The essential terms and concepts of modern SEO, artificial intelligence and search algorithms.
A software system designed to pursue goals, make independent decisions and carry out complex tasks (generating actions) rather than simply generating text. Google's MLE-STAR project is one example.
The systemic tendency of language models (LLMs) such as GPT-4 or LLaMA-3 to favour content with recent timestamps (real or artificial), which can shift Top-10 results towards more recent dates (1 to 5 years) at the expense of older but authoritative content.
A phenomenon observed in AI reranking where recent content systematically rises into the Top 40 while older content falls down the ranking (Bottom 60), creating a pivot point in the middle of the SERP.
A theoretical framework unifying online optimisation, associative memory and the architectural design of sequence models such as Titans.
A new Google algorithm (a neural system) that replaces keyword matching with an understanding of overall meaning and intent, through vectors close to the main keyword.
The long-term memory module in the Titans architecture, which stores salient facts based on their "surprise gradient".
A page's semantic quality score, computed not only from its position on one keyword but from its cumulative strength across a set of close queries (a cluster). The formula is often 1 / (60 + position).
A new model architecture able to handle enormous contexts (> 2 million tokens) and using a long-term memory based on surprise.
A concept defining the three critical ranking stages in the AI era: 1. Document Ranking (macro signals), 2. Passage Ranking (precision SEO), 3. Passage Generation (the AI answer).
A probability metric (downstream validation) applied after the AI has generated a draft. It checks whether the facts in the draft are corroborated by solid evidence (supportive search results).
A criterion assessing the intrinsic quality of content based on effort, originality, talent/skill and accuracy. It aims to reward high-value content and penalise low-effort content.
A cosine similarity score (upstream selection) between the vector of the user's query and the vector of a content segment. It determines whether a document is mathematically relevant enough to be used.
The mechanism models (such as Titans) use to decide whether to memorise a piece of information. If information is unexpected or breaks a pattern (high surprise), it is encoded; if it is predictable, it is ignored.
A Google evaluation criterion measuring the fit between the result offered and the user's intent (Fully Meets, Highly Meets, and so on).
A strategy to optimise content so it is shown directly as the single answer in conversational engines or voice assistants.
An approach that handles conversational queries and builds mega-FAQs to answer natural-language questions, in anticipation of voice assistants and agents.
The process of deconstructing a query to reveal its deep semantic and multimodal vectors, in order to reach a "Highly Meets" or "Fully Meets" score.
Breaking a topic or a complex task into precise sub-intents (often action- or decision-oriented) instead of targeting generic keywords.
The technique of aligning a specific passage of content with the numeric vectors (embeddings) close to the main keyword, to influence the AI's answer.
An SEO strategy focused on fast results: priority on decision-stage landing pages, manual/expert writing, and validating a specific angle before producing volume.
The smallest verifiable unit of information (schema: subject + relation + entity/value) an AI can isolate. Maximising the density of atomic proofs raises the confidence score.
A vector representation of different formats (text, image, video, audio) for the same search intent, essential for competitive queries.
This glossary evolves with new developments in SEO and AI.