In September 2023, our data science team made an observation that changed the direction of the entire company. While analyzing behavioral signal data from our attribution pipeline, we noticed that certain patterns of user interaction, patterns too subtle and too distributed to be visible in any single session, reliably preceded purchase events by 48 to 96 hours. Not sometimes. Not in certain verticals. Across every dataset we tested, with an accuracy that made us check our validation methodology three times before we believed the results. That observation became the foundation of what we now call the neural targeting model: a system that predicts purchase intent up to 72 hours before the user themselves may be consciously aware of their intent to buy.
The signal taxonomy
Purchase intent prediction begins with signals, and the quality of the prediction is fundamentally determined by the quality and granularity of the signal taxonomy. Most intent prediction systems in the market rely on what we call "explicit intent signals": search queries containing commercial keywords, visits to pricing pages, engagement with comparison content, demo request form submissions. These signals are valuable, but they are lagging indicators. By the time a user searches for "best compliance software pricing," they are already deep in a purchase process. The targeting opportunity window has narrowed to hours, and every competitor with a basic retargeting pixel is competing for the same impression. Our model operates on a different class of signals entirely. We call them "pre-intent signals," behavioral patterns that emerge in the 48 to 96 hours before explicit intent manifests. These signals are invisible to traditional analytics because they do not look like purchase behavior. They look like ordinary browsing. The difference is in the micro-patterns.
The strongest predictor in our model, and the one that surprised us most, is what we call micro-hesitation patterns. A micro-hesitation is a pause of 150 to 900 milliseconds during a scrolling or navigation action, where the user's attention briefly catches on a piece of content before continuing. Traditional analytics does not capture these events because they fall below the threshold of what most tracking systems consider meaningful engagement. The user did not click. They did not hover for a measurable duration. They did not scroll back. They simply paused, imperceptibly, and moved on. But in aggregate, across multiple sessions over a 72-hour window, the frequency and distribution of micro-hesitations against specific content categories form a pattern that is powerfully predictive of purchase intent. In our validation studies, micro-hesitation patterns alone achieve 61.3% accuracy in predicting purchase events within 72 hours. Combined with the full signal taxonomy, accuracy reaches 84.7%.
Beyond micro-hesitations, the model ingests signals across five additional categories. Session architecture signals capture the structural patterns of how a user navigates: the ratio of deep-page visits to homepage visits, the branching factor of their navigation path (linear versus exploratory), and the temporal spacing between sessions (accelerating visit frequency is a strong pre-intent signal). Temporal engagement signals capture when a user engages: shifts in time-of-day patterns, weekend-to-weekday engagement ratio changes, and the emergence of "research sessions" that are longer and more focused than the user's baseline browsing pattern. Cross-domain signals, collected through our partner network of 14,000 publisher sites, capture the user's broader browsing context: are they visiting category-adjacent sites more frequently? Are they engaging with educational content in the product category? Are they reading reviews of competing products? Kinematic signals capture changes in how the user physically interacts with their device: scroll speed distributions shift when a user moves from casual browsing to purposeful research, and pointer movement becomes more deliberate and target-oriented. Finally, environmental signals capture contextual factors: device type shifts (users who begin researching on mobile but switch to desktop are 3.4x more likely to purchase), network environment changes, and time-zone-adjusted engagement patterns.
The model architecture
The neural targeting model is built on a modified transformer architecture with 3.2 trillion parameters, trained on 18 months of behavioral signal data from across our customer base. The core architecture consists of three components: a signal encoder, a temporal attention mechanism, and an intent prediction head. The signal encoder is a multi-modal embedding network that converts raw signals from all six categories into a unified 1,024-dimensional representation. Each signal category has its own specialized encoder, a convolutional network for kinematic signals, a recurrent network for temporal sequences, a graph attention network for cross-domain navigation patterns, and these category-specific representations are fused through a cross-attention layer that learns the correlations between signal types. This fusion layer is critical because the predictive power of individual signals is modest, but the interactions between signals are highly informative. A micro-hesitation pattern combined with a session architecture shift combined with a temporal engagement change produces a prediction that is far more accurate than any of those signals alone.
