An AI lexicon for modern SEOs

The arrival of AI Mode and the growth of generative search have transformed how people use search engines and how those engines process and present information.

The old approach to SEO, focused on keywords, rankings, and fixed search results, no longer meets the demands of this new environment.

Modern Search is driven by large language models, intelligent content retrieval, and responses crafted on the fly using signals from a constantly shifting web.

We must rethink how we talk about and approach SEO work.

AI system and process terms

As search engines move away from predictable, rule-based retrieval toward more fluid, AI-driven synthesis, it is crucial to understand how these systems handle user queries.

Query Fan-Out

Query fan-out is the process by which a single user query triggers the generation of multiple synthetic queries designed to capture a broader spectrum of potential user intent.

Rather than treating the original query as a standalone instruction, AI-powered systems like Gemini use it as a prompt to explore related questions, reformulations, and co-occurring intents.

Synthetic queries

Synthetic queries are AI-generated variations of a user’s original search query, created to better understand the full scope of user intent.

These are not queries typed by the user but inferred, reformulated, or expanded versions created by large language models (LLMs) as part of the search system’s reasoning process.

They aim to surface content that satisfies a range of possible meanings, needs, or preferences behind the initial query.

Intent signature

An intent signature is the AI system’s internal understanding of what the user hopes to accomplish when they submit a query.

It goes beyond the exact words used and forms a composite picture of intent, drawing from the query itself, the user’s context, such as device, location, and session history, and the interpretive capabilities of large language models.

This internal profile shapes every step in the search process, including how synthetic queries are created, what kinds of content are pulled in, and which formats are favored in the final results.

Custom corpus

A custom corpus is a temporary, query-tailored set of documents a search engine puts together in response to a user query and related synthetic queries.

Instead of combing through the entire web, AI-driven search systems build a targeted micro-index comprising content that closely aligns with the user's inferred intent signature.

This curated set becomes the foundation the system draws from to generate its final responses, whether AI-generated summaries, source citations, or answer boxes.

Latent Query Space (LQS)

Latent Query Space refers to the unseen web of related queries and user intents surrounding an original search.

These are alternative phrasings, potential follow-up questions, co-searched terms, past search patterns, and contextually inferred goals that the user might never type out. AI systems use this broader space to anticipate needs and shape results beyond the literal query.

LQS functions like the dark matter of search. It influences what content surfaces and how relevance is measured, even though it does not appear in keyword tools or search volume reports.

Chunking (content chunks)

Content chunking is the process of dividing content into smaller, standalone sections that have clear meaning.

These “chunks” make it easier for large language models to interpret, extract, and repurpose information.

Chunkability measures how effectively your content supports this structure. In AI-powered search, strong "chunkability" increases the chances your content will be selected for generative answers, cited in AI summaries, or surfaced in response to a specific sub-intent within a larger query.

Semantic triple

A semantic triple is a structured format representing a fact using three parts: subject, predicate, and object.

• Subject: the entity or concept
• Predicate: the type of relationship
• Object: the associated entity or attribute

This simple structure allows AI systems and search engines to understand and organize relationships between entities in a way that machines can interpret.

Semantic triples form the core of knowledge graphs, helping map out how concepts connect and enabling more accurate, context-aware search results.

Search intent "constellations"

A Search Intent Constellation is a group of tightly connected user intents that revolve around a central query.

Rather than treating a query as a standalone request, AI-powered search systems interpret it as part of a broader, dynamic cluster of possible goals, preferences, comparisons, and next steps that mirror real-world search behavior.

This constellation acts like a blueprint for what the user could aim for, whether it’s information, evaluation, or action. It guides how content is retrieved, answers are generated, and the formats presented in the results.

Tokens (tokenization)

Tokens are the basic building blocks that large language models like GPT-4, Gemini, and Claude use to read and generate language.

Depending on how the model breaks down text, a token can be a whole word, part of a word, a punctuation mark, or even a single character.

Turning natural language into these small units is called tokenization, which allows the AI to understand and work with text.

You can think of tokens as the currency that LLMs operate on. Every message you type in and every response the AI gives back is measured in tokens, not by words or characters.

SEO concepts

The fundamentals of SEO are undergoing a significant shift. Success is no longer tied strictly to keyword relevance but is increasingly shaped by how well content aligns with user context, semantic meaning, and interaction patterns..

Visibility matrix

The Visibility Matrix offers a fresh perspective on how content appears in AI-driven search environments.

Instead of relying on a single, linear ranking, like being in the top spot for a specific keyword, it considers various factors influencing retrieval and relevance.

These include user context, synthetic queries, intent categories, content formats, and prior user interactions.

In this model, visibility is not tied to one query alone. It depends on how well your content performs across a broad and shifting network of queries, intents, and format combinations, reflecting the more complex nature of modern search behavior.

