Machine Learning

Machine learning (ML) is the branch of artificial intelligence in which computer systems learn patterns from examples and use them to make predictions or decisions, instead of following rules a programmer wrote by hand. A spam filter that learns from millions of labeled emails, a ranking system that learns which results satisfy searchers, and a large language model that learns language by predicting the next word are all machine learning.

The field is usually divided by how the system learns:

• Supervised learning trains on labeled examples (this email is spam, this one is not) and learns to predict the label for new inputs
• Unsupervised learning finds structure in unlabeled data, such as clusters of similar customers or topics
• Reinforcement learning learns by trial and reward, and its human-feedback variant, RLHF, helped turn raw language models into usable assistants
• Self-supervised learning creates its own training signal from raw data, for example by hiding words and predicting them, and is how LLMs are pre-trained

Deep learning is the part of ML that uses large neural networks with many layers. Nearly all recent progress in language, images and speech comes from deep learning, and the transformer architecture published in 2017 is the design behind current LLMs.

Search engines have used ML for years. Google's documentation of its ranking systems names several ML systems: RankBrain, which helps relate words to concepts so pages can match queries without exact wording; neural matching, which compares concept representations in queries and pages; and BERT, which helps understand how combinations of words change meaning. Its MUM model is used for specific applications rather than general ranking. On top of these ranking systems, Google's AI Overviews and AI Mode use Gemini models to retrieve, read and summarize pages, and other AI search products such as ChatGPT, Copilot and Perplexity combine retrieval systems with LLMs in a similar way. Retrieval itself increasingly relies on learned embeddings, numeric representations that place similar meanings close together, which is what makes semantic search possible.

Two consequences matter for anyone trying to be found. First, learned systems reward patterns rather than individual tricks. A ranking or retrieval model trained on enormous amounts of data generalizes from signals of relevance, quality and user satisfaction, so tactics that exploit one narrow signal tend to stop working when the model is retrained. Second, learned systems are hard to reverse-engineer. Nobody outside Google or OpenAI can list the exact weights, and Google's own guidance warns that no third-party tool has access to its internal ranking or AI systems. The useful approach is empirical: change something meaningful, then measure rankings, citations and AI mentions over time.

ML also explains why AI answers vary. Generative models sample from probability distributions, so the same prompt can produce different wording, different recommended brands and different sources from one run to the next. Treating AI visibility as a statistical measurement, averaged over many prompts and repeated runs, is a direct response to how these systems work.

For most marketing and content teams, deep ML expertise is not required. What helps is knowing which inputs these systems learn from (the text, links, structure and engagement of the public web) and which outputs to monitor.