The taxonomy determines what you can learn
If every customer video is labelled 'UGC', the tag may organise files but tell you little about which creative mechanisms work. Useful fields distinguish features that could change a marketing decision, such as the demonstration, proof, offer or call to action.
Define those fields before asking a model to fill them. Clear allowed values and examples make the output easier to compare across assets and over time. Otherwise, apparent patterns can come from inconsistent labels rather than the creative.
Read the average alongside the important fields
The percentage shown here is mean tag-field correctness per output, converted from the published 0–10 score. It is not the percentage of ads with every tag correct. That distinction matters when one wrong field can change the analysis.
For example, a wrongly detected offer could put an ad into the wrong comparison group even when most descriptive tags are right. Check the fields that your next decision depends on, especially brand, product category, format, offer and CTA.
Keep the model and definitions together
The current tagging prompts compare a compact allowed-values contract with a more fully defined taxonomy. The published model ranking combines setups, so it supports model shortlisting without establishing which prompt arm is responsible for the result.
When you deploy tagging, keep the taxonomy version with the output and test a small known set when definitions change. That gives downstream analysis a stable meaning for each field and makes discrepancies much easier to investigate.
How Benmark tests this
Models label image and video ads using a defined taxonomy. The published score averages tag-field correctness per output; critical-field checks distinguish a mostly correct record from one suitable for downstream analysis.
Test inputs, instructions and scoring
What the model receives
The fixed media pack contains five images and five videos. Models return structured tags such as format, brand, product category, offer and CTA. Coverage varies across the published rows, so the results table retains each model's recorded output count.
What changes between prompts
Two prompt conditions are defined. Baseline receives allowed values and the JSON shape; the scaffold adds full definitions and supporting-evidence requirements for specified positive tags. Published model scores currently combine setups, with no complete paired comparison available on the site.
How the result is scored
Each output's score is its fraction of correctly resolved applicable fields, using exact checks and a semantic matcher where specified. The percentage shown here is the published mean 0–10 score multiplied by ten: mean correctness per output, not a pooled field-weighted accuracy or the percentage of entirely correct ads. The critical fields include format, brand, product category, offer and CTA; a high average can still contain a consequential error.