What Data Does Artificial Intelligence Use? Definitive Glossary

Introduction: Understanding AI data terminology

At Pickastor, our analysis shows that most e-commerce businesses working with AI tools encounter the same fundamental barrier: they struggle to act on AI-driven recommendations because the underlying concepts remain unclear. Understanding what data artificial intelligence uses, and why it matters, is the first step toward making smarter decisions with these systems.

Unstructured Data

Information without a predefined format or organization, including images, videos, audio files, free-form text, and social media posts. AI systems require specialized processing techniques to extract meaningful patterns from unstructured data.

This glossary exists to close that gap.

Why AI data terminology matters for e-commerce

Artificial intelligence is no longer a technology reserved for large enterprises with dedicated data science teams. SMB store owners, marketplace sellers, and agency consultants are all working with AI-powered tools on a daily basis, whether for product optimization, demand forecasting, personalization, or search ranking. Yet the vocabulary surrounding these systems, terms like training data, feature engineering, or inference, often goes unexplained.

That creates a real business problem. When you cannot interpret what an AI system is doing or why it is making a specific recommendation, you lose the ability to evaluate its output critically, configure it effectively, or explain its value to stakeholders. Fluency in AI data concepts is not a technical luxury. It is a practical business skill.

What this glossary covers

This reference covers four interconnected areas of AI data knowledge:

• Data types: The categories of information AI systems ingest, from structured numerical data to unstructured text and images
• Collection methods: How data is gathered, labeled, and prepared before it ever reaches a model
• Processing techniques: The steps AI systems use to clean, transform, and interpret raw data
• Core AI terminology: The foundational concepts that describe how models learn, generalize, and produce outputs

Each entry is written to stand alone. You do not need to read this glossary from start to finish. Use it as a reference you return to whenever a term requires clarification.

Who this glossary is for

This resource is designed for e-commerce owners, enterprise merchandising teams, digital agencies, and marketplace sellers who work alongside AI systems and want to engage with them more confidently. No prior technical background is required.

How to use this glossary

This glossary is organized alphabetically across several sections, each covering a distinct cluster of terms. You can move directly to the letter range most relevant to your immediate question without reading sequentially.

Test Data

A separate dataset used to evaluate how well an AI model performs on new, unseen information. Test data is kept distinct from training data to provide an unbiased assessment of the model's accuracy and generalization capabilities.

Alphabetical organization and navigation

Terms are grouped into lettered sections, clearly labeled in each heading. If you are looking for a specific word, scan the section heading that matches its first letter and locate the entry within it.

Definition format

Every entry follows the same structure: a single opening sentence that defines the term precisely, followed by a short expansion that adds practical context. This keeps each definition self-contained, so you can read one entry without needing to read others first.

Cross-references and related terms

Where a term connects closely to another concept in this glossary, you will find a See also: note at the end of the entry. These cross-references help you build a fuller picture of how AI data concepts relate to one another. For a broader foundation on the subject, Everything You Need to Know About Data for AI is a useful companion resource.

Quick reference and deeper reading

If you need a fast lookup, focus on the bold term and its opening sentence only. When a topic requires more careful study, such as understanding When AI Data Leaks Happen: A Critical Case Study, the full entry and its cross-references will give you the detail you need.

Core AI data concepts: A-D

This section covers the foundational vocabulary that underpins how AI systems acquire, interpret, and learn from information. Each term below is defined to stand independently, so you can read entries in any order depending on your immediate need.

Structured Data

Information organized in a predefined format with clear rows, columns, and relationships, such as spreadsheets, databases, and CSV files. Structured data is easily searchable and analyzable by AI systems because it follows consistent schemas and formats.

Annotated data

Annotated data is a dataset in which each record has been enriched with descriptive labels, tags, or metadata that tell an AI model what the data represents, making it possible for the model to learn meaningful patterns rather than raw signals.

Annotation is the bridge between raw information and machine understanding. A photograph of a running shoe, for example, carries no inherent meaning to an algorithm. Once a human or automated process attaches labels such as "footwear," "athletic," "size 10," and "white," the image becomes annotated data that a model can use to recognise similar products in future.

