> For the complete documentation index, see [llms.txt](https://connectyai.gitbook.io/documentation/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://connectyai.gitbook.io/documentation/visual-intelligence.md).

# Visual Intelligence

This page explains how Connecty's Visual Intelligence works under the hood: what it reads, how signals are extracted, how they're joined to performance and audience data.

### What it analyzes

Visual Intelligence reads every visual surface your customer sees, across three layers:

1. **Video Ads Intelligence**, at the frame, scene, and beat level, for ads running on Meta, TikTok, and YouTube.
2. **Image Ads Intelligence**: composition, headline and hook text, product framing, offer, social proof, and CTA.
3. **Landing Page** **Intelligence**: each ad clicks through to: message match against the creative's claim, offer continuity, and the conversion path.

The unit of analysis is not the ad. It's the moment inside the ad, joined to the outcome that follows it.

### Data inputs

Analysis runs on the creative assets themselves plus platform-reported performance data from your connected ad account. Connecty is read-only: it reads your account, it never writes to it. There is no Connecty pixel and no tracking tag on your site or in your ads.

* **Self-Serve**: Meta ad account via Facebook login. Creative, performance, and delivery data at the campaign, ad set, and ad level.
* **Enterprise**: adds revenue and customer context from Shopify, Stripe, GA4, and WooCommerce, which extends the analysis from engagement and conversion into retention, repeat purchase, and LTV cohorts.

Time to first output: connecting takes about five minutes; analysis starts immediately, and the first complete recommendations typically arrive within a few hours depending on account size and historical data volume.

### Example of video signals

| Layer               | What's extracted                                              |
| ------------------- | ------------------------------------------------------------- |
| Hook & Opening      | Hook mechanism, hook family, hook strength, opening frame     |
| Visual & Production | Scene-change rate, product visibility, human presence, pacing |
| Narrative & Message | Persuasion structure, cognitive load                          |
| Proof & Trust       | Proof points, trust signals, and when they appear             |
| CTA & Conversion    | CTA presence, type (verbal, visual), and timing               |
| Performance         | The layer that joins all of the above to outcomes             |

### Contextual role decoding

Most video AI stops at entity detection: a face appears at 0:03, the product is visible in 40% of frames. Connecty adds an interpretation layer on top. It decodes what each entity is doing in that specific video and whether that moment drives retention or causes drop-off.

The difference shows up in the questions you can answer. Entity detection answers "in which videos is my red scarf visible?" Connecty answers "in which videos is the red scarf acting as the hook, or as a key moment that improves retention?" Same object, different insight. One tells you it's on screen; the other tells you whether it's earning its screen time.

### Retention mapping

Every creative moment is timestamped and mapped against audience retention: first face, first product, first text, first benefit, brand mention, first proof, verbal CTA, visual CTA. This pinpoints the micro-moment where viewers drop or commit, so a diagnosis names a timestamp, not a vibe.

### Statistical grounding

No generic benchmarks. Every signal is judged against your own top performers, inside a matched cohort: same optimization goal, same objective, same funnel stage. Comparisons use within-cohort statistics (z-scores with empirical shrinkage) so that low-volume ads don't produce false winners and high-spend ads don't drown out real signals. Recommendations that involve trend claims are gated with statistical trend tests before they surface.

Benchmarks are living: as new data lands, cohorts are recomputed and every ad is re-benchmarked against your evolving top performers.

### Audience and funnel-stage context

Connecty derives each ad's audience and funnel-stage context (TOFU, MOFU, BOFU) from ad-set targeting, campaign structure, and available delivery breakdowns. Every creative signal is then evaluated inside that context, because the same signal means different things at different stages: a weak hook at TOFU, where you're fighting for cold attention, is fatal; the identical hook at BOFU is often fine. The same creative is evaluated separately for prospecting, returning, and ready-to-buy audiences.

### Performance profiles

Each ad is classified into a performance profile within its matched cohort. The profile names where the engagement-to-conversion-to-retention journey breaks, so the recommended fix matches the failure mode:

* **Weak Hook**: viewers never engage; the fix is a re-cut of the opening.
* **Clickbait Drop-off**: strong hook, sharp retention collapse after it.
* **Missing CTA**: engagement without a conversion path.
* **Landing Page Leak**: strong ad, mismatched page; the fix is page-side, not a wasted re-cut.
* **Impulse Clicker**: clicks that don't convert or convert poorly.
* **Ugly Winner**: weak engagement metrics, excellent customers and LTV.
* **Scaling Winner**: performing and ready for more budget.

Profiles matter because several of these look identical in a dashboard while demanding opposite actions.

### Outputs

**Beat-by-beat re-cut briefs.** The ad broken down beat by beat with timestamps. Each beat gets its category, the spoken line, a keep, cut, or rewrite verdict, and a concrete script-plus-visual rewrite. Every rewrite carries a confidence tier (High, Medium, or Test) and is benchmarked against your own top performers for that audience and stage. The output format is a creator brief, ready to send.

**Creative generation.** Connecty generates images and precise video scripts built from your winning formula per funnel stage. Generation is always evidence-first: diagnosis of why an ad underperforms, then a draft of the fix. Never from a blank prompt.

**Ranked recommendations.** Visual Intelligence findings feed the same recommendation stream as the rest of Connecty: ranked by impact, with evidence and reasoning attached, delivered daily without prompting. Every recommendation waits for your review; your team decides and implements.

### Human-made and AI-generated video

The analysis is identical regardless of how a video was produced. Most analyzed content today is human-made: creator UGC, founder videos, and edited variations. AI generation raises the volume of creative brands test, which raises the value of knowing which creatives to make more of.