The first generation of contextual targeting matched ads to keywords on pages. The second generation used topic classification to map content to audience interest categories. The third generation, which we have been developing and deploying for 18 months, does something categorically different: it identifies what an audience wants next, not just what they are reading now.
Keyword contextual targeting is well understood: match ads to pages containing specific terms. The problem is that language is ambiguous. A page about "investing" could be read by a 22-year-old interested in their first brokerage account, a professional fund manager, or a retiree managing a pension. The keyword "investing" tells you the topic, not the reader's need or their proximity to a purchase decision.
Topic classification improved on this by grouping pages into audience interest categories (personal finance, tech, sports, and so on), which allowed more nuanced audience matching. But topic classification still operates at the content level: it tells you what the page is about, which is one signal about the reader, not a direct signal about their intent or their next desired action.
Intent signals are not derived from what a page is about. They are derived from the contextual journey a reader is on: what brought them here, what they have been reading recently, and what the content structure suggests they want to do next.
Our v3 contextual system is built on three layers of language model analysis that operate simultaneously on every page we evaluate.
Standard topic classifiers ask "what is this page about?" Our semantic depth analysis asks "at what level of sophistication is this topic being treated, and who reads content at this depth?" A beginner-level personal finance article and an expert-level analysis of yield curve inversions are both classified as "finance" by standard systems. Our depth analysis distinguishes between them and targets each to the appropriate audience segment.
Depth analysis uses a fine-tuned large language model to score content on technical depth, assumed prior knowledge, and vocabulary complexity. These scores are mapped to audience segments built from observed correlation between content depth and audience characteristics (verified via data clean room partnerships with publishers who have registered audience data).
Intent signals come from the structure and language patterns within content, not just its topic. Content that uses comparative language ("X vs Y," "which is better," "how to choose") signals consideration-stage intent. Content with action-oriented language ("how to set up," "step by step," "getting started") signals lower-funnel intent. Content that uses evaluative language ("review," "our experience," "after 6 months") signals post-consideration intent from other readers and often high-intent new visitors.
We extract these patterns at the sentence and paragraph level, producing intent stage scores for each page: awareness (highest of funnel), consideration, decision, and post-purchase. Advertisers can target by intent stage, which is something no keyword or topic system can provide.
The most novel component is journey modelling: using the content graph (how pages link to each other, how users navigate between them) to model the typical reader journey that leads to a specific page. A user who arrives at a product review after reading three comparison articles is in a different intent state than a user who arrives at the same review directly from search. Our journey model assigns intent weights based on navigation patterns, allowing advertisers to reach audiences at specific points in their research journey even without cookie-based user tracking.
We have been running controlled comparisons of v3 contextual targeting against keyword contextual, topic contextual, and (where available) cookie-based audience targeting across 8 advertiser verticals over 12 months. The results are consistent.
V3 contextual delivers 28 percent higher conversion rates than topic-based contextual targeting and 19 percent higher conversion rates than keyword targeting on comparable inventory. More significantly, v3 contextual closes 65 to 70 percent of the performance gap between cookie-based audience targeting and contextual-only targeting. For advertisers in cookie-deprecated environments, this recovery is material.
The performance gap between v3 contextual and cookie-based targeting is smallest in high-intent categories (financial services, automotive, B2B technology) and largest in awareness-focused categories (CPG, entertainment) where audience-based targeting has advantages that content-based signals cannot fully replicate.
Because v3 contextual operates entirely at the page level, it requires no user data, no cookies, no device fingerprinting, and no cross-site tracking. Every inference is made from the content, not the audience. This is not a compliance feature added on top: it is the fundamental architecture of the system.
For advertisers navigating cookie deprecation, this is significant. V3 contextual provides durable targeting that works on any browser, any device, in any geography, without any identity infrastructure requirements on the advertiser's side. It is not a substitute for first-party data programmes where those are available, but as a fallback for unaddressed inventory it is a meaningful improvement over what has existed before.
Monthly analysis on programmatic trends, publisher strategies, and advertiser playbooks. No fluff. Just data and insight from the team building at the frontier of adtech.
No spam. Unsubscribe anytime. Monthly cadence.