Why This Is Different from Search Intent
Search intent has been the gold standard for purchase signal since the mid-2000s. The reasoning is simple: someone typing "best project management software for remote teams" has stated their need explicitly. The query itself is the signal.
AI chat produces something structurally different. Instead of a single query, a conversation develops over multiple turns. The user refines their thinking, asks follow-up questions, and gradually expresses more specific needs. The intent gets clearer across the session, not all at once at the start. That temporal structure is something search cannot replicate, and it is one of the things that makes conversational advertising technically interesting.
When we built the signal layer inside Velocity, we had to rethink what "intent" means when a user is in the middle of a dialogue rather than composing a query. This post describes what we actually extract, how it gets translated into actionable advertiser signals, and where it still falls short compared to more mature channels.
The Four Signal Dimensions Velocity Surfaces
Topic category
The most basic signal is what the conversation is about. Velocity classifies each turn against a taxonomy of roughly 180 topic categories. These map to the IAB Content Taxonomy standard, which allows advertisers already running category-targeted campaigns on display or native to apply the same logic without rebuilding their targeting configurations. The mechanics of how those categories are scored per turn are covered in the technical overview of the relevance pipeline.
Category targeting is not revolutionary on its own. What is different is the resolution. A publisher's website might be generally about personal finance. A conversation on that publisher's assistant might be specifically about refinancing a home mortgage, which sits in a much narrower category with a distinct set of relevant advertisers. The conversation carries more targeting resolution than the page URL alone.
Purchase intent level
Not every on-topic turn carries the same purchase signal. A user asking "how does a 401k work" is in a different mental state than one asking "what is the contribution limit for a 401k in 2025 and where do I open one." Both involve retirement savings, but the second is much closer to a decision point.
Velocity scores each eligible turn on a purchase intent axis using a 0 to 1 scale. Advertisers can set a minimum intent threshold for their campaigns, which means ads only appear on turns where the purchase intent score meets the configured floor. A financial services advertiser might set a floor of 0.65 to filter out purely informational queries and reach only the users who are actively evaluating options.
Intent scoring is probabilistic. We are estimating from linguistic patterns, not reading user minds. Calibration matters, and we adjust the scoring model periodically as we accumulate more matched-conversion data from our early advertiser cohort.
Conversation stage
Velocity also tracks where the user appears to be in their decision process, across three coarse stages: awareness (learning what something is), consideration (comparing options), and decision (ready to act). This is distinct from the intent score. A user with high purchase intent who is still in the comparison stage may respond differently to a testimonial-forward ad than one who is already in the decision stage and needs a specific offer.
Advertisers can target by conversation stage independently of topic and intent score, or combine all three. In practice, most advertisers start with topic plus intent score and add stage targeting after they have accumulated enough campaign data to see which combination performs.
Negative signals
We also extract signals that indicate an ad should not appear. If a user has just expressed frustration, stated that they already purchased something, or is asking about a sensitive topic where advertising would be inappropriate, the relevance engine suppresses the ad slot for that turn regardless of topic match. We treat negative signals as equal-weight to positive ones. Getting the negative cases right matters as much as getting the matches right, both for advertiser brand safety and for publisher trust.
How This Compares to Social Behavioral Data
Social platforms built their targeting on behavioral history: what you liked, shared, searched, bought, and what your connections did. That data is rich and it has produced effective targeting for a long time. The tradeoffs are well known now: privacy pressure, consent fatigue, signal degradation as identifiers erode.
Velocity uses none of that. We work with context, not history. What is being discussed right now, in this session, at this moment. The upside is that it is genuinely privacy-safe. We do not need to know who the user is, we need to know what they are asking about. The downside is that we lose the longitudinal picture. Social targeting can model whether someone is likely to buy a new laptop based on three months of prior behavior. Conversational targeting can only see what they are asking about in the current session.
We are not claiming conversational context beats behavioral data in every scenario. For advertisers running broad awareness campaigns where historical affinity modeling works well, behavioral targeting on social channels is probably still the better option. Conversational targeting is strongest when a user is already in the middle of an active research or decision process, which is precisely when behavioral history from three months ago becomes less relevant anyway.
What Advertiser Reporting Looks Like
Advertisers using Velocity see campaign performance broken down by topic category, intent score range, and conversation stage at the turn level. This lets you see not just overall CTR but where your clicks are coming from. An ad that gets a 2.4 percent CTR on consideration-stage turns and 0.8 percent on awareness-stage turns is telling you something useful about where to concentrate budget.
We do not have post-click attribution built into the platform yet. That is on the roadmap. Currently, advertisers need to apply UTM parameters or their own attribution tooling to connect ad clicks from Velocity to downstream conversions. For advertisers who already run structured UTM frameworks across channels, this is a minor integration step. For those without attribution infrastructure, it is a gap that limits optimization depth.
Where This Goes
The honest answer is that conversational intent data is still developing as a discipline. What we have now is a strong foundation: contextual relevance, purchase intent scoring, and conversation stage classification. What we will have in 12 to 18 months includes better conversion attribution, category taxonomy refinement based on actual performance data, and more granular signal extraction from multi-turn dialogues.
If you are an advertiser running search and social campaigns, conversational advertising is not a replacement for those yet. It is an addition: a way to reach users at high-intent moments that search misses because the user formulated their need as a conversation rather than a query. For advertisers in high-consideration categories where the research process is long and dialogue-heavy, it is worth running experiments now rather than waiting for the channel to mature.
