The R4 Signal Model: How Cursive Turns Raw Intent Into High-Confidence Buying Signals

When a company visits a review site, reads a product comparison article, or searches for vendor alternatives, that behavioral signal has genuine value — it suggests the company is actively thinking about a purchase decision. The challenge is that the same category of signal (a page visit, a content consumption event) is also generated by competitors doing research, students writing papers, journalists covering the industry, and job seekers evaluating career options. Without processing, raw behavioral signals are too noisy to act on.

Most intent data platforms solve the noise problem incompletely. They aggregate signals at scale, run them through keyword-matching models, and deliver a list of accounts scored "high intent" — without addressing the fundamental question of whether those signals represent genuine buying activity. The result is a conversion problem that the industry has largely attributed to sales execution rather than signal quality.

The R4 Signal Model is Cursive's answer to the signal quality problem. It is a four-layer processing framework that every signal passes through before reaching a customer's sales team. Each layer addresses a specific failure mode in how raw behavioral data is turned into actionable buying signals.

R1 — Relevance: The Signal Feed

The quality of an intent data output is bounded by the quality of its input. If the signal feed is broad and indiscriminate, no amount of downstream processing can fully recover the accuracy lost at the source. This is the fundamental problem with co-op data networks.

Co-op networks work by recruiting thousands of publishers to share their audience behavioral data in exchange for access to the aggregate pool. The incentive for publishers to join is the size of the network; the incentive for the platform is coverage. Both incentives push toward volume rather than relevance. The result is a signal network where a significant proportion of observed behaviors are generated by visitors to sites that have nothing to do with B2B buying research.

Consider what happens when a general technology news publication is in the co-op. Their readers are primarily curious technologists, industry observers, and general readers — not buyers in active evaluation cycles. When one of those readers clicks on an article mentioning CRM software, that click becomes an "intent signal" that flows downstream to every CRM vendor subscribed to the network. The reader was not evaluating CRM; they were reading tech news. But they have now been scored as in-market.

Cursive's R1 Relevance layer is a curated signal feed built from 28,000 domains selected specifically for their contribution to genuine B2B buying research. The selection criteria are not reach or traffic volume — they are the behavioral characteristics of the audiences visiting those domains, the nature of the content produced, and the correlation between signals from those domains and confirmed purchase decisions.

The +28% relevance improvement over co-op data means that a larger proportion of signals that enter the R4 pipeline correspond to actual buying research activity — before any additional processing has been applied. This is the feed quality advantage that downstream layers then amplify.

R2 — Reasoning: Buyer-Stage Classification

Even with a high-quality curated feed, a behavioral signal is underspecified without context about what it means for buyer intent. Two visitors from the same company can generate signals from the same domain category — one is doing preliminary awareness research with no near-term purchase decision, the other is actively comparing vendors in the final stages of evaluation. Keyword-only models score both identically.

R2 Reasoning applies URL-level buyer-stage classification to every signal before scoring. Rather than flagging that a page about "marketing automation" was visited, the model examines the specific URL, its content structure, and the visiting company's profile to classify the buyer stage.

The buyer-stage taxonomy used by R2 spans five stages:

The "6x depth" reference describes the number of signal attributes the R2 model evaluates per URL compared to keyword-only systems. A keyword system evaluates one attribute: keyword presence. R2 evaluates six: topic classification, content type, intent modifier language, page structure signals, visiting company firmographic context, and historical correlation of similar URL patterns with confirmed purchase events. Each attribute is weighted and combined into a buyer-stage score that is substantially more predictive than keyword presence alone.

R3 — Recognition: Industry Cohort Models

A model trained on aggregate cross-industry data learns the average buyer research pattern across all verticals. This average is useful in the general case but systematically wrong for specific industries. B2B SaaS buyers evaluating a sales intelligence tool read different content, in a different sequence, on different platforms, than healthcare administrators evaluating a patient communication tool — even if both are technically "in-market" for software with similar feature categories.

The implication is that a single aggregate model will produce different error rates across different industry verticals. Verticals that closely match the aggregate pattern (typically the highest-volume industries that dominated the training data) will see better signal quality. Verticals that diverge from the aggregate pattern will see systematically worse signal quality — false positives and false negatives that reflect the aggregate model's blind spots.

R3 Recognition addresses this by training separate industry cohort models per vertical. Each cohort model learns the research patterns specific to buyers in that industry — the sites they visit, the content types they consume at each buying stage, the sequence of signals that historically predicts a purchase decision. The model is vertical-aware from day one, which means customers in specific verticals receive signals calibrated to their buyers rather than calibrated to the generic B2B buyer.