For e-commerce teams, annotation quality directly determines how well an AI can categorise inventory, generate product descriptions, or match search queries to listings. Common annotation types include:

• Image annotation: bounding boxes, segmentation masks, or classification tags applied to visual assets
• Text annotation: sentiment labels, named-entity tags, or intent markers added to written content
• Structured annotation: additional columns or fields appended to tabular records to clarify meaning

The cost and consistency of annotation are persistent challenges. Research suggests that annotation errors can propagate through an entire training pipeline, producing models that confidently make the wrong predictions. Quality control processes, including inter-annotator agreement checks, are therefore as important as the annotation work itself.

See also: Data labeling, Training data

Bias in training data

Bias in training data is a systematic skew present in the examples used to train an AI model, causing the model to produce outputs that favour certain groups, categories, or outcomes over others in ways that do not reflect reality.

Bias is one of the most consequential issues in applied AI, and it originates almost entirely in data rather than in algorithms. If the product images used to train a visual search model predominantly feature light-coloured backgrounds, the model may perform poorly on images with dark or complex backgrounds, not because the algorithm is flawed, but because it never saw enough counter-examples to learn otherwise.

For e-commerce operators, training data bias can surface in several practical ways:

• Representation bias: certain product categories, price points, or customer demographics are over- or under-represented in historical data
• Measurement bias: the way data was collected systematically distorts the signal (for example, click-through data that reflects past ranking positions rather than genuine customer preference)
• Label bias: human annotators apply labels inconsistently based on subjective judgement

Addressing bias requires deliberate data auditing before training begins. Diverse, balanced datasets and regular model evaluations against held-out test sets are the primary defences. The broader challenge of whether AI systems have enough varied data to train on is closely related to this problem, since scarcity of diverse examples is itself a driver of bias.

See also: Annotated data, Training data

Categorical data

Categorical data is information that represents discrete groups or classes rather than continuous numeric values, such as product type, colour, brand, or country of origin, and it requires specific encoding techniques before most AI models can process it.

Unlike numerical data, categorical data has no inherent mathematical order in most cases. "Red," "blue," and "green" are not quantities; they are labels. AI models that operate on numbers need these labels converted into a format they can compute with. The two most common encoding approaches are:

1. One-hot encoding: each category becomes its own binary column (1 if present, 0 if absent), which works well for a small number of distinct values
2. Ordinal encoding: categories are assigned integer values, appropriate only when a genuine order exists (for example, "small," "medium," "large")

For marketplace sellers managing large catalogues, categorical data governs how

Processing and transformation: E-L

products are filtered, ranked, and recommended across every major platform. Understanding how that data is processed leads directly to the next layer of AI vocabulary: the techniques that convert raw inputs into forms a model can actually learn from.

Training Data

The dataset used to teach an AI model to recognize patterns and make predictions. Training data is intentionally labeled or annotated to help the model learn the relationship between inputs and desired outputs. The quality and size of training data directly impacts model performance.

Embeddings

An embedding is a dense numerical representation of an object, such as a word, product, or user, that captures its meaning and relationships in a compact vector format. Rather than treating each item as an isolated value, embeddings position it within a multi-dimensional space where similar items cluster together.

Why embeddings matter for e-commerce:

• A product embedding might place "running shoes" close to "athletic socks" and "sports insoles" because they share purchasing context
• User embeddings encode browsing and purchase history into a single learnable representation
• Search engines use embeddings to match queries with relevant products even when the exact words do not overlap

Embeddings are the backbone of modern recommendation systems, semantic search, and personalisation engines. When an AI platform surfaces the right product for the right customer, embeddings are almost always doing the heavy lifting behind the scenes.

See also: Feature extraction, Vector representations

Vector representations

A vector representation is the specific numerical array that encodes an object's properties in a format that mathematical operations can process. Every embedding is a vector, but not every vector is an embedding. Simple vectors might encode just a handful of product attributes, while embedding vectors can contain hundreds or thousands of dimensions.

