GEO becomes easier once you can trace the path from question to answer. Part 3 explains LLMs, prompts, tokens, context windows, chunks, embeddings, grounding and RAG without technical fog. You will then know what your website must do.
By Kacper Ruta · GlasBox Studio, Malters/Lucerne · Reviewed: 13 October 2026.
Part 3 of the GlasBox GEO Handbook builds on the sustainable strategy from Part 2. We now open the machine only as far as your business needs. Part 4 will separate mentions from citations and explain how to measure both.
An LLM generates language. The chat interface, search, tools and source access belong to the product layer.
A large context window holds plenty of material. False or conflicting documents remain false.
Clear sections help readers and may help retrieval. Artificial text fragments are not a GEO shortcut.
Short definition: A generative answer system combines a language model with instructions, context and sometimes retrieved sources to produce a new answer.
01 — Which terms do you actually need for GEO?
GEO requires eight core terms and a clear view of the process. The mathematics can stay in the engine room. You must still recognise where information enters, disappears or gains an unsupported addition.
Imagine a supplier explaining your business with twelve technical terms but no decision path. Ask them to show every term against one real customer question. If the answer stays vague, a polished dashboard will not rescue it.
| Term | Short meaning | Relevance to GEO |
|---|---|---|
| large language model (LLM) | generates language from learned patterns | formulates the answer |
| prompt | provides the task and context | defines the immediate question |
| token | is a processing unit | affects volume and cost |
| context window | contains the usable working context | limits material and output |
| chunk | is a retrieved content passage | carries one piece of information |
| embedding | represents meaning as a number vector | enables similarity search |
| grounding | ties an answer to sources | reduces unsupported additions |
| RAG | retrieves sources before generation | brings current data into the prompt |
Google defines these terms in similar ways in its current machine-learning glossary, although products implement them differently. [1] For management, the process matters more than a brand name.
02 — An LLM is the engine, not the whole vehicle
A large language model is a trained model that processes and generates probable token sequences. A product adds the interface, search, file access, tools, rules and source display. Confusing both layers leads to weak GEO conclusions. [1][2]
Suppose one product run uses no web search while another retrieves current sources. The answers may differ even when the technical engines are related. Your brand therefore meets a complete system, not merely a model.
| Layer | Job | Example | GEO question |
|---|---|---|---|
| model | process and generate language | selected language or multimodal model | which knowledge and context are available? |
| product | control use and tools | ChatGPT, Claude, Google Search with AI features | does it search, cite or only generate? |
| answer run | handle one request | a chat with history and sources | what information was available this time? |
| your source | provide a verifiable claim | service page, study, profile or document | is the claim correct, reachable and clear? |
Model names and product features change faster than your company information. Anchor the strategy in reliable sources rather than a version number.
03 — How does an AI answer emerge in simple terms?
An AI answer emerges through input, context, possible retrieval, selection and generation. No provider reveals every internal step. The following model therefore maps the parts you can inspect, not a secret algorithm.
A system splits inputs into tokens and processes them within the available context. When a product retrieves external information, it selects useful passages and adds them to that working context. The model then generates the answer step by step. [3][5][6]
Not every answer run uses the web, the same search or identical sources. A repeated test can therefore return another result. GEO monitoring should record the prompt, product, date, language, location and visible sources.
Why a short explanation of Transformers and self-attention is enough
A Transformer is a neural architecture that processes relationships within token sequences through attention mechanisms. Self-attention estimates which parts of the available context matter for other parts. [1][10]
One practical GEO rule follows: a term needs the right neighbouring context. ‘Payroll’ alone remains open. ‘Payroll administration for a limited company in Lucerne’ joins the task, company form and region.
Test rule: Ask three questions about every claim: which input, which context and which source were present? Without those details, an AI answer is difficult to audit.
04 — How do training, context and live retrieval differ?
Training shapes the model, context supplies one answer run, and retrieval brings in selected external information. Each layer operates at a different time and carries different risks. Separating them makes content planning more realistic.
Training data includes text, code, images or other material from which a model learns patterns. Training does not create a tidy, searchable bibliography inside the model. A later answer cannot automatically identify the training document behind a phrase.