The temporal attention mechanism is where the architecture diverges most significantly from standard transformer designs. Purchase intent is not a static property, it evolves over time. A user who will purchase in 72 hours behaves differently from one who will purchase in 24 hours, and both behave differently from one who will not purchase at all. Our temporal attention layers operate over a sliding window of the user's last 14 days of behavioral data, with attention weights that are dynamically modulated by the predicted time-to-purchase. The model does not just predict whether a user will purchase, it predicts when, and it uses that temporal prediction to weight the importance of recent versus historical signals. Early-stage pre-intent (72-48 hours out) is characterized by subtle signal shifts that require the model to attend broadly across the full historical window. Late-stage pre-intent (24-0 hours out) produces stronger, more concentrated signals that allow the model to focus attention on recent sessions. This temporal attention mechanism improved our 72-hour prediction accuracy by 12.4 percentage points compared to a standard transformer baseline with flat attention.
Validation methodology
Extraordinary claims require rigorous validation, and a claim that we can predict purchase intent 72 hours in advance is extraordinary enough to warrant detailed methodological transparency. Our primary validation approach is temporal holdout testing. We train the model on behavioral data up to a cutoff date, then evaluate predictions against actual purchase events that occurred after the cutoff. This is not the same as random train-test splitting, which can leak temporal information and produce inflated accuracy metrics. Our temporal holdout approach ensures that the model is evaluated exclusively on future events that it could not possibly have seen during training. We run this validation continuously: every week, the model's predictions from 72 hours ago are compared against actual outcomes, and the results are published to an internal accuracy dashboard that our customers can access.
The headline accuracy number, 84.7% at the 72-hour horizon, deserves unpacking. This is measured as the area under the precision-recall curve (AUPRC), not simple accuracy, because the class distribution is heavily imbalanced (the vast majority of users do not purchase in any given 72-hour window). At our default operating threshold, the model achieves 78.2% precision and 71.4% recall, meaning that 78.2% of users flagged as high-intent actually purchase within the window, and 71.4% of actual purchasers are correctly identified in advance. For comparison, the best competing system we benchmarked against (6sense's intent prediction) achieved 52.1% AUPRC on the same dataset. The accuracy varies by vertical and purchase type. E-commerce transactions, which tend to have shorter consideration cycles and more pronounced behavioral signals, see accuracy as high as 91.3%. Enterprise B2B purchases, which involve multiple stakeholders and longer decision cycles, see accuracy around 73.8%. The model performs best when it has at least three sessions of behavioral history for a given user, the cold-start problem remains our biggest accuracy limitation for new visitors.
How autonomous agents use the predictions
The neural targeting model does not operate in isolation. Its predictions feed directly into the autonomous agent fleet, where they inform real-time targeting decisions. When the model identifies a user as high-intent at the 72-hour horizon, the agent assigned to that user's segment receives a structured intent signal that includes the predicted time-to-purchase, the confidence score, the primary signal drivers (which categories of behavior are most contributing to the prediction), and a recommended engagement intensity score. The agent uses this information to make autonomous decisions about bid levels, creative selection, channel allocation, and engagement frequency. A user predicted to purchase in 72 hours might receive a soft-touch awareness impression. A user predicted to purchase in 12 hours might receive a high-urgency conversion-oriented message with an incentive. The agent calibrates its approach to the predicted stage of the user's decision process, something that would be impossible with manual campaign management at any meaningful scale.
The results speak to the model's impact. Across our customer base, campaigns informed by neural targeting model predictions achieve a 4.1x higher conversion rate than campaigns using standard demographic and behavioral targeting. Cost-per-acquisition is 62% lower. And perhaps most importantly, the timing of engagement, hitting users at the right moment in their decision process rather than bombarding them throughout, results in a 34% reduction in ad fatigue metrics (measured by frequency-to-negative-engagement ratios). Quilmark reported that after deploying the neural targeting model across their North American campaigns, 68% of their conversions came from users who were identified as high-intent before they ever visited Quilmark's website. These were users who had not yet entered the traditional funnel, but whose behavioral signals indicated they were about to. The model allowed Quilmark's agents to reach them first, before competitors, before the user even knew they were in-market.
We are continuing to invest heavily in the neural targeting model. The next major milestone is extending the prediction horizon from 72 hours to 7 days, which would allow agent fleets to begin nurture sequences significantly earlier in the purchase process. Early experiments suggest this is feasible for verticals with longer consideration cycles, though accuracy at the 7-day horizon currently sits at 67.2%, below our production threshold. We are also exploring multi-stakeholder intent prediction for enterprise B2B, where purchase decisions involve buying committees rather than individual users. The goal is to predict not just individual intent, but collective organizational intent, by identifying correlated behavioral signals across multiple users within the same company. This is the frontier, and it is where we believe the most transformative targeting capabilities will emerge over the next 12 months.
