Here is the problem with luxury advertising in one number: fewer than 1% of site visitors purchase on their first session when the average product costs €1,000 or more. Traditional campaign optimization relies on conversion data to train bidding algorithms. But with 20-50 purchases per week, there is not enough signal. The algorithm starves. It bids blindly. Budget bleeds.
We solved this with a custom AI predictive model for DV360 programmatic campaigns. Instead of waiting for the rare purchase event, we scored thousands of daily micro-conversions, the small behavioural signals that precede a purchase, and fed those predictions directly into the bidding engine.
The result, verified in a controlled A/B test with equal budget allocation: +106% revenue. +100% ROAS. +36% sales. -25% CPA (Source: Making Science / Google).
Same budget. Same audiences. Same creative. Twice the revenue. The only variable was better signal.
Why Google's Bidding Breaks at Luxury Price Points
Google's automated bidding is brilliant at scale. Feed it thousands of weekly conversions and it will find patterns humans cannot see. The machine learning needs volume. It is a statistical engine.
Luxury does not have volume. A campaign generating 30 purchases per week at €1,500 average order value gives the algorithm almost nothing to learn from. It cannot distinguish between a serious buyer spending 12 minutes examining a €3,000 coat and a casual visitor who bounced after 8 seconds. Both look roughly the same in conversion-sparse data. So the algorithm bids conservatively everywhere, treating all visitors as medium-probability prospects.
The common workaround is broadening the conversion definition. Count add-to-cart as a conversion alongside purchase. But this introduces noise rather than signal. Someone adding a €200 accessory to cart has fundamentally different intent than someone adding a €5,000 watch. Treating both as equivalent "conversions" degrades bid quality rather than improving it. We tried this approach before building the predictive model. It did not work.
The Micro-Conversion Scoring Architecture
The insight behind the model is straightforward. Luxury customers telegraph intent before they buy. They view products in the same brand repeatedly. They check the size guide. They add to wishlist. They return to the same product page across multiple sessions. Each action carries different predictive weight.
We built the model in collaboration with Google and Making Science. It assigned weighted scores to behavioral signals:
Product views received a base score, modulated by the price tier of the product viewed and time spent on the page. Under 5 seconds got a low weight. Over 45 seconds got a high one. Add-to-cart events scored higher, further weighted by product value and cart persistence across sessions. A product that stayed in cart for 72 hours signaled stronger intent than one removed within the hour.
Wishlist saves, size guide interactions, return visits to the same product, and cross-category browsing within a single brand all fed the composite score. The weighting was not arbitrary. We calibrated against 18 months of purchase history, running regression analysis on Google Cloud BigQuery to identify which signal combinations most reliably preceded actual transactions. The processing volume was substantial: millions of event rows per day, cross-referenced against purchase outcomes over rolling 90-day windows.
The first version of the model failed. Badly. We had overweighted add-to-cart events based on the assumption that cart behavior in luxury mirrors mass retail. It does not. Luxury customers use the cart as a bookmark. They add €8,000 worth of products with no intention of buying all of them. The cart is a consideration list, not a purchase commitment. Our initial model scored these users as high-propensity and the algorithm bid aggressively on them. Conversion rates did not improve. Cost went up.
The recalibration took three weeks. We downweighted cart events and introduced cart persistence as the differentiator: a product that stayed in cart for 72+ hours across multiple sessions was a genuine purchase signal. A product added and removed within 24 hours was not. That single adjustment was the largest contributor to the model's eventual performance. The lesson was humbling. The algorithm is only as smart as the assumptions baked into the training data. Getting those assumptions wrong produces confidently wrong bidding.
The composite score was then fed to DV360 in real time. Instead of optimizing for 30 weekly purchases, the bidding algorithm now optimized against thousands of daily high-propensity signals. Suddenly it had enough data to bid intelligently.
| Signal | Weight Factor | Why It Matters |
|--------|--------------|----------------|
| Product page view > 45s | Medium | Time investment signals genuine interest, not casual browsing |
| Size guide interaction | High | Checking size indicates purchase consideration, not just aspiration |
| Add-to-cart, persists 72h+ | Very High | Cart persistence across sessions is the strongest non-purchase signal |
| Wishlist save | Medium-High | Conscious intent to return, weaker than cart but stronger than view |
| Return visit to same product | High | Repeat consideration within 7 days predicts conversion within 30 |
| Cross-category within brand | Medium | Exploring a brand's range indicates brand affinity, not just product interest |
| Pageview under 5s, single page | Negative | Bouncing quickly reduces propensity score, filters casual traffic |
The A/B Test
Claims about AI effectiveness are easy to make. Proving them requires discipline. We ran a properly controlled experiment: equal budget, same audiences, same creative, same timeframe. One group received AI-optimized bidding based on propensity scores. The control group received standard DV360 bidding.
