. If your site is minimal, you can supplement the extraction with manual edits. Most users find the extraction 80-90% accurate on the first pass and spend 2-3 minutes fine-tuning specific fields.