# Website Visitor Intent Scoring: A Practical Model for Founder-Led Teams > Score visitor intent as weighted, decayed probability. Keep contact fit and live intent separate, then route score bands to the right action. - **Author**: George Borelli - **Published**: 2026-08-13 - **Category**: Sales - **URL**: https://heyzinc.com/blog/website-visitor-intent-scoring --- ```tldr Score visitor intent as accumulated, weighted, decayed probability -- not a single signal and not a black box. Keep contact fit (who they are) separate from live intent (what they are doing now), weight signals by reliability, subtract for negative signals, and route score bands to observe, message, alert, or human takeover. Then validate the model against what actually happens and tune it. ``` Most founder-led teams I know swing between two mistakes with website traffic. The first is treating every pageview like a lead. A visitor hits pricing, a notification fires, someone drops what they are doing to start a chat. Three times out of four the visitor was skimming, or it was a competitor, or the tab was idle in the background. The team gets alert fatigue, the visitor gets a creepy message, and within a week the alerts get muted. The second mistake is the opposite: ignoring behavior entirely until a form comes in. By then the visitor has often left, the question that brought them to pricing is unanswered, and the conversation that could have happened this afternoon becomes a follow-up email next week -- if it happens at all. Both failures come from the same place. They treat intent as a switch that is either on or off, instead of a probability that accumulates and fades. A single pageview is a hint. A repeat visit to pricing from a tracked link, with dwell and a return session, is a decision-grade signal. The difference is the accumulated weight of several signals, adjusted for recency. That is what a score is for. Not a magic number, and not a black box that tells you who will buy -- a small, transparent model that turns behavior into something you can tune and act on with proportionate confidence. ## What an intent score actually is Visitor intent scoring is a methodology for ranking visitors against a scale that represents how likely they are to be evaluating or ready to act. The shape of it is borrowed from lead scoring, which has long separated two kinds of data: explicit data provided by or about the prospect (company size, industry, job title, geography), and implicit data derived from monitoring behavior (site visits, content downloads, email opens). The [established lead-scoring tradition](https://en.wikipedia.org/wiki/Lead_scoring) calls these "explicit" and "implicit," and the rule-based version assigns point values to each attribute and sets thresholds at which a lead is considered worth engaging. The difference for a founder-led website team is that you are usually scoring visitors who have not filled a form and may be completely anonymous. You have almost no explicit data. What you have is implicit, behavioral, and live. So the score has to be built mostly from behavior, and it has to be honest about what behavior can and cannot tell you. A score is not a prediction that someone will buy. It is a structured way to raise the probability that a visit is worth a specific response, and to lower it when the evidence is weak or fake. That distinction matters because the cost of acting is real -- your attention, and the visitor's tolerance for being interrupted. ## Contact fit versus live intent Before assigning weights, separate two inputs that most teams blur together. **Contact fit** is who the visitor is and whether they match your ideal customer. Firmographic signals: company, industry, size, role, region. This is the explicit layer. For anonymous website traffic it is often thin or absent, but where you have it -- a tracked link that carried a company, a form filled on a previous visit, a known account -- it tells you whether this is the kind of buyer you can actually serve. **Live intent** is what the visitor is doing right now. Behavioral signals: which page, how deep, how many times, in what sequence, from what source. This is the implicit layer, and for live website visitors it is usually your richest input. These combine, but not symmetrically. A great-fit visitor reading your homepage for the first time is a nurture or list-building opportunity, not an interruption candidate. A stranger -- possibly the wrong company entirely -- who is on pricing for the third time this week and scrolling carefully is a conversation worth starting now, because the live intent is high even if the fit is unknown. If you score them the same, you get both false positives and missed conversations. A third input sits beside these: attribution context, or where the visit came from. The web's default `Referrer-Policy` is `strict-origin-when-cross-origin`, which means cross-origin referrers are trimmed to the origin and sometimes suppressed entirely, so raw referrer data is often origin-level or missing. A [tracked