Creative pre-testing with predicted visual attention & ad-effectiveness scoring

Testing an ad, pack or landing page with an eye-tracking or survey panel takes weeks and costs thousands per creative, so most assets ship untested. Saliency models trained on large eye-tracking datasets predict an attention heatmap in seconds, measure whether the logo and message are noticed, and flag cluttered layouts. Marketers can compare variants before spend, including SMBs without a research budget. It is a prediction of perception, not a measurement of sales.

Problem & value

• Today: creative quality is judged by taste, agency opinion or, for large brands, panel-based pre-tests (eye tracking, surveys, facial coding, EEG) that take weeks. Media budgets are then spent on creatives whose logo is overlooked or whose headline is never read.
• AI value: results in seconds per asset, so many variants can be screened and design iterated before production is finalised. Vendors position this as a cheap first filter and a way to decide which few assets deserve a full panel test.
• Evidence level: Neurons reports its attention model reproduces eye-tracking results with over 95% accuracy, trained on a database from more than 100,000 participants [1]. These figures are vendor-reported; no independent benchmark on real ad creatives was found. Academic saliency benchmarks (MIT/Tuebingen) measure fixation prediction on natural images and are the better yardstick for the underlying technology [5][6]. Attention is a necessary but not sufficient condition for persuasion, so treat scores as diagnostics.
• Distinct from: REC-007 optimises with live campaign data; this predicts human perception with no audience data. GEN-001 / GEN-004 generate assets but do not evaluate them.

Variants

• Predictive attention tools (Neurons, Dragonfly AI, Attention Insight, Brainsight, EyeSee): upload an image or video, get heatmaps, AOI shares and clarity scores; the core variant.
• Brand-level ad-effectiveness scoring (Kantar LINK AI and similar): models trained on large archives of tested ads that predict outcomes such as brand impact, from the creative only [4].
• Packaging and shelf pre-testing: render a pack in a simulated shelf and predict standout.
• Web and app UX: predicted attention on landing pages and e-mail templates (compare VIS-025 for video tagging).
• Out-of-home and retail media: check legibility at distance for billboards, DOOH, in-store displays.
• SMB self-service: a Canva or Figma plug-in check before posting; free or low-priced tiers.

How it works

1. creative (image / video / mock-up)
2. pre-processing (resize to device/viewing context, frame sampling for video)
3. saliency model (CNN/transformer trained on fixations)
4. heatmap + scan-path estimate
5. AOI detection (logo, face, text, CTA via object detection + OCR)
6. metrics (AOI attention share, time-to-first-notice, clarity / cognitive load)
7. benchmark vs. category norms and previous winners
8. recommendation + optional VLM critique text
9. human creative review
10. optional panel validation of the shortlisted variant
11. launch

• Saliency model: a deep network trained on eye-tracking fixations, combining bottom-up features (contrast, colour, edges, faces, text) with learned top-down priors; output is a probability map over the image. DeepGaze-class models are the open reference [6].
• AOI and brand detection: object detection and OCR locate logos, faces and text; Neurons describes automatic AOI detection for brands, text and faces [1].
• Scoring layer: weights attention metrics into an index (clarity, focus, brand visibility); the weighting is vendor-proprietary and should be calibrated against your own KPIs.
• VLM critique: a multimodal LLM can turn the numbers into design advice. It does not predict gaze itself and can hallucinate, so keep the saliency model as ground for the numbers.
• Human-in-the-loop: creative lead and brand manager decide; tools suggest. Calibration against a small panel test on 10-20 past creatives is recommended before trusting a score threshold.

Data preparation

The main inputs are the creatives themselves (images, video, mock-ups) in the DAM, on file shares, in agency mails or in design tools, plus for calibration the results of past campaigns (click rates, recall, brand lift from agency reports or Excel) and brand guidelines. Typical gaps: only final exports in odd sizes, no link between a creative and its campaign result, several versions under inconsistent names, no channel and format context (mobile feed, poster) and no agreed category benchmarks.