What parameters and self-supervised learning mean
A parameter is a learned numerical value that helps shape model behaviour, not a stored paragraph. Self-supervised learning derives learning signals from the data itself. Neither parameter count nor training method reveals whether a product will retrieve your current page. [1]
| Layer | Timing | What it provides | Typical risk |
|---|---|---|---|
| training | before use | learned patterns and capabilities | stale, incomplete or biased patterns |
| context | during the request | prompt, history, files and tools | excessive, conflicting or disordered material |
| retrieval | during the request | selected external passages | wrong source or unsuitable passage |
| generation | during the answer | newly composed output | plausible addition without enough evidence |
Inference means using a trained model to produce a specific output. This phase matters most to your SME because the question, product rules, current sources and model behaviour meet here.
05 — Why can a convincing answer still be wrong?
A hallucination is a plausible-sounding output that is factually wrong or unsupported. Confident language does not prove truth. Names, figures, dates, quotations and URLs deserve independent checks. [1][2]
Imagine an answer about forming a limited company in Lucerne. The process and tone sound professional, yet a deadline or authority address is invented. The error hides because the prose raises no alarm.
| Claim | Risk without a source | Safer response |
|---|---|---|
| general explanation | medium | check a representative sample |
| current price or deadline | high | open the original source and date |
| legal or tax obligation | very high | require review by the responsible expert |
| verbatim quotation | high | locate the wording in the original |
| URL or contact detail | high | open the link and inspect the destination |
Tokenisation does not automatically cause false URLs. An invented path appears when a model generates a convincing character sequence without reliable retrieval. The slash is innocent; missing verification is not. [2][7]
Open every stated URL.
Compare figures with the original source.
Check the region, date and scope.
Require expert approval for legal, financial and health claims.
Allow the system to state that information is missing.
Store the prompt, answer, sources and review date.
06 — Grounding and RAG reduce risk but cannot remove it
Grounding ties an answer to supplied or retrieved source material. Retrieval-augmented generation, or RAG, first finds relevant information, adds it to the prompt and then produces an answer. [1][6]
Suppose a system must explain your current onboarding list. RAG can retrieve the right page instead of relying only on training patterns. The answer can therefore become fresher and easier to inspect.
| Step | System task | Your task |
|---|---|---|
| 1 | understand the request | know the real customer question |
| 2 | retrieve sources | keep the page reachable and unambiguous |
| 3 | select passages | write claims with their context |
| 4 | augment the context | state date, region and limits |
| 5 | generate the answer | inspect both wording and source |
| 6 | enable action | provide a working next step |
OpenAI describes semantic retrieval through vector stores, while Google explains RAG as a common grounding technique. Both descriptions expose the same principle: retrieval and generation are separate steps. [5][6]
Why a retrieved source is not yet proof
Grounding can reduce false output but cannot eliminate it. A source may be stale, a passage may lose context, or the answer may summarise it incorrectly. Anthropic therefore recommends uncertainty, supporting passages and further verification. [7]
A displayed source first proves only that a visible relationship exists. Part 4 on mentions and citations will test whether the linked page supports the specific claim.
07 — A good prompt is a brief, not a truth machine
A prompt is the input that sets the task, context and desired response format. A precise prompt improves fit. It cannot repair missing facts or weak sources. [1]
The weak prompt ‘What does a company need for payroll?’ leaves the region, legal form and goal open. A better prompt states the Lucerne company, twelve employees, transition date and required result.
| Component | Weak | Better |
|---|---|---|
| role | no guidance | for a managing director without jargon |
| task | ‘explain payroll’ | prepare a payroll-administration handover |
| context | no details | Lucerne limited company, twelve employees |
| boundary | unrestricted answer | state any missing facts clearly |
| format | unspecified | checklist with sources and next step |
| freshness | unclear | verify as of 13 October 2026 |
Anthropic recommends clear structure, relevant examples and distinct source sections for complex tasks. [7] The same principle applies internally: give colleagues a sound brief before polishing the prompt.
Name the audience, task and required decision.
Provide the region, date and relevant starting point.
Separate instructions, sources and examples visibly.
Request evidence for verifiable claims.
Allow an honest ‘not enough information’ response.
Define the format, length and next step.
