How Knack reads a creative

Most creative reporting tools read your ad names. Knack reads your ads.

That sounds like a small distinction. It decides whether the analysis survives contact with a real account.

Why naming conventions break

The standard approach is to parse structure out of the ad name: UGC_Testimonial_Hook3_Sarah_v2. It works beautifully in a demo and degrades the moment real people touch the account.

It degrades when a freelancer joins and uses a different scheme. When an agency hands the account over and the new team inherits six months of someone else's shorthand. When someone ships twelve variants on a Friday afternoon and names them final, final2, final_USE_THIS. Within a quarter, the taxonomy that the entire analysis rests on is half-fiction.

And the failure is silent. The tool doesn't tell you it stopped understanding your account. It keeps producing confident charts about categories that no longer mean anything.

What Knack does instead

Knack pulls the actual creative — the image, or a frame from the video — and looks at it. A vision model tags what is visibly there across nine dimensions:

The instruction to the model is explicit about this: tag what you can see — who is on screen, where it was shot, how it is cut, what the copy is doing — and do not infer from the ad name, because the names are unreliable and often inherited from another agency.

Two things that don't come from the model

Not everything should be inferred, and pretending otherwise is how tools end up confidently wrong.

Owned versus partnership is derived from the ad object itself — the Instagram account the creative was published from. That is a fact, not a judgement, so asking a model to guess at it would only add error. It is also available immediately, before any tagging has run.

Funnel stage comes from the campaign objective. An ad in an awareness campaign is judged on hook rate, hold rate and CTR. An ad in a sales campaign is judged on ROAS, CAC and CVR. This matters more than it sounds: an upper-funnel ad has no ROAS by design, and a tool that ranks everything on one leaderboard will quietly recommend killing the top of your funnel.

From tags to a brief

Tags on their own are trivia. The point is what they let you group.

Once every ad carries structured attributes, performance can be aggregated by attribute rather than by ad. The question stops being "which of these forty ads worked" and becomes "which kind of ad worked" — which is the only version a creative team can act on, because you can brief an angle and you cannot brief an ad ID.

Typical outputs: talking-head founder content outperforming creator UGC at the bottom of the funnel on this account. Every winner shot in a real location, every loser in a studio. Offer-led angles winning on CTR and losing on CVR, which is a landing page problem rather than a creative one.

What it deliberately won't do

Knack won't tell you an ad is good on forty impressions. Creatives below a spend and volume floor are flagged too early to call rather than ranked — a cheap winner on thin data is held back instead of promoted, because the evidence for most creative decisions is much weaker than it feels.

It won't assign a single quality score either. Every flag is a specific, checkable condition — this frequency, that CTR gap, this trend over that window — and at most one is raised per axis, so an ad is never simultaneously described as gaining and fatiguing. You can hover any flag and see the rule and the numbers that produced it. A score you can't interrogate is a number you have to take on faith, and creative decisions are too expensive for faith.

And it won't claim to measure incrementality. Platform-reported conversions are directional. Knack tells you which ads are doing the work relative to each other; whether the channel is profitable is a different question needing a different instrument, and the two get conflated often enough to be worth saying out loud.