Brand Kit auto-fill from website URL:
your brand, learned in 30 seconds.

Open the Brand Kit setup tab. Drop your website URL into the single field at the top. Hit auto-fill. We do the rest:

• Fetch the live page at the URL you pasted (with redirects honored, paywalls respected).
• Read the page and its own stylesheets, the HTML plus the CSS your site ships with it, so your brand colors are found even when the site only declares them in CSS variables or on its buttons and headings.
• Extract the assets, colors, fonts, logo URL, Open Graph metadata, sample copy from the hero section.
• Pre-fill the Brand Kit form , every field that has a sensible value gets filled in. You see exactly what was detected before you save.

Total time from paste to saved brand kit: typically under 30 seconds. You can edit any field after auto-fill, auto-fill is a head start, not a contract.

The color extractor looks in two places, in order:

1. Brand pictures, the theme color the page declares, its Open Graph image, its logo and its favicon. A clear brand color there wins.
2. The page's own CSS, when the pictures carry no brand color (a site built from type and color often has none), we read its stylesheets: the colors it names for a role, like a primary, brand or accent variable, then the colors of its banners and of its buttons, links and headings.

The output is your Brand Kit's primary, accent, and background colors. Rendered carousels, OG images, and overlay templates start using them on the very next generation.

Two more high-leverage detections happen in the same pass:

• Logo, we pull the most likely logo image (from <link rel="icon">, the navbar logo slot, or the OG image fallback) and stage it in your Media Library so the carousel maker and Image Studio can apply it to every generated visual.
• Brand fonts, the typography pass identifies your headline and body fonts (Google Fonts and the common system stacks) and saves them to the Brand Kit. The carousel maker, OG image renderer, and PDF exporter all consume them.

Brand voice is the field new users skip the most often, and the field that has the biggest effect on output quality. Auto-fill gives the LLM a head start by pulling sample copy from the hero section, the about page (if reachable from the homepage), and recent blog post titles:

• Hero copy becomes the first “voice reference” in your Brand Kit, short, on-brand, written deliberately.
• About-page copy (when present) becomes the second voice reference, usually a longer, more narrative example of how you talk about yourself.
• Detected vocabulary, proper nouns and product names that appear repeatedly get added to your preserve-terms list so the AI never substitutes them away.

Auto-fill gets the Brand Kit to ~80% accuracy in ~30 seconds. The last 20% is your call:

• Confirm the colors. The UI shows three to five candidate colors with a swatch preview. Pick or remove any.
• Tighten the voice references. Edit the auto-pulled samples or paste your own (your best post, your favorite paragraph from a newsletter, anything that reads unmistakably as you).
• Add forbidden terms. Auto-fill doesn't guess these, competitor names, off-limits topics, tone you actively avoid.

Save once. Every future generation, every carousel, every OG image, every repurpose run pulls from this Brand Kit. You set it up once and the platform applies it everywhere, forever.

Drift's closed-loop AI compounds over time, but only if it has something to compound around. A new user who skips Brand Kit gets generic posts that sound like every other LLM's output. A new user with a 30-second auto-filled Brand Kit gets posts that already read like their voice on day one, and the closed-loop signal then tunes those further.

The longest a brand kit takes to set up by hand is ~5 minutes and half of new users skip it. Auto-fill removes that drop-off. Net effect on the activation funnel: cold visitor lands on /, generates a sample post, signs up, pastes their URL, sees branded output the AI knows is theirs, all inside the first 90 seconds.

The auto-fill route is read-only and respects standard web conventions:

• No paywall bypass. If the page returns a 401/403/paywall, we extract what's public and tell you what we couldn't reach.
• No login-walled pages.We only fetch what an unauthenticated browser would see.
• No deep crawl. We fetch the URL you pasted plus, optionally, the linked About page. We don't recursively crawl the rest of your site.
• No third-party share. The scraped content is stored only in your account's Brand Kit. We never use it to train any cross-customer model.

A Brand Kit is only worth filling in if the renderer honours it, and for a while ours did not. In July 2026 a customer told us they could not see their brand in the output. They were right, and we could prove how right by scoring it: we run every render against the brand it was supposed to use and measure how much of the brand actually landed.

Re-running the same scorer over the same brands, before and after the engine rebuild on 29 July 2026:

• Mean brand fidelity 0.468 to 0.791. Same scorer, same brands, only the renderer changed.
• Pass rate 44% to 90%. The share of renders that carried enough of the brand to count.
• One gym brand went 0.001 to 0.827. At 0.001 every card came out in the product's indigo and the brand's own primary colour was absent from all eight.

The score now runs in the live quality gate, so a render that loses the brand is flagged rather than quietly shipped. We are publishing the before-number as well as the after-number because a number with nothing to compare it against is not evidence.