Will your ad perform? Grade it in 10 seconds

How the score is calculated

The grader runs two models on your image, and they do different jobs. The first one produces the number. The second one explains it.

1. A vision model reads the image. Your upload is scaled down to at most 1,024 pixels and passed through CLIP ViT-B/32, a pretrained image encoder whose weights we never change. It turns the picture into an embedding: a list of numbers that describes what is in it, from faces and products to colors and text blocks.
2. A small model predicts engagement. A gradient boosted model (LightGBM), trained on 13,622 real Meta ads and their real engagement, maps that embedding to a predicted engagement rate. It sees only the image. It does not know your brand, budget, targeting or the ad text outside the picture.
3. The prediction becomes a percentile. Your score is where that prediction ranks against the 13,622 ads in the training set. A score of 70 means your creative is predicted to beat 70% of them.
4. A second model writes the breakdown. A vision language model looks at the same image and lists what works and what to fix. It is told the score band but cannot change the score, and the score still comes back if the breakdown fails.

The model is called Treviso. We validated it the hard way: held out by account with 5-fold cross validation, so every reported number comes from ads and advertisers it had never seen. Budget and targeting context alone predicted almost nothing (0.53 pairwise, close to a coin flip), while the creative lifted ranking accuracy to 0.78. A sibling model we trained to predict conversion cost barely beat random (0.59), which is why this page scores engagement, the thing the image controls, and never claims to predict sales.

A worked example

Say you run a restaurant and you have two versions of a weekend brunch ad.

Version B is predicted to out-engage 71% of the corpus; version A only 38%. The breakdown for A would point at what the score is reacting to: no single focal point, text too small to read on a phone, and nothing human in the frame. If you only have budget to test one, B is the better bet. If you can test both, run them in the same ad set and let Meta's delivery confirm it.

How to read your result

• 80 and above, top performer. The image already does what high-engagement ads do. Spend your effort on the offer and the landing page.
• 60 to 79, strong. A good creative. The fixes listed are usually about clarity: fewer words, a bigger product, a face.
• 40 to 59, average. It will deliver, but you are likely paying more per click than you need to. Try one or two of the fixes and grade again.
• Under 40, below average. Rework the image before you spend on it. Start with the first fix in the list.

The most useful way to use the grader is comparative. Grade three or four variants of the same ad, launch the top two, and keep the winner. Pair it with the Facebook ad mockup generator to see each variant in the feed before you grade it, and with the ad hook generator when the problem is the message rather than the picture.

Limits worth knowing

The score ranks predicted engagement: likes, comments, shares and clicks relative to reach, compared with real Meta ads. It is not a click-through rate, a cost per result or a sales forecast. It reads still images only, so for a video ad, grade the opening frame. It knows nothing about your audience, so an ad for a niche product can score lower than it will actually perform with the right people. And like any model trained on past ads, it rewards what worked before; a genuinely new format can surprise it.

When the score tells you an ad needs work, the AI ad generator can build new versions from your website, and recreate ad creatives remakes an ad you already like in your own brand. If you are ready to run it, the budget calculator tells you what a test should cost in your industry.