Search Intent Optimization: a plain-English glossary of SEO, AEO, GEO and AI terms

Search and AI work has acquired an impressive amount of terminology in a very short time. Some of it describes real technical systems. Some of it is useful shorthand. Some of it is a way of making familiar work sound as though it has just arrived from the future.

This is my working glossary for the terms I actually use with clients and in build work. It is deliberately practical. If a term has no settled definition, I say so. If a tactic is often oversold, I say that too.

A new label does not create a new discipline. The useful question is still: what does the system need to understand, retrieve, verify or help someone decide?

Start with the job, not the acronyms

I use Search Intent Optimization as the plain-English umbrella for the work of making a business easy for the right buyer to find, understand, trust and choose, wherever they begin their research.

It has four connected jobs:

1. Be found: make useful information technically accessible, clearly organised and relevant to a real search or research need. 2. Be understood: make the business, its relationships and its claims easy for people and systems to interpret correctly. 3. Be trusted: support important claims with direct experience, useful source material and independent corroboration where it matters. 4. Be chosen: give a serious buyer the proof, commercial clarity and next step they need to make a decision.

SEO, AEO and GEO still describe useful parts of the job. They are not competing channels, successive levels of a pyramid or substitutes for thinking about the buyer and the decision.

The specialist labels

Search Intent Optimization

Search Intent Optimization is not another platform-specific discipline. It is a way to keep the work connected to the buyer's job. Technical SEO supports being found. Entity work and structured information support being understood. Useful answers and source material support trust. Good user experience and commercial clarity support the final decision.

The name is only useful if it keeps those connections visible. It should not be used to claim that every search journey can be mapped neatly or that one metric can represent a complex buying decision.

SEO

Search engine optimisation is the work of making a site easier for search engines to discover, understand and present for relevant searches. It covers technical access, content quality, information architecture, search intent, entity clarity and measurement. Good SEO is not a trick for forcing a ranking.

AEO

Answer engine optimisation is a loose label for improving the chance that content can help answer a direct question in search, assistants or other answer surfaces. In practice it overlaps with good information design, clear answers, credible sources, entity clarity and conventional SEO. There is no single agreed AEO standard or universal score.

GEO

Generative engine optimisation is another loose label, usually for work intended to help a business be understood and appropriately represented in generative search and AI answers. I use it to mean entity reconciliation, technical entity foundations and authority architecture. It is the outward-facing entity and source-credibility part of the wider job, not a separate channel with its own magic switch.

LLMO

Large language model optimisation is marketing shorthand for similar work. It is not a formal technical discipline. Ask what the proposed work actually changes, how it will be observed and which outcomes it cannot promise.

AIO, DEO and SXO

AI optimization or AIO is too ambiguous to be useful without an explanation of the task. Decision engine optimization or DEO often renames ordinary decision support and then invents a KPI layer around it. Search experience optimization or SXO usually points to sensible work on usability, trust and conversion. Use the more specific task name where possible.

AI visibility

A useful umbrella phrase for whether a brand, product, person or source appears in selected AI answer surfaces. It is an observation, not a mature universal KPI. A credible report records the prompts, locations, date, platform, citations and limits of the sample.

AI Overview and AI Mode

Names used by Google for AI-powered search experiences. They are product features, not generic names for every AI answer. Their behaviour, availability and presentation can change, so a strategy should not depend on one frozen version of the interface.

Search foundations

Crawler and crawling

A crawler is software that automatically requests pages and follows links to discover web content. Crawling is that discovery process. Being crawled does not guarantee that a page will be indexed, ranked or shown to a user.

Crawlability

Whether a crawler can reach important pages and resources reliably. Broken links, accidental blocks, slow servers, bad internal architecture and unrendered JavaScript can all get in the way.

Rendering

The process of turning a page's HTML, CSS and JavaScript into the page a user or crawler can see and interpret. JavaScript-heavy sites need testing because the source response and the finished page are not always the same thing.

Indexing

The process by which a search engine processes content it has found and makes it eligible for retrieval in its index. Indexing is not a promise of traffic or rankings.

Ranking

The order or prominence a result receives for a particular query, audience, place and moment. There is no one permanent rank for a page.

SERP

Search engine results page. The results interface for a query, including ordinary web results and potentially maps, shopping, video, images, featured results or AI features.

Query and keyword

A query is what someone typed or asked. A keyword is a research label used to group queries with similar language or intent. Treating a keyword as a single fixed user need is a common planning error.

Search intent

The likely job behind a query. Someone may want a definition, comparison, supplier, proof point, instruction, location or solution to a problem. Intent is a hypothesis that should be checked against the actual results and the audience, not guessed from a phrase alone.