link](https://heyzinc.com/blog/tracked-links-versus-utm-parameters) recovers that context where the visit originated from one -- a campaign, a DM, a community post -- which is why tracked-link context is one of the most reliable inputs you can put into a score. ## Weighting the signals A useful score weights signals by how reliably each one predicts evaluation, not by how easy it is to collect. In rough tiers -- tune these to your own audience: - **Low:** a single pageview, especially homepage or blog. It tells you someone arrived. It tells you almost nothing about why. - **Low to medium:** time-on-page, scroll depth, repeated interaction. Stronger than a bare pageview, but inflatable by a background tab or an idle session, so treat it as an approximation. - **Medium:** a page sequence. Pricing to product and back to pricing means the visitor is comparing and evaluating, not browsing. The sequence does more work than any single page. - **Medium to high:** repeat visits. A visitor who comes back is signaling interest a single bounce cannot. The catch is recognition -- you need a stable identifier to know they returned. - **Highest:** tracked-link context plus a repeat visit. The source conversation is known, the visitor returned, and you have attribution context to ground the opening message. This is the strongest combination most small teams will see. Sequences and combinations beat any single signal. A pricing pageview is a hint. A pricing pageview followed by a return visit from a tracked link, with dwell on the same section, is a decision. That is what [present-tense visibility that makes a score actionable](https://heyzinc.com/blog/real-time-visitor-tracking) is really for -- not watching a feed, but feeding a model that decides whether this minute is the right one to act. ## Recency and decay A score that never forgets is useless. A pricing visit two weeks ago should not still fire an alert today, because intent is not a lifetime total -- it is a present-tense state. Decay makes the score reflect what is happening now rather than what happened once. Concretely: each signal's contribution fades over time. You can think of it as a half-life. A pricing pageview might be worth its full weight for a few hours, half that by tomorrow, and near zero within a week. A repeat visit resets and adds to the weight. The exact half-life is a design choice that depends on your sales cycle -- a 30-day SaaS trial and a six-month enterprise evaluation do not decay the same way. One honest limit on "now": even mainstream real-time analytics is not instantaneous. Google Analytics documents its [Realtime report](https://support.google.com/analytics/answer/9271392) as a "best effort" service with no formal service-level objective, notes that app data is batched on the order of minutes, and warns that attribution processing is limited in Realtime. So when I say "live intent," I mean present enough to change what happens next -- not perfect ground truth about what a specific person is doing this second. That is still the only window in which acting on intent is possible. ## Negative signals A scoring model that only adds points will overreact. You also need to subtract and suppress. Common negative signals: - **Idle or background tabs.** Time-on-page looks high, but the visitor left the tab open and walked away. Engagement signals (scroll, repeated interaction, returning to a section) are more reliable than raw time alone. - **Bot and scraper traffic.** Crawlers can hit pricing and product pages in volume and inflate everything. Filtering them out of the scoring input is not optional. HeyZinc, for example, filters bot and scraper traffic out of visitor notifications by default -- a capability, not a marketing claim, and one worth checking in any tool you use. - **Careers-page bounces.** A visitor who lands on jobs and leaves is usually a job seeker, not a buyer. - **Single-bounce exits.** One page, no scroll, immediate exit. The lowest possible intent. - **Recognized non-buyer patterns.** Competitor research, repeated visits that never progress, traffic from regions or segments you cannot serve. Negative signals do two things. They keep the score honest by removing fake or irrelevant weight, and they protect the team from alert fatigue. A model that fires on every pricing pageview gets muted within a week. A model that fires only when positive signals accumulate and negative signals are absent stays useful for months. ## Action tiers: observe, message, alert, hand off A score is only useful if each band maps to a specific, proportionate action. Otherwise it is just a number in a dashboard. - **Observe (low).** Record the visit, let it inform the score, do not interrupt. Most visits live here, and that is correct. - **Message automatically (medium).** Start a contextual conversation through chat. The message should acknowledge what the visitor is doing without claiming to know who they are. This is where [proactive outreach for website visitors](https://heyzinc.com/blog/proactive-outreach) earns its place -- a soft, relevant nudge while they are still present. - **Alert the team (high).