Your internal template prompt is not the real GEO target. The many questions asked by real people matter more. Your content must remain useful even when nobody uses your perfect briefing.
08 — Tokens and context windows govern space, cost and focus
A token is a unit into which a model splits inputs and outputs. The context window is the complete working space available for one answer run. Both vary across models and products. [3][4]
A token may represent a character, word part, word or punctuation mark. Word count and token count are therefore different. Language, spelling, encoding, images and tool data change usage. [3]
A context window works like a desk, not an unlimited company archive. System instructions, conversation history, files, search results, tool output and the new answer share the available space. [4]
| Component | Uses context? | Management consequence |
|---|---|---|
| system instructions | yes | keep rules short and clear |
| conversation history | usually | old branches can distract |
| uploaded documents | yes | provide only relevant versions |
| retrieved chunks | yes | record selection and origin |
| tool results | yes | limit and structure raw data |
| generated answer | yes | leave enough output space |
API costs may price input, output, cached and reasoning tokens separately. Current rates change and do not belong in a permanent GEO rule. Measure total consumption on representative tasks instead of comparing only list prices. [3]
Why more context does not guarantee more quality
A larger window allows longer documents, but irrelevant context can obscure important evidence. Anthropic calls this quality loss as context grows ‘context rot’. Selection still matters more than raw volume. [4]
Do not provide an entire drive when preparing a proposal. Supply the current service description, valid pricing, exclusions and two suitable examples. The AI needs a clear desk, not a digital attic.
09 — A useful chunk carries meaning and boundaries together
A chunk is a coherent content passage that a retrieval system can store or select. Useful chunks contain enough context for correct interpretation. The product still controls the exact boundaries.
Suppose a passage says only ‘The deadline is 30 days’. Without an event, region and date, the figure is useless. A robust section states what the deadline covers and where the authoritative source sits.
| Passage | Retrieval value | Reason |
|---|---|---|
| ‘The deadline is 30 days.’ | weak | event and scope are missing |
| ‘For this internal handover plan, allow 30 days.’ | better | purpose and example status are clear |
| heading plus self-contained answer | strong | topic and claim stay connected |
| table without an introduction or unit | weak | columns lose meaning outside the page |
| figure with region, period and source | strong | claim remains verifiable |
Anthropic showed in its own retrieval tests that missing document context can weaken chunks. The following table reproduces its published failure rates for retrieval within the top 20 results. [8]
| Tested method | Top-20 retrieval failure | Relative reduction |
|---|---|---|
| embeddings | 5.7% | baseline |
| contextual embeddings plus BM25 | 2.9% | 49% |
| additional reranking | 1.9% | 67% |
The figures come from the domains and configurations tested by Anthropic. They document that retrieval experiment, not a general GEO effect or an expected value for your website.
Why you should not cut your website into artificial mini-chunks
Google does not require artificial micro-passages for its generative Search features. Its systems can understand several topics on one page. Page and section length should therefore follow the audience and subject. [9]
Good structure still matters because people need headings, paragraphs and tables. Keep one central thought per section, while retaining its definition, scope and limitation together.
Open each section with a complete core claim.
State the subject, region and period when they alter meaning.
Keep a heading beside its corresponding answer.
Explain abbreviations at first use.
Give tables an introduction, unit and conclusion.
Avoid pronoun-heavy paragraphs without a clear referent.
Write for readers, not a guessed chunk length.
10 — Embeddings find meaning, not automatic truth
An embedding is a numerical representation that lets systems compare content for similarity. Related meanings may sit close together despite sharing few words. Proximity signals relevance, not accuracy or quality. [1][5]
A query for ‘payroll handover for a small Lucerne limited company’ may retrieve payroll-onboarding content. The exact phrase need not appear. Domain, legal form, region and working step jointly shape meaning.
| Term | What it is | What it does not prove |
|---|---|---|
| embedding | numerical representation of content | subject-matter accuracy |
| vector | the resulting number sequence | meaning that humans can read |
| vector space | comparison space for representations | a public map of the system |
| vector store | index for those representations | automatic source quality |
| similarity score | computed proximity to a query | approval or recommendation |
Multimodal systems can also turn images, audio or video into processable representations. [1] GEO needs no special rule in response: the image, caption and surrounding text should carry the same real claim.
