How to Choose the Winning Ad Variant

Gather every variant you want to evaluate in one place, and include your current champion as the control. Label each variant with the single hypothesis it represents ('variant B: CTA moved to upper-center,' 'variant C: headline rewritten as specific offer'). Without labels, you lose attribution — when a winner emerges, you need to know what design choice produced it. Aim for 5–10 variants per batch; fewer than 5 doesn't justify the filter work, more than 10 hits diminishing returns.

• →Include the current champion as variant 0 — it's your control
• →Label each variant with the one thing that makes it different
• →5–10 variants is the sweet spot; more than 10 adds noise without adding learning

Upload all candidates to GazeIQ (or any attention-scoring tool) and capture each variant's overall 0–100 score plus the five sub-scores: CTA visibility, headline salience, visual hierarchy, edge avoidance, and clutter penalty. This is your quantitative baseline and the input to every downstream filter. Capture the scores in a spreadsheet with variants as rows and dimensions as columns — you'll reference this through all four filters.

• →Use the same tool for every variant — don't mix scoring systems
• →Capture all 5 sub-scores, not just the overall — filter 2 depends on the sub-scores
• →Score the control champion too, so you have a baseline to beat

Apply the first filter: drop any variant with an overall score below 65 OR any single sub-score below 50. The threshold isn't arbitrary — in our audits, variants below those numbers almost never outperform a control in live testing. The 65 floor catches generally weak creative; the 50 sub-score floor catches variants with a hidden failure mode (for example, an overall score of 72 dragged down by an edge-avoidance sub-score of 48, meaning the CTA is in the safe-zone trap zone).

• →Double threshold: overall ≥65 AND every sub-score ≥50
• →The variant that passes this filter but has any sub-score below 60 goes on the watch list — not the shortlist
• →If your control champion fails this filter, you have a bigger problem than variant selection

Walk through each surviving variant against five psychology principles: (1) one dominant focal point (figure-ground), (2) ≥4.5:1 headline contrast (preattentive processing), (3) mobile-first legibility at 375px viewport (80% of impressions), (4) Gestalt grouping (related elements close, unrelated elements separate), (5) CTA on a natural gaze terminus (after the product or headline, not before). A variant that scores well quantitatively but fails a principle often has hidden fragility — the saliency model flagged the hotspot, but the principle violation predicts poor downstream behavior (dwell, recall, conversion).

• →Apply all 5 principles to each variant; any 'fails' gets demoted or dropped
• →Especially watch for CTA-before-product ordering — a common scoring-model miss
• →Mobile legibility check: shrink the preview to 50% and verify the headline is still readable

Show each remaining variant to 2–3 humans who are NOT on the creative team. Show it for 1.5 seconds (use a phone stopwatch), then ask one question: 'What is this ad selling?' If 2 of 3 can answer correctly, the variant passes. If they can't, the variant's attention architecture works but its message delivery doesn't — and it won't convert regardless of click. This is the cheapest, most-skipped filter in ad testing. It takes 5 minutes and catches the creatives that game scoring models without communicating the offer.

• →Use people outside the team — your designer and PM know what you're selling; your roommate doesn't
• →1.5 seconds is the real scroll window; 3 seconds gives you false positives
• →A 2-of-3 pass rate is the minimum; a 0-of-3 or 1-of-3 pass kills the variant

For each variant that has passed filters 1–3, compute a predicted win probability: what's the chance this variant will beat the control in a live A/B test? A simple method: weight the overall attention score (50%), the psychology check score (30% — number of principles passed out of 5), and the human comprehension rate (20% — 2-of-3 = 67%). Normalize across all survivors so probabilities sum to 100%. The variant with the highest win probability is your top pick; the next-highest is your second pick. Document the math so anyone on the team can reproduce the ranking.

• →Weighting: 50% attention score, 30% principle check, 20% human comprehension
• →Normalize win probabilities across surviving variants so they sum to 100%
• →A variant with <20% predicted win probability rarely earns its test budget

Promote the top 1–2 variants to paid-testing status. One primary (the highest predicted win probability), one challenger if your budget supports it. Archive the rest in a shared creative library with their scores, principle check results, and human test results attached. The log is what compounds — after 5–10 selection cycles, you'll see which sub-scores in your account best predict live CTR for your audience, and the win-probability calculation will sharpen. That compounding is what turns ad-creative testing from an expense into a capability.

• →Top 1 = primary paid test; top 2 = challenger at smaller budget share
• →Don't soft-launch 4 variants simultaneously — you dilute learning
• →Log every rejected variant with the filter that killed it — patterns emerge over cycles

Why not just pick the variant with the highest attention score?

Because attention scores can be manipulated by creative choices that don't translate to downstream conversion. A creative with a giant red arrow pointing at nothing will score well on hot-zone concentration but won't convert. The 4-filter framework combines quantitative scoring (filters 1 + 4) with qualitative checks (filters 2 + 3) specifically to catch variants that game the model. Every filter is cheap; skipping any of them is expensive in the long run.

How is 'predicted win probability' calculated?

A simple weighted composite: 50% from the attention score (normalized 0–1), 30% from the psychology principle check (count of principles passed out of 5, divided by 5), and 20% from human comprehension rate (fraction of human testers who correctly identified the offer). Sum and normalize across surviving variants so probabilities add to 100%. More sophisticated versions train a model on your account's historical pre-test vs live-CTR data, but the simple weighted composite works well from day one and improves as you accumulate cycles.

Can I apply this framework to video ads?

Yes. For video, treat the first frame, a mid-video frame, and the CTA frame as three separate scoring events, then take the minimum of the three (a weak first frame tanks the whole video regardless of later content). For the human sanity check, show the first 3 seconds of the video — that's the real scroll window on Reels and TikTok.

How often should I run a selection cycle?

Frequency depends on creative volume. If you're producing 10+ variants per week, run selection weekly. If you're producing 5–10 per month, run bi-weekly. The discipline matters more than the cadence — running selection on Monday every two weeks, even with fewer candidates, outperforms running sporadic big batches.