Key characteristics of vector representations:

• Dimensionality: the number of values in the array, which determines how much information the vector can capture
• Distance metrics: vectors that are numerically close (measured by cosine similarity or Euclidean distance) represent objects that are conceptually similar
• Scalability: vector databases allow AI systems to search millions of representations in milliseconds

For e-commerce teams, vector representations underpin every "customers also bought" module and every visual similarity search feature. The quality of those vectors depends directly on the quality of the underlying product data.

See also: Embeddings, Feature engineering

Feature engineering

Feature engineering is the process of selecting, transforming, and creating input variables that improve a model's ability to learn patterns from raw data. It sits between data collection and model training, and it is widely regarded as one of the most consequential steps in the entire machine learning pipeline.

Common feature engineering techniques:

1. Normalisation and scaling: adjusting numerical values so that large-magnitude features do not dominate smaller but equally important ones
2. Interaction features: combining two variables to capture relationships that neither expresses alone, such as multiplying price by discount rate to create a "perceived value" signal
3. Temporal features: extracting day of week, hour, or seasonality flags from timestamps
4. Aggregation: summarising historical behaviour, for example, a customer's average order value over the past 90 days

Poor feature engineering is one of the most common reasons a technically sound model underperforms in production. Investing time here consistently returns better results than simply adding more data or increasing model complexity.

See also: Feature extraction, Labeled data

Feature extraction

Feature extraction is the automated or semi-automated process of deriving meaningful signals from raw, high-dimensional data such as images, text, or audio. Unlike feature engineering, which relies on domain expertise to construct variables manually, feature extraction uses algorithms, often neural networks, to discover useful representations on their own.

Feature extraction in practice:

• A convolutional neural network extracts edges, textures, and shapes from product images without being told what to look for
• A language model extracts semantic signals from product descriptions, identifying tone, category relevance, and keyword intent
• Audio models extract pitch, rhythm, and timbre from voice search queries

For marketplace sellers, automated feature extraction means that a well-structured product listing, complete with high-resolution images and detailed descriptions, gives AI systems more raw material to work with, which directly influences ranking and visibility.

See also: Embeddings, Neural network training data

Labeled vs unlabeled data

Advanced data types: M-R

This section defines the core data types that power advanced AI systems, from the contextual layers wrapped around raw inputs to the continuous streams that enable real-time decision-making. Understanding these terms helps e-commerce teams make smarter decisions about how they structure, collect, and leverage their data assets.

Semi-Structured Data

Information that contains some organizational elements but lacks the rigid schema of fully structured data. Examples include JSON files, XML documents, and HTML pages. Semi-structured data bridges the gap between raw unstructured content and fully organized databases.

Metadata

Metadata is data that describes other data, providing contextual information such as file size, creation date, author, format, and source. AI systems use metadata to organize, filter, and prioritize inputs before processing begins.

For e-commerce applications, metadata is far more than administrative detail. When an AI system evaluates a product image, it does not just analyze the pixels. It also reads the file name, upload timestamp, alt text, and associated category tags. This surrounding context shapes how the model interprets and ranks the asset.

Common metadata types in e-commerce AI:

• Descriptive metadata: Product titles, category labels, brand names
• Structural metadata: File formats, image dimensions, character counts
• Administrative metadata: Upload dates, modification history, source platform
• Behavioral metadata: Click timestamps, session duration, scroll depth

See also: Feature extraction, Raw data

Model training data

Model training data is the dataset used to teach an AI system how to perform a specific task, exposing the model to input-output pairs so it can learn patterns, relationships, and decision rules.

The quality and diversity of training data directly determines what an AI model can and cannot do. A model trained exclusively on product data from one retail category will generalize poorly to others. For this reason, leading AI platforms invest heavily in curating broad, representative training datasets before deployment.

Key characteristics of effective training data:

• Volume: Sufficient examples to capture meaningful variation
• Diversity: Coverage of edge cases, not just common scenarios
• Accuracy: Correctly labeled or verified ground truth
• Recency: Updated to reflect current language, trends, and user behavior

See also: Labeled vs unlabeled data, Validation sets

Validation sets

A validation set is a portion of data held back from model training and used to evaluate model performance during development, helping teams detect overfitting and tune hyperparameters before final testing.

