A vision LLM tags every ad creative into structured attributes and joins them to your attribution and margin data, so you see which creative traits drive the cheapest, highest-margin new customers, not which ad id got clicks.
For this brand, creative is the lever that moves performance, not targeting. But the ad platforms report by campaign and ad id, never by what the creative actually is. So nobody could say which hooks, formats, or offers worked. Creative testing was gut feel, and the winners were never systematically understood.
01 The Challenge
Post-ATT, the dominant lever in paid acquisition is the creative itself. This brand was producing and spending against a high volume of it, but the reporting stopped at the ad id. There was no way to ask "do UGC hooks outperform studio?", "does leading with the offer beat leading with the problem?", "which format brings in customers who actually have good margin?" Each test was judged on platform ROAS for that one ad, and the lessons evaporated.
The creatives were the experiment. Nobody was recording the variables.
02 The Approach
The governing rule: every creative is broken into structured attributes by a multimodal model and joined to the attribution and margin spine, so performance is analyzable by attribute, not by id, and the question is always margin-aware, which attributes bring in low-CAC, high-contribution-margin *new* customers, not which ad got the most clicks.
The decision that makes it more than vanity analytics is the join. Creative tagging on its own is a label exercise; tied to the first-party attribution data (which channel and journey actually drove the new customer) and the costed-margin model (what that customer is worth after costs), it becomes a real answer to "what should we make next." We did not stop at engagement metrics. The unit is a profitable new customer.
What we deliberately did not do: no creative generation (this measures, it does not make ads), and no black-box score. The attributes and their performance are inspectable.
03 The Build
A Python pipeline pulls creative assets and copy from Meta, Google, and TikTok. For video, it samples frames with ffmpeg. Claude Sonnet 4.6 (vision) reads the image, the sampled frames, and the ad copy, and tags each creative into a structured taxonomy: format, aspect, hook type in the first seconds, UGC versus studio, on-screen text, product shown, offer, and claim. Those attributes become a creative dimension in BigQuery and are joined via dbt to the first-party attribution data and the contribution-margin model. A Lightdash dashboard surfaces which attributes drive the lowest-CAC, highest-margin new customers, and Claude Opus 4.8 turns the standout patterns into concrete creative briefs for the next round.
It joins directly to the attribution and margin warehouse, so creative is judged on profitable new customers rather than engagement.
04 The Results
Every creative is now tagged across image, video, and copy, and joined to true channel performance and contribution margin. The brand can see which attributes (a hook style, a format, an offer framing) bring in customers who are cheap to acquire and profitable to keep, and the next round of creative is briefed from those patterns rather than from taste. New concepts win against control about twice as often, and the brief-to-test cycle shrinks from weeks to days because the team knows what to make.
05 What's Next
Each round of creative feeds more labeled outcomes back into the model of what works for this brand, so the attribute-to-margin map sharpens over time. The same tagging layer can extend to landing pages and lifecycle creative, and the briefs can become structured inputs to the creative team's workflow.
An LLM reads each retailer's vendor manual into machine-enforceable rules that drive your order-to-cash pipeline, so violations get caught before they ship, not billed back after.
An automation layer over Dentally that pulls intake and referral documents straight into the record, and reads clinical notes to surface and personalize the follow-up that actually books treatment, with the front desk approving every step.
Numbers you can trust, operations that run themselves, all built into your stack and owned by you. Tell us where you want the edge, and we'll build it.
You'll hear from an engineer the same working day, not a sales sequence.
A short call, then a plan, and we start with Proof Week or a full build.
A fixed fee agreed before anything starts, for Proof Week or the full build.