Analyze Competitors’ Creatives with Vision AI

Reading five of a competitor’s creatives is easy: open the profile, scan them, jot down notes. Analyzing competitors’ creatives at scale is another story: ten rivals each posting 25 or 30 pieces a month add up to nearly 300 ads, and nobody on the team is going to catalog the theme, offer and palette of 300 images by hand. What happens in practice is familiar: creative analysis turns into sampling (“I looked at a few posts, I think they’re betting on video”) and the brief inherits the guesswork.

Collection is no longer the bottleneck. Public ad libraries and automated profile scanning handle the capture, as we detailed in the guide on how to spy on competitors’ ads. The bottleneck now is the reading: turning a pile of images into an answer to three questions. What is the rival pushing? With what hook? With what look?

This article shows what a computer vision AI extracts from each creative after the collection happens, how those fields become a comparable pattern, and where automated analysis still doesn’t replace a trained eye. If you’re still building your ad database manually, start with the step by step for competitive analysis with the Meta Ads Library and come back here: the topic now is what to do with the material you’ve collected.

What the caption doesn’t say (and the artwork delivers)

A caption is cheap text: short, generic, easy to swap. The real creative decision lives in the image. A lingerie store writes “new arrivals” in the caption and prints “3 items for R$99” on the artwork. A supplements brand posts “a hard workout deserves a reward” while the image carries a coupon badge, a free shipping banner and the product tub taking up half the frame. Anyone analyzing only the captured text misses exactly what matters:

• Price and terms printed on the artwork, which rarely show up in the caption.
• Which product is featured: a single hero, a bundle, or a lifestyle scene with no product at all.
• The presence of a human face: a creator and a customer on screen tell one strategy; a studio still tells another.
• The campaign’s visual code: the palette that screams clearance is recognizable before any word.

That’s why ad analysis based only on text and hashtags stays shallow. The image is the channel where the competitor communicates the offer; ignoring it means reading half the message.

What the AI extracts when analyzing competitors’ creatives

In Batedor, every collected creative runs through a vision model that returns structured fields, always the same ones, for any image. That’s what makes 300 pieces comparable to one another: instead of 300 opinions, you have 300 rows with the same columns.

In the panel, each analyzed creative appears as a card: the image, the theme and mood badges, the CTA and face tags when detected, the dominant colors and the one- or two-sentence summary. None of this requires opening the original post: the structured read is already done by the time you arrive.

Two layers: the campaign’s type and the campaign’s look

Vision doesn’t work alone. Before it, every detection already goes through automatic campaign classification, which sorts the finding into 16 types: discount coupon, free shipping, stock clearance, product launch, giveaway, installment payment, limited-time offer, among others. That layer reads the content of the detection; vision reads the artwork. The value shows up at the intersection:

• Stock clearance + urgent mood + text taking up half the artwork: a real clearance, with the terms printed on it. It deserves a commercial response.
• Product launch + Lifestyle theme + zero CTA: warm-up phase. The rival is seeding; the offer comes later, and you have time to react.
• Coupon in the classification + Lifestyle artwork with a face: the discount is hidden in the caption or on the creator’s profile. A bet on influence, not on a billboard.

Sorting these combinations by hand would mean reading each piece twice. With the two automatic layers, the triage arrives done and your time goes to the decision.

From the single creative to the pattern: the visual playbook

A single piece is misleading. What defines a competitor’s creative direction is the distribution: of everything they’ve published, how much is offer, how much is brand, how much is education? In the Batedor panel, each competitor’s Visual playbook page aggregates the analyzed creatives and answers all at once: dominant theme, percentage of pieces with a human face, percentage with a call to action, average on-artwork text, aggregated dominant palette and the recurring visual vocabulary. It updates daily, as new creatives come in.

Two example reads, with very Brazilian categories. For a supplements brand: Promotional theme in 7 out of 10 ads, urgent mood, almost every piece with a CTA and a high average of on-artwork text. That’s a performance operation fighting for clicks; competing in the same visual code leaves you indistinguishable, and the cheap opening might be the educational content it doesn’t make. For a skincare brand: Lifestyle theme, warm mood, a face in almost every ad and very little text. This one is buying brand, and probably only pushes an offer on seasonal dates; the turning-point signal appears when its aggregated palette darkens and the CTA percentage rises.

That’s the second use of the playbook: detecting a change of direction. A rival that spends weeks in neutrals and suddenly concentrates black, yellow and an urgent mood is heating up for a date. If that happens in October, you know what’s coming in November.

How this changes your brief

A brief with no competitive data turns into a catchphrase (“we want something impactful”). A brief with a visual playbook turns into a testable hypothesis. Take a Father’s Day brief written in July:

1. Relative positioning: if all three direct rivals are on urgent Promotional, fighting over the same billboard is an auction. Either you win on the offer, or you contrast on purpose (educational, a face, a warm tone).
2. Visual hook: if nobody in the category uses a human face and UGC barely shows up, a creative with a real customer is cheap open space to test.
3. On-artwork text benchmark: if the category runs clean artwork and the fastest-growing rival prints the price, test price on the artwork before ruling it out for aesthetics.
4. Timeline: the rival’s shift in palette and mood weeks before the date is the trigger to bring your production forward, not to copy it.

One caveat that keeps the brief honest: the playbook shows what the competitor is trying, not what’s working for them. You don’t see anyone’s ROI, budget or conversion. The read is about creative direction, and direction becomes your hypothesis to test, not a ready-made answer.

What automated analysis doesn’t do

Real limits, so you can calibrate expectations before turning it on:

• Video isn’t watched: for Reels and videos, what enters the analysis is the cover frame, not the whole script.
• On-artwork text is an estimate: the percentage is approximate, good for comparing pieces to one another, not for measuring pixels.
• The AI is sometimes wrong: an ambiguous creative can land in “Other” or get a debatable theme. The aggregated pattern dilutes those errors; the single piece doesn’t.
• No performance metrics: the rival’s spend, reach and conversion stay invisible, here and in any tool.
• LGPD scope: only public content from the monitored profiles enters the analysis.
• Availability: vision analysis is part of the Professional plans and up (details on plans).