Search demand

The evidenced level and shape of interest in a topic or query set. Third-party volume estimates are useful inputs, not precise counts of buyers.

Search appearance

How a result is displayed, including its title, snippet, URL and eligible rich features. Better appearance can help a result communicate relevance. It is not a replacement for the underlying page being useful.

Snippet

The text preview shown with a search result. Search engines may generate it from the page rather than using the meta description exactly as written.

Meta title and meta description

HTML metadata that helps describe a page in search contexts. A title is an important signal and user-facing label. A meta description is a suggested summary, not a ranking switch and not a guarantee of the displayed snippet.

Canonical URL

The representative URL for materially duplicate versions of a page. A `rel=canonical` tag expresses a preference. Search engines can use different signals and may choose another canonical, so it needs checking rather than faith.

Redirect

A server instruction that sends a request from one URL to another. A permanent redirect is usually the appropriate tool when a page has moved for good. Chains, loops and irrelevant destination pages create waste and confusion.

robots.txt

A file at the site root that gives participating crawlers instructions about which URLs they should not crawl. It is not a reliable way to keep sensitive material private, and it does not automatically remove an already-known URL from search.

noindex

A directive asking compliant search engines not to include a page in their search index. It is different from blocking crawling. A page must usually be accessible to a crawler for its noindex instruction to be seen.

XML sitemap

A machine-readable list of important URLs that helps search engines discover them. It is a hint and an inventory, not an instruction to index every URL.

Internal linking

Links between pages on the same site. Good internal links make useful paths visible to people and machines, explain relationships and help important pages be discovered. A giant automated web of weak links is not information architecture.

Information architecture

The deliberate structure of a site's topics, pages, labels and paths. It answers what belongs together, what is distinct and where someone should go next. It is one of the places where SEO, UX and entity clarity meet.

Structured data

Standardised, machine-readable data embedded in a page to describe its visible content and relationships. It can help systems interpret what a page is about and can make certain search appearances eligible. It should accurately describe the page, not invent a better version of the business.

Schema markup and Schema.org

Schema markup is the common name for structured data using the Schema.org vocabulary, often expressed as JSON-LD. Schema.org offers a shared vocabulary. Each platform decides which types and properties it uses, so valid markup is not a promise of a rich result, citation or knowledge panel.

Entity

An identifiable real-world thing, such as a person, company, product, place, organisation, publication or concept. An entity can have names, aliases, relationships, dates and independent references. A keyword is not necessarily an entity.

Knowledge graph

A structured representation of entities and their relationships. Search engines, platforms and businesses can each maintain their own version. There is not one public master graph that every system reads in the same way.

Knowledge panel

A search feature that presents information about a recognised entity. It is a platform feature, not a general certificate that all entity work is complete.

E-E-A-T

Google's shorthand for experience, expertise, authoritativeness and trust. It is a framework for thinking about helpful, reliable content, not a visible score to chase. Trust is the most important part of the framework.

YMYL

Your Money or Your Life. A label for topics where inaccurate information can significantly affect health, safety, financial stability or civic life. These topics need particularly strong accuracy, sourcing and editorial care.

Core Web Vitals

A set of user-experience measurements relating to loading, interactivity and visual stability. They are useful diagnostic signals. They do not turn a weak page into a good answer or a good business into a trusted entity.

JavaScript SEO

The part of SEO concerned with sites whose meaningful content, links or metadata depend on JavaScript. The practical work is testing what is served, rendered, crawlable and indexable rather than assuming the framework got it right.

Log-file analysis

Analysis of server request logs to understand which URLs bots and users actually request. It is one of the better ways to replace a crawl-based guess with evidence about crawler behaviour on a site.

International SEO and hreflang

International SEO helps a site serve the right language or regional version to the right audience. hreflang is a technical annotation that signals language and regional alternatives. It requires accurate reciprocal setup and does not fix duplicate, weak or poorly localised content.

Migration

A material change to a site's platform, URLs, content structure, domain, language setup or tracking. A migration needs an inventory, redirect plan, technical checks, ownership, launch gates and post-launch monitoring. It is not just a design release.

Entity and AI-answer work

Entity reconciliation

The work of establishing that scattered references point to the same real-world entity, or documenting that they do not. It may involve aliases, old domains, former names, people, products, places, dates and conflicting public records.

Entity disambiguation

Making clear which of several similarly named entities is meant. A name alone is rarely enough. Useful disambiguators include place, sector, history, relationships, official identifiers and attributable source material.

Entity architecture

My term for the connected work that makes an entity easier to recognise and verify. It combines identity reconciliation, clear first-party representation, structured relationships and corroborating evidence. It is not a schema-only exercise.

