Guide · by Divisadero
How to Analyze Meta Ads with AI
AI can now read ad creative at scale, watch the video, read the claim, judge the hook. Here is the exact, repeatable method to turn a competitor's Meta ads into a creative plan. No gatekeeping: this is the whole recipe.
First, what the Meta Ad Library can and can't tell you
The Ad Library is the richest free source of competitive ad data anywhere, but it exposes creative and runtime, not performance. You can see exactly what an ad says and how long it has run; you cannot see spend, impressions, or conversions. Every honest AI analysis is built on that line: you are reading craft and strategy, and using long runtime as the one real signal that an ad is probably working.
The recipe, in five steps
This is the same pipeline we run, given away in full. The hard part isn't any single step, it's doing all five consistently, for every competitor, every week (more on that below).
Pull the creative
Start from the Meta Ad Library, the public, searchable record of every ad running on Facebook and Instagram. Filter to your competitors and pull each ad's creative: the image or the full video, the primary text, headline, landing destination, and how long it has been running.
How: The Ad Library has a free web UI and an API. For analysis at any scale you want the media files themselves (the video, not just a thumbnail), because the creative is the signal.
The honest caveat: Know the limit up front: the Ad Library shows you creative and rough runtime, NOT spend, impressions, CTR, or ROAS. You can infer that a long-running ad is probably working (brands kill losers), but you cannot read performance directly. Anyone who claims an AI knows your competitor's 'winning ads' from the Ad Library alone is guessing, treat runtime as the one real performance proxy.
Turn each ad into structured data
This is the step AI unlocked. A multimodal model can now watch a 30-second video ad and extract what used to take a human analyst: the hook (first 3 seconds), the core format (problem/solution, testimonial, demo, founder story), the offer, the on-screen claims, the implied target persona, the production style, and the emotional register.
How: Feed the model the media plus the ad text and ask it to fill a fixed schema, the same fields for every ad, so the output is comparable. A consistent taxonomy is what turns 3,000 ads from a pile into a dataset.
The honest caveat: Pin the schema and the definitions. If 'hook' means something different ad to ad, your aggregates are noise. Spell out each field once and reuse it verbatim.
Score the creative
With every ad structured, score it on a few creative dimensions, hook strength, persuasion, execution/production, and credibility (does the claim read as trustworthy, or does it overreach?). A 0-100 scale per dimension makes ads sortable and gaps visible.
How: Give the model an explicit rubric for each dimension and score against it, not vibes. The most useful comparison is craft vs. credibility: when an ad is beautifully made but scores low on credibility, it is quietly losing trust, and that gap is an opportunity.
The honest caveat: These are directional creative-quality reads, not validated performance metrics, and the model has taste, not ground truth. Use them to rank and triage, not to declare a winner. Always keep the score next to the ad so a human can sanity-check it.
Aggregate to find patterns and gaps
The payoff isn't any single ad, it's the distribution. Roll the structured data up: what formats dominate the category, which themes are saturated, where credibility lags craft, and, most valuable, the coverage gaps, the personas, products, and angles competitors are NOT addressing.
How: Count format mix, plot average score by month to see if the category's craft is rising or decaying, and cross-tabulate format against persona to find empty cells, the proven format nobody has aimed at a specific audience yet.
The honest caveat: Gaps are hypotheses, not guarantees. A format/persona combo might be empty because it doesn't work, or because no one has tried it. The data tells you where to look; a test tells you if it's real.
Turn it into a creative plan
Finish by converting the analysis into a ranked list of concrete things to make next: 'launch a 7-day-diary UGC series in the proven swap format,' not 'do more UGC.' Each opportunity should name the action and the evidence behind it.
How: Prioritize by leverage: high-frequency proven formats applied to an uncovered angle beat novel formats with no proof. The output is a short, prioritized test plan, the actual deliverable.
The honest caveat: This is where most one-off analyses die: you build the plan once, ship two ads, and the snapshot is already stale.
Two prompts to start with
Paste these into a multimodal model with the ad attached. Keep the schema identical across every ad, comparability is the whole point.
Structure an ad
You are an ad-creative analyst. Watch this ad and return JSON only:
{ "hook": "<the first-3-second hook, verbatim or described>",
"core_format": "problem_solution | testimonial | demo | founder_story | listicle | other",
"primary_claim": "<the main benefit claimed>",
"claim_has_disclaimer": true | false,
"target_persona": "<who this ad is clearly for>",
"offer": "<promo / guarantee / none>",
"production_style": "ugc | studio | mixed | animation" }Score the creative
Score this ad 0-100 on each dimension, with one sentence of reasoning each:
- hook: how strongly the first 3 seconds earn attention
- persuasion: how compelling the argument/offer is
- execution: production craft and clarity
- credibility: how trustworthy the claims read (penalize unbacked health claims)
Return JSON: { "hook": n, "persuasion": n, "execution": n, "credibility": n, "notes": "..." }The catch: an analysis is a snapshot
You can run this recipe yourself, and you should. But the moment you finish, it starts going stale. New competitor ads launch every week, winners fatigue, and the gaps you found get filled. Doing this once is a project; keeping it current, every brand, every ad, every week, is a pipeline. That continuous version is exactly what we built.
Divisadero runs this pipeline across Health & Wellness DTC brands on Meta continuously, so the analysis is always current and the opportunities are always fresh.
Divisadero · Competitive Meta ad intelligence for Health & Wellness brands. Explore every scored ad at divisadero.co/explore/brands.