R4 — Restrictions: Noise Suppression

Even with a curated feed, URL-level classification, and vertical cohort models, a meaningful portion of behavioral signals in any intent dataset are generated by non-buyers. Competitors researching the market. Job seekers evaluating companies. Students and academics doing course research. Journalists and analysts covering the industry. Internal employees generating signals from their own company's IP ranges. Each of these visitor categories produces behavioral signals that look superficially similar to buyer research but represent no purchase intent.

R4 Restrictions is the suppression layer that identifies and removes these signal types before they reach scoring or delivery. The filter operates across several suppression categories:

• Competitor suppression: Companies whose firmographic profile places them in the same vendor category as the product being tracked are suppressed. A competitor researching your pricing page is doing competitive intelligence, not evaluating a purchase.
• Internal traffic suppression: Known internal IP ranges, employee email domains, and behavioral patterns consistent with internal traffic are removed to prevent your own organization's activity from inflating apparent intent.
• Job seeker suppression: Visitors whose behavioral pattern is consistent with career research (careers page visits, Glassdoor cross-referencing, company culture content) rather than product research are filtered.
• Academic and student suppression: Institutional email domains, university IP ranges, and behavioral patterns consistent with coursework or academic research are removed.
• Media and analyst suppression: Known media company domains, analyst firm IP ranges, and behavioral patterns consistent with editorial research are filtered.

The practical effect of R4 Restrictions is a substantial reduction in signal volume combined with a substantial increase in signal precision. What was a list of 10,000 "intent accounts" from a co-op feed with keyword matching becomes a list of 500 high-confidence accounts after all four R4 layers have been applied. The 500 accounts are in genuine active buying stages. The remaining 9,500 were noise.

How the Four Layers Compound

The R4 layers are not independent filters applied in sequence — they are compounding quality multipliers. A signal that passes R1 (came from a relevant domain) is then enriched by R2 (buyer-stage classified), refined by R3 (scored against vertical-specific patterns), and cleaned by R4 (non-buyers removed). Each layer adds information and removes uncertainty. The combined effect is substantially larger than any single layer.

The Super Pixel: On-Site Identity Layer

The R4 Signal Model processes off-site behavioral signals — activity on the 28,000-domain curated feed. Cursive's Super Pixel is the complementary on-site layer: a first-party visitor identification pixel that identifies the specific individuals visiting your website.

The two layers are designed to work together. Off-site R4 signals tell you which companies are actively researching your category across the web. The Super Pixel tells you which specific individuals at those companies are actively researching your product directly. A company that appears in your R4 signal set (off-site intent) and whose employee visits your pricing page or comparison content (on-site Super Pixel) represents a high-confidence convergent signal: a named individual at an account that is both category-researching and product-evaluating.

The Super Pixel achieves a 70% visitor identification rate — meaning 70% of anonymous visitors to your website are identified by name, email, job title, and company rather than remaining anonymous. For a website receiving 10,000 monthly visitors from ICP accounts, this represents approximately 7,000 named, contactable leads per month that would otherwise leave as anonymous sessions.

Signal Quality as a Compounding Advantage

The case for the R4 Signal Model is ultimately a case for signal quality as a compounding advantage. Teams that operate on high-noise intent signals spend sales cycles reaching accounts that are not in a buying stage, burning outreach capacity on false positives, and progressively losing confidence in the tooling. Teams that operate on high-quality signals spend the same outreach capacity on accounts that are genuinely in active evaluation, see higher response rates, build faster pipeline, and compound their advantage as the model learns from those outcomes.

The difference between a 10,000-account intent list at 12% buyer accuracy and a 500-account list at 94% buyer accuracy is not just a difference in list size. It is a difference in what happens to the sales team that has to work those lists. The first team works through 10,000 accounts, finds that the vast majority are not in a buying cycle, and eventually de-prioritizes the intent data workflow. The second team works through 500 accounts, finds that most are genuinely evaluating, builds confidence in the signal, and scales the workflow.

Signal quality is where the leverage is. The R4 model is Cursive's approach to delivering it consistently.

If you want to see what R4 signal quality looks like on your own traffic, run a free estimate of how many named leads your site is generating and losing. Or start with the Super Pixel and layer in the full R4 off-site signal model from there.

For context on the specific failure modes that R4 was designed to address, read the companion post on why intent data fails.