OpenAI describes semantic search as a method that can find suitable content despite few shared keywords. Its search results include relevant chunks, similarity scores and the source file. [5]
Why keywords remain useful
Semantic search does not replace clear wording. Names, product codes, places, standards and specialist terms often supply decisive precision. A system can expand meaning, but it should not need to guess your company name.
Write naturally and precisely. State the service, audience, Lucerne, language and boundaries where they belong. Do not repeat terms mechanically; add useful information instead.
11 — Conversation turns a keyword into a decision path
Conversational systems retain parts of the earlier exchange as context and answer follow-up questions from it. One chat can therefore move a customer from orientation through comparison to action. [1][4]
Imagine a managing director who first asks about payroll administration. She then adds twelve employees, two cross-border workers and a January transition. Every follow-up changes the suitable answer and potential source selection.
| Conversation stage | Example prompt | Required information |
|---|---|---|
| understand | ‘What does outsourced payroll cover?’ | scope and definitions |
| narrow | ‘Does it suit twelve employees in Lucerne?’ | audience, region and limits |
| compare | ‘What remains in-house?’ | roles and handover points |
| verify | ‘Which documents are needed to begin?’ | current checklist and source |
| act | ‘How does an initial meeting start?’ | contact route, preparation and appointment |
A classic keyword list captures only part of this path. Build a prompt panel containing opening questions, follow-ups and objections. Use it for testing, not for producing dozens of near-duplicate pages.
Why local Lucerne details must appear early
Local suitability often depends on small facts: service area, language, meeting format, travel and responsibility. Do not bury them in the footer. Connect each detail with the relevant service and action.
Google names maintained business details and a Google Business Profile as foundations for local visibility. [9] A profile cannot replace a clear service page; both must describe the same real business.
12 — Lucerne example: From thin copy to a reliable answer source
This fictional example continues the Lucerne fiduciary firm from Parts 1 and 2. Example values are not benchmarks. The team rebuilds its ‘Payroll administration’ page without promising an AI recommendation.
The starting point
The existing page contains a sales paragraph and a contact form. It states no audience, handover process, required documents, exclusions, responsible expert or review date. People must ask, and retrieval systems find few self-contained claims.
| Reviewed field | Starting value in the example | Risk |
|---|---|---|
| clear core claims | 2 of 12 | answer remains generic |
| claim with a region | 0 of 12 | local suitability remains open |
| claim with a visible source | 1 of 12 | freshness is hard to inspect |
| defined next step | 1 of 3 | action path breaks |
| named responsible person | 0 | maintenance remains unclear |
The baseline covers only twelve predefined claims and three action paths. It does not measure a GEO ranking. The team can still compare the page cleanly after the revision.
The revised page structure
The new page opens with an answer: who the service fits, what it includes and where it applies. Handover, documents, limits, expert contact, sources and booking follow. Each section resolves one customer question.
| Section | Opening claim | Required detail |
|---|---|---|
| suitability | which firms the service fits | region, size and language |
| scope | which tasks are covered | inclusions and exclusions |
| handover | how the transition works | steps and responsible people |
| documents | what the firm must provide | list, format and timing |
| limits | when further advice is needed | clear escalation point |
| contact | how the first meeting starts | person, channel and preparation |
The structure is not a secret ranking factor. It reduces information loss for people and creates complete sections for different retrieval paths.
The three-layer test
| Layer | Test | Pass condition |
|---|---|---|
| source | review twelve core claims on the page | every claim correct, dated and approved |
| answer | run a fixed prompt panel in defined products | claim and boundary reproduced correctly |
| action | open visible links and contact paths | destination works on mobile in DE-CH and EN |
The team tests ten documented conversation paths under identical starting conditions. Ten is a GlasBox example size, not an industry standard. Each run stores the prompt, product, language, date, answer and visible sources.