** Notify a founder or teammate, ideally through a companion or mobile notification, once there is a reply worth responding to. The alert should carry the score, the signals behind it, and the source context. - **Human takeover (highest).** A teammate joins the conversation by text or a live website call. Reserved for the band where the accumulated evidence and the response so far both justify a person's time. The thresholds between bands are a judgment call, not a universal rule. They depend on your team's size, your tolerance for interruption, and your audience. Set them conservatively at first -- it is cheaper to miss a borderline visit than to burn out on false positives -- and move them as you learn what each band actually produces. Do not let anyone tell you there is a proven threshold that maximizes conversion. There is not, for your specific situation. ## Validation: the model is a hypothesis A scoring model is a hypothesis about which behaviors predict which outcomes. Like any hypothesis, it has to be checked against reality. After the model runs for a while, review scored visits against what actually happened. Did the high-band visits produce real conversations? Did they qualify? Did any close? Did the medium-band automated messages get replies, or get ignored? Did the observe band contain visits you wish you had acted on? That review is where the model gets better. If repeat visits from a tracked link almost always produce a conversation but pricing-page dwell alone almost never does, you raise the first weight and lower the second. If careers-page bounces are leaking into the alert band, you strengthen the negative signal. The model is not set once; it is tuned against outcomes over time. Keep the honesty here. Scoring creates the precondition for timely, proportionate engagement. It does not by itself win customers. If your real bottleneck is message match, trust, or offer clarity, a better score will not fix it -- [response time is one of several conversion bottlenecks](https://heyzinc.com/blog/why-your-website-traffic-is-not-turning-into-customers), and it may not even be the one you are facing. The score tells you when to act. Whether acting pays off still depends on your audience, your offer, and your execution. ## Turning score bands into HeyZinc actions This is where the model connects to something you can actually run. HeyZinc lets you define behavior-based intent rules that watch multiple signals together rather than firing on a single pageview. A rule can combine a pricing sequence, a repeat visit, dwell, and a tracked-link source into one threshold, so the score reflects accumulated evidence instead of a trigger. When that threshold is crossed, the score band routes to an action: observe silently, start a contextual conversation automatically, alert the team after a reply, or hand off to a human who can continue by text or a live website call. Tracked links preserve the source conversation context, so the opening message can acknowledge where the visit came from without pretending to know more than the captured context proves. A few honest qualifications. Alert delivery, including through the companion or mobile app, depends on your workspace's notification configuration and the device, so treat mobile alerts as a configured workflow rather than a universal guarantee. Intent detection and the higher-tier analytics surfaces are tied to plan and setup; not every workspace has every surface on day one. And the recognition that makes repeat-visit scoring work is probabilistic -- private browsing modes and storage eviction weaken stable identifiers, as [WebKit's tracking-prevention documentation](https://webkit.org/blog/14445/) describes -- so a repeat visit is a strong signal, not a certainty. If you want to see how [real-time lead capture for small teams](https://heyzinc.com/for-smb) maps onto a scoring model like this, or you want to talk through what your thresholds should be for your site, [talk to our team](https://heyzinc.com/contact). The point is not to watch a bigger dashboard. It is to have a score you trust enough to act on, and a clear action for each band. ## Start narrow You do not need a predictive model or a data science team to start. You need three or four signals you actually believe in, one decay rate that fits your cycle, one negative rule that cuts the obvious noise, and one threshold above which a person gets involved. Pick the signals that map to your real sales motion -- usually a pricing or product sequence, a repeat visit, and a tracked-link source. Set the threshold conservatively. Run it for a couple of weeks. Look at what each band produced. Move the weights. The goal is not a perfect score. It is a score honest enough that you trust it to decide when to stay quiet and when to start a conversation while the visitor is still there. --- - [More Sales articles](https://heyzinc.com/blog/category/sales) - [All articles](https://heyzinc.com/blog)