Technical entity foundation

The accessible information architecture, accurate on-page facts, structured data and clear relationships that give machines a consistent first-party account of an entity. It cannot make an untrue claim credible.

Authority architecture

The evidence around an entity that gives a system reason to take it seriously: useful original material, attributable expertise, history, independent validation, case studies and relevant references. It is earned through the quality and consistency of the underlying work.

Corroboration

Independent sources that support a factual claim or relationship. Repeating a claim across pages controlled by the same business is not the same thing as independent corroboration.

First-party source

Material published or directly controlled by the entity being described. It is essential for current factual representation, but it has an obvious interest in its own claims. Important claims may need independent support too.

Independent validation

Relevant evidence from a source that is not controlled by the entity. The source's relevance, standards and proximity to the claim matter more than raw link count.

Citation

A visible reference from an answer surface to a source. It shows that the source appeared in that answer at that time. It does not prove a stable ranking, causal relationship, endorsement or commercial result.

Citation-worthy source

Not a formal platform category. A practical description of a source that is specific, traceable, current enough for the claim, useful to the question and credible in context. No one can guarantee a system will cite it.

Retrieval

The process of selecting information that may be useful for a query or task. Search engines retrieve web documents. Internal AI systems can retrieve approved business information. Retrieval is selection, not proof that the selected text is correct.

Grounding

Connecting an AI response to supplied or retrieved sources so that claims can be checked. It can improve traceability, but it does not eliminate bad sources, misinterpretation or fabricated conclusions.

llms.txt

A proposed convention for a plain-language file intended to help language models navigate a website. It is not a web standard and there is no broad public evidence that it reliably improves AI visibility. Treat it as a low-risk supplement only after the important pages, links and facts are already sound.

Zero-click search

A search where the user gets enough information from the results interface to avoid visiting a website. It can be useful to the user and still reduce visits. Measure it in context, not as an automatic failure or success.

Semantic SEO

An umbrella phrase for designing content around meaning, entities, relationships and intent rather than repeating exact words. The useful version is just clear, well-structured communication backed by genuine subject matter knowledge.

Topical authority

Another useful but non-standard phrase. It describes a site's demonstrated depth and reliability on a subject area. It is not created by publishing a fixed number of pages, clustering keywords or adding internal links on their own.

Measurement without pretend certainty

Baseline

The starting measurement against which a later change is compared. It needs a metric, value, reporting period, source and clear scope. "Visibility was low" is not a baseline.

Prompt set

A documented group of questions used to observe AI-answer behaviour. A useful prompt set has a stated audience, decision stage, location, language and reason for each question. Changing the questions changes the measurement.

Citation rate

The proportion of observed answers that cite a specified source, domain or brand. It can be a useful observation for a defined prompt set. It is not a universal measure of authority, demand or revenue.

Share of voice

A comparison of how often a brand appears against competitors for a defined query or prompt set. The definition must specify the surface, sample, scoring, competitors and period. Without that, it is a persuasive phrase rather than a metric.

Observable, inference and causal claim

An observable is something recorded directly, such as a cited answer on a date. An inference is a reasonable explanation that remains open to revision. A causal claim says a change produced an outcome. These must not be blurred, especially in AI search where systems and interfaces change often.

Confidence

The degree to which the evidence supports a finding. Confidence should rise with repeatable observations, traceable sources and plausible alternatives being considered. It is not a substitute for evidence.

Experiment log

A dated record of a hypothesis, change, evidence, observation, limitation and next test. A visible experiment log is more credible than retrospectively turning every change into a success story.

AI build and delivery terms

Large language model, or LLM

A model trained to predict and generate language-like sequences. It can be very useful for transformation, classification, drafting and reasoning within a defined task. It does not hold a dependable internal database of current facts.

Foundation model

A broadly trained model that can be adapted to many downstream tasks. Large language models are one kind of foundation model.

Inference

The act of running a trained model to produce an output. In ordinary product conversation, this is often what people mean when they say "using the model".

Token

A unit of text a language model processes. Tokens are not exactly words. Token limits influence cost, latency and how much material a model can consider in one request.

Context window

The maximum amount of input and output a model can handle in a single run. Large context is not a licence to throw an entire shared drive at a model. The relevant source, task and acceptance criteria still need to be selected.

Prompt

The instructions and input given to a model. A good prompt can clarify a task. It cannot replace missing evidence, a decision owner or a definition of done.

Context engineering

The more useful name for preparing the information a model needs: task, constraints, trusted sources, tools, output structure and boundaries. It is operational design, not incantation writing.