Nobody can guarantee that a product will use the page. The team can guarantee that its own source is correct, understandable and reviewable. Responsible GEO begins there.
| Result after testing | Decision |
|---|---|
| source correct, answer correct | observe and repeat regularly |
| source correct, answer wrong | investigate retrieval, context and product limits |
| source unclear, answer wrong | repair the source first |
| answer correct, action broken | repair the contact or conversion path |
| result varies | increase the sample and claim no effect |
The result table blocks one comfortable excuse: a correct answer cannot repair a broken contact path. GEO ends with verifiable customer utility, not with polished prose.
13 — What should you do on Monday morning?
On Monday, choose one important service and follow one customer question through to action. One morning is enough for an initial diagnosis. Do not produce twenty new pages yet.
Choose one revenue-relevant service with clear responsibility.
Write down five real opening questions and five follow-ups.
Mark the answer, region, date, boundary and contact path on the page.
Open every cited original source and every link.
Rewrite the first paragraph as a self-contained answer.
Review every section for complete context.
Test three documented conversation paths in one defined product.
Store errors as tasks with a responsible person and deadline.
Then allow one week for expert approval, technical checks and the English adaptation. A wider prompt panel becomes worthwhile only after the core source is sound.
GlasBox working rule: Repair the source first, retrieval second and presentation last. Otherwise you are polishing an answer on uncertain foundations.
14 — Frequently asked questions
The recurring questions concern practical boundaries more than mathematics. These answers deliberately separate the technology, source and business effect.
Must I understand how a neural network is trained?
No. For GEO, you need the path through input, context, retrieval and answer. Model training matters when you build or assess the systems themselves.
Is ChatGPT the same as GPT?
No. ChatGPT is a product, while GPT names a model family. The product can add search, tools, rules and interfaces.
Does RAG prevent every hallucination?
No. RAG can supply current, reviewable information, but retrieval and summarisation can still fail. Critical claims require checking against the original.
Does every page need short chunks?
No. Google does not require artificial micro-sections for its generative Search features. Write complete, clearly organised answers at the length people need. [9]
Are embeddings a new ranking factor?
No. Embeddings are a representation and similarity-search technique. Their use proves neither a confirmed ranking factor nor a guarantee of a mention.
How long should a good prompt be?
As long as necessary and as short as possible. The goal, context, boundaries, sources and format must be clear. Repetition without new information only consumes attention and tokens.
Can a large context window replace our knowledge base?
No. A large window holds more material but manages neither validity nor responsibility. Your knowledge base still needs versions, sources, access rules and review cycles.
Why do identical prompts produce different answers?
Several factors may change: model, product, conversation history, search results, selection, system rules and generation. Record every test run in full.
15 — Conclusion: GEO starts with the path to an answer
GEO becomes manageable when you separate the model, product, context, retrieval and source. An LLM generates language, grounding supplies material, chunks carry passages and embeddings help retrieval. No single component proves truth.
You cannot open the platform, but your source remains reviewable. Verify claims, boundaries and dates, then document every answer run. Part 4 will show when a brand is merely mentioned and when a source receives a genuine citation.
The rule for Part 3: Do not optimise for a guessed algorithm. Build complete, current and verifiable information for people, then test how systems retrieve and reproduce it.
Sources and editorial basis
Technical definitions and product claims use the following primary sources, reviewed on 13 October 2026. The process model, Monday checklist and Lucerne example are GlasBox methods. Two diagrams are unmeasured concept models; the sole results table uses clearly bounded Anthropic test data.
Definitions of LLM, prompt, token, grounding, hallucination, embedding, self-attention and RAG.
Why probabilistic token generation creates fluent language and a risk of precise fabrication.
Token categories, model and language differences, context limits, API usage and cost implications.
The context window as working memory, what it contains and the risk of quality loss with excess context.
Semantic search using embeddings, vector stores, chunks, similarity scores and source files.
Grounding and RAG as retrieval, prompt augmentation and subsequent generation.
Allowing uncertainty, binding claims to supporting passages and retaining checks for critical information.
The experiment, limits and published retrieval failure rates for contextual retrieval methods.
RAG, query fan-out and the clarification that artificial chunking is not required for Google Search.
The primary paper introducing the attention-based Transformer architecture.
Are your pages understandable, retrievable and verifiable?
We clarify your starting point, the most important sources and the next verifiable step. We agree the scope and deliverables of any audit in the proposal.
Request an initial conversation