Preparation steps: - Collect the last 10-20 creatives with measured outcomes as a calibration set (PREP-018, PREP-020). - Name and tag assets with campaign, channel, format and version in the DAM (PREP-010, PREP-009). - Export campaign results into one tidy table and join them by asset ID (PREP-011, PREP-005). - Define which elements count as areas of interest (logo, CTA, product, headline) once, with the brand team (PREP-019). - Keep a small gold set of well-performing creatives as reference (PREP-016). - Check that uploads are not used for vendor training.

Effort: low, about 10-20 percent of the project; the calibration step is the one that matters. See GUIDE-003 for a quick check.

Implementation paths

Which path for whom

• smb: A - attention-prediction SaaS from EU vendors delivers heatmaps and scores within days for a standard process, with no ML staff needed.
• enterprise: B - integrating scores into the DAM and creative-ops workflow and combining them with a VLM critique and own brand benchmarks calls for APIs and orchestration.
• developer: C - open saliency models (DeepGaze family) and open VLMs run on a single HW2 machine; accuracy on ad layouts is lower, so calibrate on own data.

Path A — Buy (recommended)

• Attention-prediction SaaS: Neurons Predict (Denmark, EU), Attention Insight (Lithuania, EU), Dragonfly AI (UK), Brainsight (Canada), EyeSee (Italy/Netherlands, EU; combines panels with AI) [1][2][3].
• Ad-effectiveness benchmarking: Kantar LINK AI (UK/global, with large tested-ad norms) [4]; similar offers come from other research agencies.
• Choose A when: you need results this week, have no ML team and assets are mostly images or short videos. Check: EU data processing, whether uploads are used for training, API access, brand-library and team seats, validation evidence provided.

Path B — Build on cloud APIs

• Vendor APIs (Neurons, Attention Insight and others offer API or bulk upload) for pipeline integration into DAM or creative-ops tools.
• Multimodal LLMs (Claude, Gemini, GPT, Mistral Pixtral; EU regions where offered) for critique, message-clarity checks and "what would a first-time viewer notice" prompts, combined with a saliency model for the numbers.
• Orchestration with n8n or a small service; store scores per asset ID in the DAM.

Path C — Self-host (open source)

• Saliency: DeepGaze IIE / III (PyTorch, open code), SALICON/MIT1003-trained models, and pysaliency for benchmarking; video saliency models such as TASED-Net or ViNet [5][6].
• AOI: YOLO-class detectors or open-vocabulary detectors (Grounding DINO, OWLv2) plus PaddleOCR for text; open VLMs such as Qwen-VL for critique via vLLM.
• Hardware: HW2 is ample; saliency inference is light, so even HW1 works for images.
• Caveat: open models are trained on natural scenes, so ad-specific behaviour (text, logos, layouts) is weaker than in commercial models trained on ad-relevant data.

Path D — Custom training

Worth it only for agencies or brands with proprietary tested-creative archives: fine-tune a saliency model on your own eye-tracking sessions, or train an outcome model (tabular or CNN features → recall, brand lift) on historical pre-test and campaign results. Needs hundreds to thousands of creatives with measured outcomes and consistent test protocols.

State of the art (2026)

• Saliency prediction is mature on natural images; benchmarks such as MIT/Tuebingen track progress on fixation metrics (AUC, NSS, KL) [5]. Commercial tools extend this to ads, packaging, video and web, with AOI-level reporting.
• Trend: vendors add generative features, such as AI-written improvement suggestions and variant generation in the same tool, linking to GEN-001 / GEN-004 and creative-ops platforms.
• Limits and failure modes: (1) predicts free-viewing or generic task attention; real context (scrolling speed, sound, prior brand knowledge, intent) changes gaze; (2) strong at "what is seen", weak at "what is understood or liked"; (3) text, faces and high contrast dominate, so unusual designs may be mispredicted; (4) cultural and reading-direction differences (e.g. RTL scripts) are under-represented in training data; (5) vendor accuracy claims are self-reported; (6) effectiveness scores trained on past ads favour conventional creative and may penalise novelty.