The AI group delivered +106% revenue. Not +10%. Not +25%. Double.
ROAS improved by 100%. Sales volume increased by 36%. CPA dropped by 25% (Source: Making Science / Google). These results came exclusively from better signal processing. No creative changes. No audience expansion. No budget increase. Just a bidding algorithm that finally knew who to pursue.
One detail worth noting: the model did not just find more converters. It found higher-value converters. Average order value in the AI-optimized group was meaningfully higher than in the control group. The propensity scoring, by weighting signals like price tier of products viewed and cross-category brand exploration, naturally biased the system toward customers with luxury purchase patterns rather than just any converter.
Building the Model: What We Learned
After deploying and refining this approach across campaigns managing tens of millions of euros in annual ad spend, several principles emerged that determine success or failure.
Data Pipelines Before Data Science
A simple model on clean data will outperform a complex model on messy data. Every time. The first phase of any predictive advertising project should be data pipeline architecture. Every micro-conversion event captured accurately. Every timestamp correct. Every user identity resolved across devices. We spent more time on data infrastructure than on model development. That ratio was correct.
Luxury-Specific Signal Weighting
Generic propensity models trained on mass retail data fail in luxury. Purchase consideration periods are weeks or months, not hours. Higher price can increase desirability rather than suppress it. Brand affinity is a stronger predictor than category browsing. Return visit cadence follows different patterns.
We built custom weighting that reflected these realities, and recalibrated quarterly as seasonal purchase patterns shifted. A model trained on holiday shopping behavior performs poorly during resort season. Luxury has more behavioral variation across seasons than mass retail because the purchase motivation varies more (gifting vs. self-purchase, vacation wardrobe vs. office wardrobe).
Real-Time Scoring Changes Everything
A propensity score that updates daily is useful. One that updates in real time is transformative. The difference: catching a high-intent user during their active browsing session versus bidding on a stale signal 24 hours later. We invested heavily in real-time data tracking that fed DV360 within minutes of user actions, not in overnight batch processing. The latency difference translated directly to conversion rate.
Always Validate With Equal-Budget Testing
Never trust a model without controlled validation. We have seen predictive models that looked impressive in backtesting and failed in production because the training data contained biases invisible until the model encountered real traffic. The discipline of equal-budget A/B testing is non-negotiable. It is the difference between science and marketing.
Privacy and the First-Party Data Advantage
Does predictive advertising still work as cookies disappear and privacy regulations tighten? Yes. Actually better.
The micro-conversion signals that feed this model (product views, add-to-cart, wishlist saves, size guide clicks) are all first-party data. Collected on the brand's own domain. Subject to the brand's own consent framework. Independent of third-party cookies, cross-site tracking, or any mechanism that privacy regulations restrict.
What changes is activation. Instead of syncing propensity scores to broad programmatic networks via cookie matching, we feed signals through privacy-safe channels: Google's Privacy Sandbox APIs, server-to-server integrations, and first-party audience segments uploaded through Customer Match. The predictive advantage does not diminish. The plumbing changes.
Brands that built their analytics infrastructure on first-party data are better positioned for predictive advertising than those that relied on third-party data ecosystems. The privacy transition is accelerating the advantage of owned data.
When This Approach Makes Sense
Not every brand needs a custom predictive model. The investment is justified when:
Average order value exceeds €500, making each conversion valuable enough to warrant sophisticated targeting. Product catalog exceeds 1,000 SKUs, creating enough complexity that the algorithm needs help deciding what to show whom. Monthly site traffic exceeds 1M visits, generating sufficient micro-conversion volume for reliable predictions. Annual advertising budget exceeds €500K, where efficiency improvements produce meaningful absolute revenue gains.
Below these thresholds, Google's native Smart Bidding and Performance Max campaigns are the right starting point. The jump to custom models should happen when native tools have been optimized to their ceiling and the incremental revenue opportunity justifies the build.