An ordered list of numbers that represents a point or direction in a multi-dimensional space, and the fundamental data structure through which machine learning models represent every type of input, including text, images, user behavior, and campaign performance, in a form that mathematical operations can be applied to. Every embedding, every model prediction, and every gradient in deep learning is a vector, making the vector the atomic unit of computation in all of modern AI.
Also known as feature vector, data vector, numeric array
A vector is an ordered sequence of numbers, such as [0.3, 0.8, 0.1, 0.6], where the position of each number in the sequence is meaningful. In machine learning, vectors represent examples in feature space: each position in the vector corresponds to one feature, and the value at that position is the feature’s value for this example. A customer record represented as a feature vector might have its first element as age, second as monthly spend, third as days since last purchase, and so on for however many features are included. This vector representation is what enables mathematical operations such as distance computation, linear transformation, and dot product to be applied to any example in the dataset.
Learned embeddings, which represent words, users, products, images, and other entities in a continuous semantic space, are vectors of typically 64 to 2048 dimensions. These embedding vectors are not human-designed feature vectors; their components do not correspond to named features. Instead, they are learned from data such that the geometric relationships between vectors in the embedding space encode semantic relationships: similar entities have vectors that are close together, and certain semantic relationships manifest as consistent directional patterns in the embedding space. The dot product between two embedding vectors quantifies their similarity, enabling fast retrieval of the most similar entities to a query by identifying the vectors with the highest dot product.
Vector operations form the backbone of neural network computation. Matrix-vector multiplication applies a learned linear transformation to the input vector, projecting it into a new space. Dot products measure similarity and produce the attention weights in transformer self-attention. Gradient vectors quantify the direction of steepest increase of the loss function with respect to model parameters, pointing the optimizer toward the direction of greatest improvement. The entire forward pass of any neural network, and the entire backward pass of backpropagation, is a sequence of vector and matrix operations, making linear algebra the language in which neural networks are defined and executed.
A working ad agency using embedding-based semantic search, nearest-neighbor audience matching, or any AI system that represents content, users, or products as learnable representations is working with vectors constantly. The ability to reason about what a vector represents, what operations on vectors mean, and why cosine similarity measures the angle between vectors rather than their length provides the conceptual foundation for using these tools with understanding rather than as black boxes. The intuition that similar entities have similar vectors, and that vector arithmetic can capture semantic relationships, explains capabilities such as why semantic search finds conceptually related content even when no keyword matches exist.
Vector similarity search enables semantic product discovery that keyword matching cannot provide. A product search system that represents each product and each search query as embedding vectors, and retrieves the products whose vectors are most similar to the query vector, will surface products that are semantically related to the query even when they share no exact keywords. A query for “sustainable office accessories” will retrieve products tagged with “eco-friendly desk items” or “recycled workspace supplies” because these phrases have similar positions in the embedding space, reflecting their shared semantic territory. Keyword search would miss these matches entirely. The vector representation is what enables this semantic connectivity.
Audience vectors constructed from behavioral signals enable similarity-based look-alike targeting without predefined segment rules. Representing each user as a vector of behavioral signals (engagement rates by content category, purchase frequency by product tier, recency-weighted interaction counts), and identifying the users whose vectors are most similar to the vectors of high-value customers, produces a look-alike audience based on multi-dimensional behavioral similarity rather than discrete segment membership. The k-nearest-neighbor retrieval that identifies the most similar users operates entirely in vector space: it finds the users closest to the seed audience centroid vector using cosine or Euclidean distance. This vector-based similarity search is faster and more nuanced than rule-based segmentation for large-scale audience matching tasks.
Campaign performance vectors that represent each campaign as a numeric profile enable systematic creative pattern analysis across portfolio history. Representing each past campaign as a vector of performance metrics and creative attributes, clustering these vectors, and analyzing which attribute clusters correspond to high-performance versus low-performance outcomes translates qualitative creative intuition into quantitative pattern analysis. The geometric structure of the campaign vector space reveals which combinations of creative attributes co-occur in high-performing campaigns and which predict poor performance, providing a data-driven foundation for creative strategy recommendations that complements but exceeds what retrospective human analysis of individual campaigns can reliably surface.
An agency builds a semantic content matching system for a publishing client that needs to automatically match brand-safe inventory to advertiser campaign briefs. The client operates 14 content verticals with 80,000 to 120,000 active article URLs at any time. Advertisers submit campaign briefs describing their target audience context and content affinity requirements. The prior matching system used keyword-based contextual targeting that matched advertiser keyword lists to article metadata, producing a 22% match rate (fraction of available inventory receiving placement eligibility for any active campaign) and requiring advertisers to maintain keyword lists that often missed semantically relevant content. The agency implements vector-based semantic matching. Each article is represented as a 384-dimensional sentence embedding computed from the article title, first paragraph, and primary tags, using a pretrained sentence transformer. Each campaign brief is represented as an embedding of the advertiser-provided content context description. The system indexes all article embeddings in a vector database and retrieves the top-1% most similar articles for each campaign brief using approximate nearest-neighbor search over cosine similarity. Match rate increases from 22% to 71%: the semantic vector matching finds relevant content in adjacent topic areas that keyword lists do not explicitly cover. A human review sample of 200 matched article-brief pairs finds 91% are contextually relevant by editorial judgment, confirming that the semantic vector similarity correctly captures content affinity. The system also enables reverse search: given a high-performing article, retrieve all campaign briefs for which this article is a strong contextual match, surfacing placement opportunities that the keyword-centric system could not identify because the brief and article shared no common terms.
The generative AI foundations module covers vectors comprehensively including feature vectors, embedding representations, dot product similarity, nearest-neighbor retrieval, and the vector operations that underlie recommendation, semantic search, and audience targeting systems.