What Keywords Were Designed For
Keyword targeting was one of the most important innovations in digital advertising. Before it, online ads were placed based on demographic assumptions, site category, or geographic targeting. Keywords made it possible to match an ad to the specific thing someone was looking for at the exact moment they searched for it. The precision was genuinely new.
The underlying model is a sparse representation: a query is a short list of words, and those words are matched against a list of keywords an advertiser has bid on. The match is mostly exact or fuzzy-string based, with some semantic expansion. Google and others built sophisticated systems on top of this model, but the core idea remains: if the query contains the right words, show the ad.
That model was designed for search boxes. A search box takes a short, compressed expression of what someone wants and converts it into tokens. The keyword is a natural unit for that input type. AI chat does not work that way. Users type complete questions, follow-up requests, clarifications, and references to earlier parts of the conversation. The input is richer, longer, and structurally more like a document than a query. Applying keyword matching to a conversational turn is like using a ruler to measure temperature: the tool was not built for the thing you are measuring.
What Conversation Context Actually Carries
Consider a user on a financial planning assistant who, over five conversation turns, asks how compound interest works, requests a comparison of index funds and active managed funds, asks what expense ratio is considered high, and finally asks "which brokerage would you recommend for a first-time investor with around $5,000 to start with?"
A keyword match on "brokerage" would fire for any advertiser targeting that term. A contextual model reading the full turn sequence sees something more specific: a user who is in the active comparison stage of a first-time investment decision, who has been researching passive investing strategies, who is cost-sensitive (they asked about expense ratios), and who is at a decision point with a specific dollar amount. That is a different targeting signal than a keyword match on "brokerage."
This is the core argument for context over keywords in AI advertising. The conversation carries intent information that keywords cannot capture, because keywords strip away the surrounding meaning and the temporal structure. Context preserves both.
The Temporal Dimension: How Intent Evolves Across Turns
One of the structural advantages of conversation-level targeting is that intent can be observed as it develops, not just captured at a single point. In search, you see one query. In a conversation, you see a sequence, and that sequence has structure. A user who starts with awareness questions and moves toward comparison and decision questions is following a recognizable trajectory. That trajectory predicts what kind of ad will land well better than any single query could.
Velocity's contextual engine reads turn sequences, not just individual turns. The relevance scoring for any given turn takes into account not just what was just said, but the conversational momentum. The scoring pipeline itself, including latency targets and the embedding model, is described in the technical overview: what topic thread has been running through the session, how specific the questions have become, and whether the user's language has shifted from exploratory to evaluative. This session-level context is something keyword targeting cannot replicate by design.
We want to be direct about the limits here too. Reading intent from a sequence is better than reading it from a single query, but it is still a statistical inference. Conversations do not follow clean funnel-stage progressions. People jump between topics, change direction, and ask abstract questions in the middle of what looks like a purchase-stage session. The model does not get this right every time, and honest reporting on our early data shows a meaningful variance in intent-score accuracy across different topic categories. Categories where users follow predictable dialogue patterns (home buying, software evaluation, travel planning) score more reliably than categories where conversations are less structured.
Why Keyword Bidding Cannot Simply Be Ported to Conversations
Some advertisers and ad platforms have tried to apply keyword targeting to conversational advertising by extracting keywords from the conversation turn and matching them against keyword lists. This approach is technically easy to implement and requires minimal change to existing advertiser workflows. We tried it ourselves in our early prototyping, and the results were instructive.
The problem is ambiguity at the turn level. A conversation about retirement planning might mention keywords like "stocks," "bonds," "risk," and "portfolio" in a single turn. Each of those terms, in a search context, would fire specific advertiser bids. In a conversation context, they are all part of one thread. Running a keyword auction against all of them simultaneously would produce overlapping bids from incompatible advertiser categories, making the match selection arbitrary.
More fundamentally, conversation turns use natural language, which means the same intent can be expressed in many ways that do not share keywords. "What should I do with my savings" and "how should I invest $10,000" have almost no keyword overlap, but they carry similar intent for financial services advertisers. A keyword model misses the connection. A contextual model that understands meaning rather than matching strings catches it.
The Practical Implications for Advertisers Building Conversational Campaigns
If you are building an advertising strategy that includes AI chat as a channel, the most important shift is from keyword-centric to topic-and-intent-centric targeting. Instead of bidding on specific terms, you configure category and intent parameters that describe the kind of conversational moment you want to reach. A project management software advertiser does not need to bid on "task tracking tool" and "team collaboration software" and twenty other variants. They configure a target category (business productivity software) and an intent threshold, and the contextual engine finds the conversations that match.
This is a simpler operational model for advertisers in some ways, and a more complex one in others. Simpler because you do not need to maintain and optimize a large keyword list. More complex because evaluating campaign performance requires understanding intent-score distributions and conversion path analysis rather than just CPC and CTR by keyword. That analysis requires different tools and different intuitions than search campaign optimization.
We are building the reporting and analytics infrastructure to support this, but we want to be honest with advertisers that conversational campaign analytics is a developing discipline. The frameworks that work well for search and social campaigns are partly applicable and partly not. Advertisers who treat conversational advertising as a learning channel and invest in understanding the signal data it generates will be better positioned as the channel matures than those who apply existing frameworks wholesale and wonder why the results do not behave like search.
Where This Leaves Keyword Targeting
Keyword targeting is not going away. For search, it remains the right tool for the input format. Even within conversational advertising, there are use cases where a keyword or phrase match layer adds value on top of topic-level targeting, particularly for brand protection (ensuring competitor keywords do not trigger your ads) or for very specific product categories where the category taxonomy is not granular enough.
The argument is not that keywords are obsolete. The argument is that applying keyword logic as the primary targeting mechanism to conversational turns wastes the most valuable thing about the surface: the rich contextual signal that comes from reading what people actually say to each other, rather than the compressed keywords they type into a search box.