Dense retrieval and sparse retrieval

Dense retrieval is a modern information retrieval technique used by AI search systems to match queries and documents based on their semantic meaning, rather than exact word matches.

It relies on embeddings, numerical representations of text generated by models like BERT, Gemini, or GPT, which allow systems to understand how closely related two pieces of content are in meaning, even if they use entirely different words.

Sparse retrieval refers to traditional keyword-based search techniques, where content is retrieved based on literal term overlap between a user’s query and a document.

This includes methods like TF-IDF or BM25, where content is ranked based on how often query terms appear in the text, weighted by their rarity or importance.

With dense retrieval, if a user searches for “how to relieve back pain after sitting all day,” dense retrieval might surface content titled “ergonomic chair posture tips” or “stretches to reduce desk-related pain,” because those documents are semantically aligned with the query even if the phrase “back pain” doesn’t appear verbatim in the query.

A sparse retrieval engine would prioritize pages that contain the exact phrase “affordable standing desk” over a page that says “cost-effective ergonomic workspace solution,” even if the latter is more relevant in context.

Attributed Influence Value (AIV)

Attributed Influence Value (AIV) is a suggested metric for tracking how frequently and significantly your content appears in AI-generated outputs like summaries, direct answers, or suggestions.

It captures not just presence but the depth of influence your content has within these responses.

In some ways, AIV serves as the generative SEO counterpart to share of voice in traditional advertising, reflecting how prominently your content features in the AI's interpretation and delivery of information.

Traditional metrics like click-through rate or rank position lose relevance when users receive answers directly from AI systems.

Persistent presence

Persistent presence means a brand or website shows up regularly across different types of content and user situations, not just in quick appearances on search results pages.

In AI-driven search, where results depend on context and personalization, being visible in many formats (videos, maps, reviews, tools, and articles) increases the chances of being included in AI-generated answers.

For example, a fitness brand with trusted blog posts, good reviews on Google Maps (Google Business Profile), helpful YouTube videos, and well-cited research builds a strong presence.

Even if none of its pages are ranked first, the brand becomes a reliable source that the AI is potentially more likely to use.

LLM Infrastructure & Workflow Terms

Today’s search engines depend on complex layers of language models that analyze, organize, and respond to queries in real time.

Foundation model

A foundation model is an extensive, general-purpose AI system trained on massive and varied datasets, such as books, websites, code, research papers, and real-world conversations.

These models are the backbone for many tasks, including answering questions, summarizing text, translating languages, generating code, and powering search tools.

In search, foundation models such as Google Gemini, OpenAI’s GPT-4, Anthropic’s Claude, and Meta’s LLaMA help make sense of user queries.

They go beyond matching keywords by understanding intent, generating related questions, and guiding what information should be retrieved or shown.

For example, if someone searches “how to reduce anxiety before a presentation,” the foundation model doesn’t just find pages with those words. It uses its training to:

• Figure out what the user wants, like quick tips or long-term strategies
• Recognize whether the focus is on anxiety in general or public speaking
• Consider related searches or past behavior to add context
• Detect tone and urgency, and suggest the best format, like a checklist or video

Downstream LLM(s)

Downstream LLMs are smaller, specialized language models that step in after a foundation model has interpreted a query and identified the user's intent.

While the foundation model handles the broader understanding by identifying goals, context, ambiguity, and creating related query variations, downstream models focus on specific tasks such as summarizing content, pulling structured data, comparing items, or offering recommendations.

Depending on the query's requirements, you can think of them as expert tools the system uses.

This multi-step approach helps AI search engines deliver clearer, more useful, and human-like responses.

For example, if someone searches “electric cars under £40,000,” the system might work like this:

• The foundation model understands the user is looking for an affordable car comparison, possibly as a first-time buyer, and expands the query with related variations
• A comparison model builds a table with specs like range, price, and charging speed
• A summarization model pulls key points from reviews and product pages
• A structured data model extracts clean data from sources like JSON, product listings, or schema markup

These outputs are combined into one clear and helpful answer for the user.

State data

State data is the real-time and ongoing context that AI systems use to understand better a user’s query within the bigger picture of their search behavior.

It includes details like the user’s device, location, session activity, search history, personalization preferences, and linked account data such as past visits in Google Maps, Gmail usage, or YouTube viewing habits.

This information helps the system deliver more relevant and tailored results based on the user’s characteristics and previous actions.

Classification cues

Classification cues are the signals AI systems rely on to determine what kind of query a user has entered and what response would best match it.

These cues help sort the query into intent types like informational, transactional, comparative, exploratory, navigational, or hedonic.

Based on that classification, the system decides which tools or models to use, whether to summarize, compare, extract data, or something else.

This step is a key turning point in AI search. Once a query is classified, it shapes everything that follows, from the models activated to the response format and the kind of content that can be included.