The use of machine learning models to estimate a user’s likelihood of taking a specific action, such as purchasing, converting, requesting information, or churning, based on behavioral signals, contextual cues, and historical patterns. Intent prediction is the foundational capability behind audience targeting, lead scoring, content recommendation, and automated bid optimization in advertising and CRM.
Also known as purchase intent modeling, intent scoring, propensity modeling
Intent prediction models learn the statistical relationship between observable signals and target behaviors by training on historical data where both the signals and the eventual outcomes are known. For purchase intent prediction, training data consists of behavioral sessions paired with whether a purchase occurred within a defined conversion window. The model learns which combinations of signals, including page visit patterns, search queries, product interaction sequences, device characteristics, and temporal features, are associated with conversion. Applied to new users in real time, the model produces an intent score that estimates how likely that user is to convert given their current behavioral profile.
The conversion window definition is a consequential modeling decision that shapes what the model learns to predict. A 7-day conversion window trains a model to predict near-term intent; a 90-day window trains a model to predict longer-consideration purchases. The optimal window depends on the typical purchase cycle for the product: a model predicting intent to purchase a fast-moving consumer good should use a short window; a model predicting intent to purchase a car should use a much longer one. Using the wrong window length produces a model that is technically predicting the right label but is predicting intent at the wrong point in the purchase cycle, leading to targeting decisions that are timed incorrectly relative to when the user is actually in-market.
Multi-intent models predict multiple intent types simultaneously from the same behavioral signals, which is useful for products with multiple conversion pathways or for clients whose CRM needs to distinguish between purchase intent, information-request intent, and support-seeking intent. Hierarchical intent models predict a coarse intent category, such as in-market versus browsing, before predicting a finer-grained intent type within the in-market category. Sequential intent models use the history of a user’s intent scores over time to predict how their intent is evolving, enabling more sophisticated triggers for remarketing and CRM outreach based on intent trajectory rather than a single snapshot score.
Almost every targeting decision an agency makes involves an implicit model of user intent: which users are most likely to respond to a message, when they are most likely to be receptive, and what will be most relevant to them at that moment. A working ad agency that makes these judgments explicit through intent prediction models, rather than relying on demographic proxies or manual intuition, reaches more of the right people at the right moments and wastes less spend on users who are not in the market for what the client is offering.
Intent-based targeting outperforms demographic and interest targeting on conversion metrics. Demographic and interest segments describe who a user is; intent signals describe what they are likely to do. A user who matches the demographic profile for a product but shows no in-market behavioral signals is a worse targeting bet than a user who does not match the demographic profile but is actively researching the product category. Agencies that have moved their targeting strategies from demographic-first to intent-first report consistent improvements in conversion rates and cost per acquisition because they are reaching users whose current behavior indicates purchase readiness.
CRM scoring programs benefit from custom intent models over platform defaults. Major ad platforms provide intent audience segments based on their own behavioral data, but these segments are built on signals from platform interactions and may miss the specific behavioral patterns that predict intent for a specific client’s product. A custom intent model trained on the client’s own conversion data, even if it uses platform behavioral features as inputs, will learn the specific signal combination that predicts conversion for that client better than a generic platform segment calibrated across many advertisers.
Intent score recency is a critical feature that many models ignore. A user who showed high purchase intent 30 days ago and has not returned to the site since may no longer be in-market. Using a static intent score without recency decay produces targeting lists that include users who have already converted elsewhere, exited the consideration set, or simply moved on. Intent scores should be time-decayed based on how quickly in-market status expires for the specific product category, with high-velocity categories using aggressive decay and long-consideration categories using slower decay.
An agency manages search and display campaigns for a home security company. The client has been using a demographic targeting strategy targeting homeowners aged 35-55 in certain income brackets. Conversion rate on this audience is 1.2% on search and 0.3% on display. The agency builds a custom intent model trained on 14 months of the client’s website visitor data paired with form submission conversions. The model uses 47 behavioral features including page view sequence, calculator tool engagement, pricing page visits, comparison content dwell time, return visit frequency, and device switching patterns. On a holdout validation set, the model achieves an AUC of 0.84 for predicting 30-day form submission. The top 20% of users by intent score convert at 6.8% on search and 2.1% on display, representing a 4.7x and 7x improvement over the demographic baseline respectively. The agency shifts 60% of display budget from the demographic targeting strategy to remarketing users in the top intent score quintile. Display conversion rate on the intent-targeted audience is 2.4% versus the 0.3% demographic baseline, and the campaign achieves target lead volume with a 38% reduction in display spend.
The automations and agents module covers how to build custom intent prediction models that improve targeting precision across paid media, CRM, and personalization, including the feature engineering and model validation practices that make intent scores reliable enough to act on.