This article explains why Article 22 GDPR does not typically apply to Voyado Engage’s scoring and recommendation features and provides descriptions of how the machine-learning algorithms work. You may use this documentation to support your GDPR compliance documentation and internal assessments, and to help you respond to data subject enquiries about how their data is used.

Article 22 GDPR

Article 22 of the GDPR gives individuals the right not to be subject to a decision based solely on automated processing, including profiling, where that decision produces legal effects concerning them or similarly significantly affects them. Article 22 applies only where all the following conditions are met:

1. A decision is made about the individual.
2. The decision is solely automated with no meaningful human involvement.
3. The decision produces legal effects or similarly significant effects on the individual.

Why Article 22 does not typically apply

For Voyado Engage features such as churn scores, propensity scores, and product recommendations, the third condition is generally not met.

A churn score, for example, is an internal metric used by your marketing team to decide whether to send a win-back campaign. The consumer is not denied a service, does not lose any rights, and faces no negative consequence as a result of the score itself. If anything, a high churn score typically leads to the consumer receiving a discount or re-engagement offer.

The same reasoning applies to propensity scores and product recommendations. They inform decisions about marketing relevance, but they do not produce outcomes of the kind Article 22 was designed to protect against. The EDPB’s guidelines on automated decision-making (WP251) specifically confirm that receiving targeted marketing or personalized content does not meet this threshold.

In most implementations, these scores feed into a human-led process (for example, a marketer reviews a segment and decides on a campaign), which further takes the processing outside the scope of Article 22 (the second condition).

Algorithm transparency

This section explains how Voyado Engage's machine-learning features work: what data they use, how scores and rankings are calculated, and what outputs they produce. For each feature, the model produces a score or ranked output that Authorized Users may use for segmentation and campaign purposes.

How churn scores are calculated

The churn score predicts how likely a Contact is to stop purchasing from you, based on their purchase history, recency, engagement, and reachability. It works in three main steps:

1. What data the model looks at

For each Contact, the model analyses four types of signals: purchase history (what they have bought, how much, how often, and across how many product categories), recency (how long since their last purchase), engagement (how often they click links in your emails), and reachability (whether they can be contacted via email, SMS, or both).

2. How the model learns what “churned” means

The model does not apply a one-size-fits-all rule; instead, it learns typical buying rhythms across your customer base. If someone normally makes a purchase every 30 days but has not purchased in 90, that is a significant deviation. Specifically, a Contact is considered churned when the gap since their last purchase is substantially longer than the typical purchase pattern. For Contacts with enough repeat purchases, this is based on their own personal buying cadence. For Contacts with limited history, the model uses the average pattern across all your customers. As a result, a weekly grocery buyer is evaluated differently from someone who makes seasonal purchases.

3. What you get as output

The model produces a churn probability between 0 and 1, where a higher number means a higher risk of losing the customer. From this probability, the system also calculates:

• Estimated churn date – when the customer is expected to become inactive
• Years left as an active customer – how long they are expected to keep buying
• Customer lifetime value (CLV) – based on their purchase history and your specific margins
• Segmentation into Active (≤ 0.5), Downward (0.5–0.75), or On the Leave (> 0.75)

All scores are recalculated regularly, so they reflect each Contact's most recent behavior.

How product affinity works

Product Affinity identifies which Contacts are most likely to be interested in a specific product or set of products, based on purchase behavior patterns across the customer base.

1. How the model understands products

The model builds a map of how products relate to each other based on real purchase behavior. When customers frequently buy certain products together, either in the same transaction or across consecutive purchases, those products are placed closer together on the map. Over time, this creates a rich picture of which products naturally cluster together and which are unrelated.

Each product receives a position on this map (called an embedding). Each Contact also receives a position, based on the average of all the products they have purchased, so Contacts with similar purchase patterns end up near each other.

2. How a query works

When you select one or more products (or filter by product attributes like brand or category), the system finds where those products sit on the map and calculates an average position, which becomes your search point. It then measures how close every Contact's position is to that search point. Contacts whose purchase history aligns closely with your selected products score highest.

3. What you get as output

Each Contact receives an affinity score reflecting how well their purchase history matches the queried products. From this score, Contacts are grouped into four tiers:

• Tier 1 (highest affinity) – purchase history strongly aligned with the queried products
• Tier 2 – moderate alignment
• Tier 3 – weak alignment
• Tier 4 (lowest affinity) – purchase history has little in common with the queried products

The underlying product map is rebuilt regularly, so it reflects the latest purchasing patterns across your customer base.

How product recommendations work

Product Recommendations suggest the products each Contact is most likely to buy next, based on what they and other customers have previously purchased.

1. How the model learns which products go together

The model looks at every receipt across your customer base and learns which products tend to appear together. If customers who buy product A also frequently buy product B, the model picks up on that pattern. It does this by giving every product a position in a kind of map: products that are commonly co-purchased end up close together, while unrelated products end up far apart.

Only products that are currently in your product feed are eligible to be recommended, so the model never suggests items that are out of stock or delisted.

2. How recommendations are generated for each Contact

For each Contact, the system looks at everything they have purchased and finds the products most similar to their purchase history. This produces a ranked list of candidate recommendations, which then goes through several refinement steps:

1. Already-purchased filtering – products the Contact has already bought are removed (unless you have chosen to allow re-recommendations).
2. Behavioral re-ranking – products the Contact has recently viewed, left in an abandoned cart, or purchased repeatedly get a boost in the ranking.
3. Item promotion – if you have assigned promotion weights to certain products in your feed, those products are boosted or dampened accordingly.
4. Category diversification – the system avoids recommending multiple products from the same category, so the final list covers a broader range.
5. Similarity de-duplication – if two recommended products are very similar to each other (based on their names and descriptions), the lower-ranked one is removed to avoid near-duplicate suggestions.

Contacts who do not have enough purchase history to generate a full set of recommendations receive fallback suggestions: the most popular products across your customer base, spread across different categories.

3. What you get as output

Each Contact receives a personalized list of recommended product SKUs, ranked by relevance. The number of recommendations is configurable (typically around 10), and each recommendation also carries a relevance score and an expiry date.

A global fallback list of top-selling products is also generated for use when a Contact has no purchase history at all.

The model is retrained regularly, so recommendations stay current with your latest product catalogue and purchasing trends.

Legal disclaimer

Please note that this article is provided for general guidance only and does not constitute legal advice. Voyado is not a legal professional adviser and does not accept responsibility for any interpretations and actions based on this information.

Article last reviewed 10 March 2026 13:27