The distinction between training data and validation data is critical. If a model is evaluated on the same data it learned from, performance scores become misleading. Validation sets provide an independent benchmark, revealing whether the model has genuinely learned generalizable patterns or simply memorized its training examples.

Typical data splits in supervised learning follow an approximate 70/20/10 ratio: 70% for training, 20% for validation, and 10% for final testing. These proportions vary depending on dataset size and task complexity.

See also: Model training data, Optimization data

Multimodal data

Multimodal data refers to inputs drawn from multiple distinct data types simultaneously, such as combining text, images, audio, and structured numerical data within a single AI model or pipeline.

Modern e-commerce AI increasingly operates across modalities. A customer might submit a voice query, upload a reference image, and receive a recommendation that draws on both inputs alongside their purchase history. Systems capable of processing this combination deliver more accurate and contextually relevant outputs than those limited to a single data type.

Multimodal combinations common in e-commerce:

• Image plus text: Visual search paired with keyword refinement
• Audio plus text: Voice search transcription with intent classification
• Behavioral plus product: Purchase history combined with real-time browsing signals
• Video plus metadata: Product demonstration clips analyzed alongside structured attributes

For marketplace sellers, multimodal capability means that investing in multiple content formats, high-quality images, detailed descriptions, and video assets, creates compounding advantages as AI systems become better at synthesizing these inputs together.

*See also: Feature extraction,

Specialized terminology: S-Z

This section covers the final range of core AI data concepts, from structured databases to zero-shot learning. Understanding these terms helps e-commerce teams make informed decisions about data collection, model evaluation, and the ethical boundaries of AI development.

Raw Data

Information in its original, unprocessed form before any cleaning, transformation, or analysis. Raw data often contains errors, inconsistencies, and irrelevant information that must be addressed before it can be effectively used by AI systems.

Structured data

Structured data is information organized into a predefined format, typically rows and columns, that machines can read and query directly. It is the most straightforward input type for traditional machine learning models and remains foundational to AI systems across e-commerce.

Common examples in e-commerce contexts include:

• Product catalogs: SKUs, prices, dimensions, and stock levels stored in relational databases
• Transaction records: Order IDs, timestamps, payment amounts, and customer identifiers
• Inventory tables: Supplier codes, reorder thresholds, and warehouse locations

Structured data is highly compatible with SQL databases, spreadsheets, and enterprise resource planning systems. Its clarity and consistency make it easier to clean, label, and feed directly into predictive models. The tradeoff is that it captures only what has been explicitly defined. Nuance, context, and meaning that live outside the schema are lost.

See also: Unstructured data, Feature extraction

Synthetic data

Synthetic data is artificially generated information that mimics the statistical properties of real-world data without containing actual records from real individuals or events. It is produced using algorithms, generative models, or simulation environments.

Synthetic data has grown significantly in importance for several reasons:

1. Privacy compliance: Generating synthetic customer records avoids the legal risks of using personally identifiable information in model training.
2. Data scarcity: When real examples of a rare event, such as product returns from a specific demographic, are too few to train on, synthetic samples can fill the gap.
3. Bias correction: Researchers can deliberately generate balanced synthetic datasets to counteract underrepresentation in historical data.
4. Cost reduction: Labeling real data is expensive. Synthetic data can be generated at scale with labels already attached.

For e-commerce teams, synthetic data is particularly useful in testing recommendation engines, fraud detection models, and demand forecasting systems before deploying them against live customer data.

See also: Training data, Data augmentation

Test data

Test data is a reserved subset of a dataset used exclusively to evaluate a trained model's performance on unseen examples. It is never used during training or validation.

The purpose of test data is to simulate real-world conditions as closely as possible. A model that performs well on training data but poorly on test data is said to be overfitting: it has memorized patterns rather than learned generalizable rules.

Best practices for test data management include:

• Keeping test sets completely separate from training pipelines
• Ensuring the test set reflects the actual distribution of data the model will encounter in production
• Refreshing test sets periodically as real-world conditions change

For marketplace sellers running AI-powered pricing or inventory tools, understanding whether a vendor's model has been properly tested against held-out data is a meaningful quality signal.