System instructions

Higher-level instructions that guide an AI system's role, permitted behaviour and constraints. They are important controls, but they are not a complete security boundary against bad data or malicious instructions.

RAG

Retrieval-augmented generation combines a model with a retrieval step that brings relevant external or internal material into the model's context. It can make business knowledge easier to find and trace. It still needs source quality, permissions, evaluation and a way to handle missing information.

Embedding

A numeric representation of text, image or other content that allows a system to compare similarity. It helps retrieval find related material. Similarity is not the same as factual relevance or permission to use a source.

Vector database or vector index

Storage and search infrastructure for embeddings. It enables similarity search at scale. It is an implementation component, not a knowledge strategy.

Chunking

Splitting source material into smaller retrievable pieces. Good chunks preserve meaning, source identity and useful boundaries. Bad chunks detach a claim from its qualification or source.

Retriever and reranker

A retriever finds a first set of potentially relevant sources. A reranker reorders those candidates using a closer relevance test. Both are ways to improve selection. Neither makes weak source material trustworthy.

Tool use or function calling

A model choosing from defined software actions such as searching an approved source, calculating a value or creating a draft. The application performs the actual action. Tool access should be narrowly scoped and logged.

Model Context Protocol, or MCP

An open protocol for connecting AI applications to tools and context sources in a consistent way. It can make integrations easier to manage. It does not turn a tool into an agent, nor does it make an integration safe by itself.

Agent

A software pattern in which a model can plan or select steps, use tools and continue until a task boundary is met. The word is used far too broadly. Describe the actual permissions, tools, stop conditions and human approval points instead.

Agentic workflow

A workflow that gives a model some discretion over task sequencing or tool use. For serious client work, it needs clear scope, evidence rules, safeguards, reversible actions and an owner. "Agentic" is not a quality standard.

Workflow automation

Software that performs predefined steps based on a trigger or rule. Most useful automation is simpler and more reliable than an agentic loop. The right choice depends on how much judgement and uncertainty the task contains.

Harness

The surrounding system that makes model work controlled and repeatable: task definitions, source selection, tools, validation, evaluations, permissions and records. Models can change. A good harness retains the operational memory.

Task contract

A compact definition of the decision, requested outcome, scope, allowed sources, constraints, acceptance criteria and approver for a piece of work. It prevents a capable model from being asked to invent the brief.

Readiness check

A deterministic check that required inputs, permissions, dates, sources and owners are present before work starts. A failed readiness check is useful information, not an invitation to guess.

Source hierarchy

An explicit order of trust and relevance for evidence. For example, a signed client brief may outrank an old slide deck, and a current primary source may outrank a search summary. A hierarchy makes conflicts visible.

Provenance

The record of where information came from, when it was retrieved, what changed it and which source supports a claim. It is the difference between an answer that sounds plausible and one that can be checked.

Deterministic validation

A check with a repeatable pass or fail rule, such as validating a required field, calculation, URL or file format. Use it for what software can verify reliably. Do not pretend it can settle a strategic judgement.

Evaluation, or eval

A deliberate test of whether an AI system behaves acceptably for defined cases. A useful eval has representative inputs, a rubric or expected outcome, a record of failures and a way to rerun it after a change.

Regression test

A repeated test that checks whether a previously working behaviour broke after a change. It matters because a model, prompt, tool or source update can quietly undo an earlier improvement.

Guardrail

A constraint that reduces a known risk, such as restricting a tool, rejecting an unsafe output or requiring a source. Guardrails work best in layers. A single sentence in a prompt is rarely enough.

Approval gate and human in the loop

An approval gate requires a responsible person to review or authorise a step before it takes effect. Human in the loop is the broader idea that a person remains involved. The meaningful question is where that person can stop, correct or own the decision.

Structured deliverable

An output with defined fields, evidence references, gaps, owner and next step rather than free-form prose alone. Structure makes review, reuse and hand-off easier.

The short version

For Search Intent Optimization, make the useful thing easy for the right buyer to find, understand, trust and choose. SEO helps make it accessible and easy to navigate. GEO and AEO help make the entity, answers and evidence easier to recognise and verify. For AI build work, make the task, sources, permissions, checks and approval points explicit before asking a model to help.

The vocabulary will keep changing. The underlying discipline is less exciting and much more durable: clear context, credible evidence, useful work and someone accountable for the outcome.

Changelog

19 September 2026

First draft created as a living glossary. It labels GEO, AEO, LLMO and AI visibility as working language, and does not make unsupported claims about schema, llms.txt, citations or AI-answer outcomes.

21 September 2026

Reframed the glossary around Search Intent Optimization as the buyer-facing umbrella. SEO, AEO and GEO remain defined as useful task labels rather than competing disciplines.