Glossary of AI personalization

An A/B test is a controlled statistical experiment in which a single variable changes between two versions while every other factor stays identical. That isolation is what lets you claim a causal link between the change and the result, rather than a correlation.

An affinity profile is a numeric vector in which each dimension corresponds to a specific category, brand or attribute — interest in Nike, interest in the Sneakers category, the $70–140 price band. Values are time-decayed, so recent actions weigh more heavily than older ones.

An AI shopping assistant is a seller-side conversational assistant that advises the shopper across the retailer's catalogue and guides them to a purchase. Unlike a shopper-side AI agent, it does not act autonomously across different stores.

AOV = revenue / number of orders. Period and channel both matter: AOV varies sharply by traffic source (organic vs retargeting), device (desktop vs mobile) and audience segment (new vs returning shoppers).

Bayesian statistics in A/B testing expresses the result as the probability that a variation beats the control, which lets you decide as data accumulates instead of waiting for a fixed p-value threshold.

Behavioral segmentation groups users by their actions — purchase history, product views, clicks and visit frequency — in order to build personalized communications and recommendations.

Cart abandonment is an unfinished purchase: the shopper added products to the cart but never placed the order. The market average abandonment rate is around 70%.

The cold start problem is what happens to a recommender system when there is no historical data on a user or a product, which makes accurate personalization impossible through standard collaborative filtering. It is a data problem, not a model problem — a better algorithm does not solve it.

Collaborative filtering is a recommendation algorithm that predicts a user's preferences from the behaviour of similar users or from patterns of joint consumption between items. Unlike content-based filtering, it never looks at what a product is — only at how people interact with it.

Conversational commerce is selling through a conversational interface — chat, voice or messaging — rather than a classic storefront. It covers both scripted bots and autonomous AI agents.

CR is measured across the funnel: visitor to order (the macro rate), category to PDP, PDP to cart, cart to order. Growth in the overall rate usually comes from improving the bottleneck — the step with the heaviest drop-off.

CRO is the discipline of raising conversion systematically through a loop of data collection, hypothesis formulation, prioritization and validation by controlled experiment.

Cross-sell is the practice of offering complementary products from adjacent categories to a customer who has already chosen the main item, with the goal of raising average order value.

Unlike a CRM, which manages sales and communications, and a DMP, which handles anonymous audiences for advertising, a CDP works with identified customers and builds persistent profiles enriched in real time. Its core value is a single customer view that holds regardless of the channel an interaction happened in.

Customer lifetime value (LTV) is the projected total revenue from a customer across the entire relationship with the company, used to assess marketing efficiency and unit economics.

Dynamic content is content on site pages that renders differently for different users or segments in real time. The URL stays the same for everyone — only what is inside the page changes.

Exit intent is a trigger that fires when a visitor shows the intent to leave the site, and is used to show a personalized retention offer at that moment.

First-party data is data a company collects directly from its own users through channels it owns — website, app, CRM — which makes it the most accurate, the most current and the easiest to keep compliant.

FBT is a recommendation algorithm that suggests products based on an analysis of what other shoppers bought together inside a single order.

Hyper-personalization is personalization in which the decision is made at the level of an individual user and an individual request, drawing on a long-term profile, session context and real-time events rather than on segment membership.

Identity resolution is the process of merging the different identifiers and data sources belonging to one real person into a single unified customer profile inside a CDP.

An LLM (large language model) is a language model with billions of parameters, trained on large text corpora by autoregressive token prediction, capable of understanding language, following instructions and generating text.

Multi-armed bandit solves the exploration/exploitation trade-off: it balances trying out new variations against exploiting the best one found so far. The clearer the lead of one variation, the more traffic it receives automatically.

Personalization is the automatic adaptation of content, product selection and offers to an individual user, based on their behavioural profile, history and the context of the current session.

A personalization engine is a platform that collects behavioural data and decides in real time on the optimal content, recommendations or offers for each individual user.

Personalized recommendations are product selections generated individually for each user from their behavioural profile and interaction history. What separates them from popular or similar items is the input: the individual profile, not an aggregate of the store or the attributes of the current page.

PLP personalization is the algorithmic re-ordering of products on a category page against the preferences of an individual shopper, derived from their behavioural profile. It changes the sort order of an existing category, never its contents — the same products, ranked differently for each visitor.

Real-time personalization is the adaptation of recommendations and content to a user's actions in the current session, with a delay of under 100–200 ms.

A recommendation widget is the UI component on a site page that displays algorithmically selected products as a carousel, a grid or a list. It is the delivery surface, not the algorithm — the same widget can be driven by any recommendation strategy.

A recommender system is an algorithmic system that selects the products most relevant to a user from the catalogue, based on their behaviour, product attributes and session context. It is not a single model but a pipeline: candidate generation, ranking and merchandising rules are separate stages, each of which can be changed independently.

RAG (retrieval-augmented generation) is an AI architecture in which a language model is extended with a retrieval step: relevant documents are pulled from an external store and passed into the LLM context so the answer is built from real data.

RFM analysis is a method of segmenting a customer base by the recency, frequency and monetary value of purchases, used to personalize marketing and manage the customer lifecycle.

Sample size is the minimum number of observations per A/B test variation that makes it possible to detect the target effect at the chosen statistical power.

Semantic search is an information-retrieval method that encodes queries and documents into a shared vector space and returns results by similarity in meaning rather than by lexical match.

A single customer view (SCV) is one record per customer that brings together every available data source, so their behaviour, preferences and interaction history can be understood in full.

Social proof is a set of UI elements that display the activity and behaviour of other shoppers in order to reduce uncertainty and speed up the purchase decision.

Statistical significance is the degree of confidence that the difference between A/B test variations is real rather than the product of random fluctuation; the usual requirement is p < 0.05.

Unlike cross-sell, upsell offers a more expensive version of the same or a comparable product rather than an addition to it. For example: a shopper is looking at a $900 laptop, and the upsell offers a comparable model at $1,150 with a better screen.

Zero-party data is information a shopper voluntarily and knowingly gives a brand — through quizzes, surveys or preference settings — and it is used for immediate personalization when no behavioural history has accumulated yet.