Beyond binary logic: why “Yes/No” no longer works

For years, finding similar audiences (Lookalike) was built on rigid binary logic. The advertiser set a condition: “My target audience is people who spent 30 seconds on the site, scrolled 80% of the page, and clicked button N.” Everyone who met the condition got a “1” (target); everyone else got a “0” (non-target).

The problem with this approach is that it ignores how “warm” the audience is. A user who spent 28 seconds on the site and studied a product closely but didn’t click the button got the same “0” as a random visitor who closed the tab after a second. To capture that hidden value, modern ML architectures are moving from classification (1/0) to intent scoring (Intent Scoring), scoring every visitor on a scale from 0 to 100.

Scoring architecture: a 3D user model

For the algorithm to predict conversion probability with high accuracy, the user profile inside the system is not a flat list of actions but a three-dimensional matrix (a feature vector space):

• 1. Context vector (Where and How?)
  This includes aggregated technical and geo data: time zone, device type, browser category, and referrer (organic, direct, a specific ad campaign). It forms the baseline portrait of the visit.
• 2. Historical vector and decay function (Time Decay)
  A person’s interest is not constant. If a user searched for a fridge a month ago, current interest is zero. If it was this morning — it is at its peak. The algorithm computes cumulative interest in site categories using a mathematical time-decay function: older actions gradually lose weight, making room for fresh intent.
• 3. Current-behavior vector (micro-conversions)
  This is a dynamics analysis: scroll speed (is the person reading carefully or skimming?), interaction patterns with UI elements in the first seconds of the visit.

Adaptivity instead of rigid frames: smooth anomaly cut-off

If we score a user from 0 to 100, an “overheated” audience problem appears. A content manager who uploads products for two hours a day will hit every target action at “100 points” in raw stats — but is useless for ads.

Historically, systems solved this with a hard filter: “if on the site more than an hour — it’s a bot or an admin.” But what if you published a long, engaging article and average time on site grew organically? A hard filter will cut off your most loyal audience.

Smoothness and adaptivity: how ClikBy cuts off anomalies

Hard thresholds like “if a user stays longer than 2 hours — it’s a bot” are hopelessly outdated. If you add an engaging long read, average visit time will rise, and a hard filter will cut off your most loyal audience. In ClikBy we solved this by introducing an adaptive, smooth anomaly coefficient (A_i).

Step 1. Drop absolute numbers in favor of percentiles

The ClikBy system no longer uses hardcoded limits (for example, “maximum 50 sessions”). Instead, the algorithm continuously analyzes the data distribution over the last 3–7 days and automatically determines:

• The median (the norm) — the current standard behavior of the audience on that specific site.
• 95th percentile (P₉₅) — the boundary above which only the 5% most active (“overheated”) users sit.
• 99th percentile (P₉₉) — the zone of extreme outliers, home to content managers, parsers, and tabs left open overnight.

If user behavior changes, the P₉₅ и P₉₉ metrics shift automatically. The platform adapts to the site’s “new normal” without manual intervention.

Step 2. A smooth penalty function (decay sigmoid)

Instead of a sharp cut-off “step”, we use an inverted logistic curve. While the user’s behavior stays within P₉₅, their multiplier A_i = 1 (no penalty). As soon as they cross the anomaly boundary, the coefficient starts плавно approaching zero according to the formula:

How it works in practice (final scoring)

Suppose an online store’s rules assigned the user 80 raw points of interest (RawScore). Let’s see how ClikBy’s 3D paradigm plays out in different situations:

Champion-Challenger: algorithms competing for ROI

In industrial marketing, relying on a single model — even a smart one — often leads to stagnation. To keep improving the quality of bought traffic, teams use a model A/B testing (Champion-Challenger).

How does this work in practice? Система параллельно рассчитывает скоринг двумя разными алгоритмическими пайплайнами. Алгоритм «А» может делать ставку на глубокую историю визитов, а экспериментальный алгоритм «Б» — придавать больший вес микро-поведению в текущей сессии.

Users on whom the models agree go into the base campaign. Segments where the models disagree are split (strictly by hashed identifiers) and sent to different ad campaigns to compete. By comparing final lead cost (CPA) and the average interest score of traffic that lands on the site, the marketer can mathematically prove which algorithm works better for that business.

“None of this deep predictive math would be possible without a strict Privacy by Design concept. There are no names, IP addresses, or raw cookies in our system. We work only with irreversible hashes and behavioral vectors. The system knows that "Vector X will complete a purchase with 85% probability", but it does not know who is behind that vector. That lets us deliver unprecedented Lookalike audience accuracy while staying inside the absolute, strict bounds of GDPR and CCPA.”

Request an audit of your current Lookalike audiences