See also: Validation data, Training data

Training data

Training data is the labeled or unlabeled dataset used to teach an AI model to recognize patterns, make predictions, or generate outputs. It is the primary input that shapes a model's behavior and capabilities.

Training data quality is arguably the single most important factor in AI performance. In our experience at Pickastor, the difference between a well-optimized product listing and a mediocre one often comes down to the quality of the training data used to build the underlying recommendation or scoring model. Noisy, incomplete, or biased training data produces unreliable outputs regardless of how sophisticated the model architecture is.

Key dimensions of training data quality:

• Accuracy: Labels and values must correctly reflect reality
• Completeness: Missing fields reduce the signal available to the model
• Diversity: A narrow dataset produces a model that generalizes poorly
• Recency: Outdated training data leads to models that miss current trends

Most commonly confused terms

Even experienced teams misuse AI data terminology, and the consequences for implementation can be significant. Confusing training data with test data, or raw data with processed data, leads to flawed model evaluation, poor system design, and wasted budget. The pairs below are the most frequently conflated in e-commerce AI projects.

Training data vs test data

Training data is the dataset the model learns from. Test data is a separate, held-out dataset used to evaluate how well the model performs on examples it has never seen. Using the same data for both tasks produces artificially inflated accuracy scores and models that fail in production. In e-commerce, this distinction matters when validating a product recommendation engine: the model must be tested against real purchase events it was never trained on.

Labeled data vs annotated data

These terms are often used interchangeably, but they carry a subtle difference. Labeled data refers to any dataset where inputs have been assigned a target output (for example, "this product image belongs to category: footwear"). Annotated data is broader and includes richer contextual markup, such as bounding boxes around objects in an image, sentiment tags on customer reviews, or entity tags in product descriptions. All labeled data is a form of annotation, but not all annotated data is simply labeled.

Structured vs unstructured data

Structured data lives in defined fields with consistent formats: price columns, SKU identifiers, inventory counts. Unstructured data has no predefined schema: customer reviews, product photography, support chat transcripts. Both types feed AI systems, but they require different processing pipelines. Most e-commerce businesses hold far more unstructured data than they realize, and extracting value from it requires additional preparation steps before any model can use it.

Synthetic data vs real data

Synthetic data is algorithmically generated to simulate real-world patterns. Real data is collected from actual events and users. Synthetic data is valuable when real data is scarce, sensitive, or imbalanced, but it can introduce distribution gaps that cause models to underperform when deployed against genuine customer behavior. The two are often combined rather than used exclusively.

Raw data vs processed data

Raw data is data in its original, unmodified state: server logs, form submissions, sensor readings. Processed data has been cleaned, normalized, and structured for model consumption. The transformation between these stages is where most data quality problems are introduced or resolved. Understanding how AI and data science interact in practice helps teams allocate the right resources to each stage.

Confusing any of these pairs during system design creates compounding errors that are difficult to diagnose once a model is in production.

Quick reference table: AI data types at a glance

With so many overlapping terms in the AI data landscape, a single reference point saves time when making decisions about data strategy. The table below consolidates the core concepts from this glossary into one scannable resource, organized by category for quick lookup.

Data types by category

How to use this table

Cross-reference this table when evaluating a new AI tool or auditing an existing data pipeline. For context on how location and behavioral data is sourced in practice, the breakdown in Where Does Placer AI Get Its Data? Complete Breakdown illustrates how these categories apply to a real-world platform. Teams building e-commerce AI systems will encounter most of these types simultaneously, making familiarity with all of them a practical necessity rather than an academic exercise.

Related resources: Dive deeper into AI data topics

The glossary terms above become more actionable when paired with practical guidance on implementation, data quality, and strategy. The resources below are organized by topic so you can move directly to the area most relevant to your current project or question.

AI and the evolving data science landscape

Understanding what data AI uses is inseparable from understanding who manages that data and how roles are shifting. Are Data Scientists Being Replaced by AI? What Experts Say examines how automation is changing data workflows, which directly affects how e-commerce teams should think about data ownership and pipeline management.

Data collection and quality for e-commerce

For SMB and enterprise e-commerce teams, raw data is only as useful as the processes behind it. Guides on data collection best practices cover topics such as structured versus unstructured input, labeling workflows, and validation techniques. These resources help teams avoid the common pitfall of training AI models on incomplete or biased datasets.