Shared resources & combinations

• Same stack (object-detection + ocr-docparse + vlm): VIS-035 brand and logo detection, VIS-022 visual search, VIS-025 video tagging, VIS-018 industrial OCR.
• Upstream: GEN-001 marketing imagery, GEN-004 banners and layouts, GEN-002 video: generate many variants, screen them here.
• Downstream: REC-007 ad optimisation, which validates predictions with live results; feed A/B-test outcomes back to calibrate scoring.
• Infrastructure: one HW2 GPU server can serve saliency, detection and a VLM together.

Effort, cost & KPIs

• Path A: hours to days; subscription pricing (tiers by seats or assets; check vendor pages).
• Path B: 2-6 weeks to integrate with DAM and review workflow.
• Path C: 1-3 months incl. calibration; low GPU cost, main cost is engineering and validation.
• KPIs: correlation between predicted AOI attention and later panel or A/B results; share of assets screened before launch; creative iterations per asset; time to approval; post-launch brand recall or CTR uplift versus untested baseline; false-pass rate of low-scored assets.

People & change

Involve the marketing or brand lead (process owner and sponsor), the creative team and any external agency (suppliers), plus IT for DAM access and SaaS accounts. Management only needs to approve a small budget. Position the tool as a fast second opinion before expensive media spend, not as a judge of creative work; designers who help set up the benchmark will accept it more easily.

Works council: no in the usual case; the tool scores ads, not employee behaviour or performance. If scores are used to rank individual designers, the picture changes (depends) and §87(1) no. 6 BetrVG should be checked; a short information of the works council under §90(1) no. 3 BetrVG is good practice (GUIDE-004).

Data protection: the creatives may show people, but they are normally models with signed releases; no GDPR personal data of staff is processed. Check the vendor AVV and that uploads are not retained for training (GUIDE-005, GUIDE-010). AI literacy: explain that the heatmap is a prediction, not measured gaze, and that a VLM critique can hallucinate; AI Act Art. 4 asks employers to support staff literacy (GUIDE-006). Run a 4-week pilot on 10-20 past and current creatives and compare scores with real results before setting any threshold (GUIDE-009). Typical concerns: "it will flatten creativity", "scores are arbitrary"; keep the creative lead as final decider and review score thresholds regularly (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: minimal risk. The tool analyses creatives, not people, and no Annex III category applies; AI literacy (Art. 4) applies to staff [7]. Beware adjacent variants: panel add-ons that infer emotions from webcams of employees are an Art. 5 prohibition at the workplace; emotion recognition of consumers is otherwise high-risk or transparency-bound. Predictive attention on uploaded assets does not do this.
• GDPR: generally no personal data if assets contain no identifiable people; ads with real models or user-generated content may contain faces, so check the vendor's DPA, hosting region and training-use clauses. Panel-based add-ons (webcam eye tracking) need consent.
• Advertising law: scores do not replace compliance review (UWG, misleading-advertising and influencer-disclosure rules).
• Business risks: over-reliance on a proxy metric, optimising to "looks salient" instead of persuasive, homogenised creative, vendor lock-in on proprietary scores.
• Mitigations: calibrate on own data, keep a panel or A/B step for big bets, document thresholds, treat scores as one input among several.

Getting started

1. Pick 20 past creatives with known results (good and weak ones) and run them through 2 tools' trials.
2. Compare predicted attention to what you know: does the logo score match recall results?
3. Define AOIs and a minimal checklist (logo seen in 3 s, CTA visible, one focal point).
4. Add the check to the creative workflow before approval and log scores with campaign results.
5. After one quarter, review correlation with KPIs and adjust or drop thresholds.

Sources

1. Neurons — How Predict (attention AI) works and validation — https://knowledge.neuronsinc.com/how-neurons-ai-works
2. Dragonfly AI — predictive attention platform — https://dragonflyai.co
3. Attention Insight — AI attention heatmaps — https://attentioninsight.com
4. Kantar — LINK AI creative testing — https://www.kantar.com/campaigns/link-ai
5. MIT/Tuebingen Saliency Benchmark — https://saliency.tuebingen.ai
6. DeepGaze (open-source saliency models) — https://github.com/matthias-k/DeepGaze
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Brainsight — predictive attention tools — https://www.brainsight.app