AI implementation and product optimization

E-commerce agencies and marketplace sellers working with AI optimization tools will find implementation guides particularly useful. Platforms like the Pickastor AI Optimization Platform and its AI Score feature rely on well-prepared product data to deliver accurate recommendations. Practical implementation resources bridge the gap between glossary-level understanding and day-to-day operational decisions.

Ethical AI and data privacy

Resources on ethical AI address consent frameworks, data minimization principles, and regulatory compliance. These are essential reading for any team handling behavioral, transactional, or personally identifiable data within an AI pipeline.

Recently added terms and updates

This glossary section highlights newly added terms and recent updates made to our comprehensive resource library. Last updated in June 2025, these entries reflect the latest industry developments, emerging concepts, and refined definitions to keep our content current and relevant for users seeking up-to-date information.

This glossary is reviewed and expanded on a regular basis to reflect the pace of change in AI data practices. The following terms and clarifications have been added or revised in the most recent update cycle.

New additions

• Synthetic data generation: Added to reflect growing enterprise adoption of AI-generated training datasets as an alternative to raw customer data collection.
• Data residency: Included following increased regulatory scrutiny across multiple jurisdictions.
• Federated learning: Expanded with practical e-commerce context, particularly relevant for marketplace sellers managing distributed product catalogs.

Emerging concepts to watch

Privacy-preserving AI techniques and consent-based data architectures are reshaping how businesses approach model training. For a closer look at how leading AI providers handle data ownership and privacy obligations, see our coverage of hidden facts about OpenAI and data privacy you should know.

Check back regularly as new terms are added each quarter.

Frequently asked questions

What types of data does artificial intelligence use?

Artificial intelligence systems use a broad range of data types, including structured data (such as spreadsheets and databases), unstructured data (such as images, video, and free text), and semi-structured data (such as JSON files and emails). Understanding what data does artificial intelligence use is foundational to building reliable, high-performing models for any business application.

How is data collected for AI training?

Data is collected through web scraping, user interactions, sensor feeds, third-party data providers, and internal business systems. For e-commerce applications, this often includes product catalogs, customer behavior logs, and transaction histories.

What is the difference between training data and test data?

Training data is used to teach a model to recognize patterns, while test data is held back to evaluate how well the model performs on unseen examples. Keeping these datasets separate prevents overfitting and gives a realistic picture of real-world accuracy.

Why is data quality important for AI systems?

Poor-quality data produces unreliable predictions, biased outputs, and costly errors in production. Clean, consistent, and representative data is the single most controllable factor in determining whether an AI system delivers business value.

Can AI systems work with unstructured data?

Yes. Modern AI models, particularly deep learning architectures, are specifically designed to process unstructured data such as product images, customer reviews, and video content. Preprocessing steps such as tokenization and normalization are typically applied first.

What is synthetic data and why is it used?

Synthetic data is artificially generated data that mimics the statistical properties of real datasets. It is used when real data is scarce, sensitive, or too costly to collect at scale.

How do AI systems handle real-time data?

Real-time data is ingested through streaming pipelines and processed continuously, allowing models to update recommendations, detect anomalies, or adjust pricing dynamically without waiting for batch updates.

What is data labeling and why is it necessary?

Data labeling is the process of tagging raw data with meaningful annotations so that supervised learning models can learn from it. Without accurate labels, a model cannot distinguish between correct and incorrect outputs.

What are the ethical considerations in AI data collection?

Key concerns include informed consent, data minimization, bias in training sets, and compliance with regulations such as GDPR. Responsible data practices protect both end users and the businesses deploying AI systems.

How much data does an AI system need to be effective?

There is no universal threshold. Simpler models can perform well with thousands of examples, while large language models require billions. Data quality and relevance consistently matter more than raw volume.

Based on our work at Pickastor, e-commerce teams that prioritize clean, well-labeled product data see the fastest gains when adopting AI tools. If you are ready to put these principles into practice, the Pickastor AI Optimization Platform provides a structured starting point for assessing and improving your catalog data quality.