# W01 Slide Script — What GEO Is—and Is Not ## Production conventions - **Format:** 16:9, deep purple / ivory / muted gold; one focal idea per slide. - **Typography:** minimum 30 pt body, 44 pt title; formulas receive a spoken and text interpretation. - **Visual rule:** every diagram is an original instructional synthesis. Shape, labels, line style, and position carry meaning in addition to color. - **Evidence rule:** paper IDs are reading routes, not badges of validity. Boundaries stay on the same slide as the source use. - **Pacing rule:** the pause/check on each slide is optional only where the seminar plan explicitly compresses discussion. ## Slide 01 — The object before the tactic **On-screen text** > W01 · What GEO is—and is not > Essential question: What is the actual object of optimization? Small footer: “Observation ≠ mechanism disclosure.” **Visual** An ivory field with one generated-answer card on the right and six unlabeled purple waypoints fading behind it. A gold bracket encloses the visible answer; a hatched band covers the hidden waypoints. Do not imitate a commercial interface. **Speaker notes** Open without defining GEO. Ask learners to imagine that a product appears in an answer with a citation. Did the entity become visible, did a source become visible, did a claim become supported, or did a user outcome change? Today’s work is to separate those events before discussing intervention. State that the diagram is an abstraction, not a reconstruction of a platform. **Pause/check** Silent 30-second response: write three different things that “visibility” might mean. **Accessibility** Alt text: “A visible generated answer sits in front of six partly hidden process waypoints, emphasizing that the interface does not expose the full mechanism.” The footer is read aloud. ## Slide 02 — One favorable answer, five different claims **On-screen text** - Entity mentioned - Source cited - Claim supported - Source information absorbed - User acted Caption: “Related events; different units.” **Visual** Five vertically aligned cards, each with a distinct icon and unit label: entity–response, source–response, claim–response, claim–source–response, user–task. Thin lines show possible relations but no single arrow joins all five. **Speaker notes** Explain that one response may satisfy any subset of these events. A brand can be mentioned without its page being cited. A page can be cited while an adjacent number is wrong. A source fact can appear without visible attribution. A user can act without clicking. The scientific error is to move from one event to another without evidence. **Pause/check** Ask: “If a citation appears, which of the other four events follow logically?” Expected answer: none without additional evidence. **Accessibility** Provide the five-card content as the same ordered text list. Icons are decorative; unit labels carry the distinction. ## Slide 03 — Outcomes for W01 **On-screen text** By the end, you can: 1. Define a bounded GEO object. 2. Separate objects, stages, metrics, and outcomes. 3. Mark observable, proxy-sensitive, and latent variables. 4. Audit a claim without inventing a closed-system mechanism. **Visual** A four-rung ladder labeled Define → Separate → Mark → Defend. Each rung has a small evidence artifact beside it: scope card, object table, observability map, bounded conclusion. **Speaker notes** Frame the session as methods training. Learners will not leave with a list of production tactics. They will leave with a way to turn ambiguous claims into research objects and to state uncertainty precisely. Preview the studio outputs: a system card, visibility-chain map, metric card, and “what we cannot claim” memo. **Pause/check** Readiness poll: define retrieval and citation in one sentence each. Invite two contrasting answers. **Accessibility** The ladder order is repeated numerically. State that height indicates instructional sequence, not importance. ## Slide 04 — A working definition **On-screen text** **GEO:** the study and responsible intervention of source visibility, use, attribution, and downstream effects in generative search and recommendation systems. Constraints: governed sources · plural outcomes · named surface · evidence and risk limits **Visual** The definition sits inside a purple frame. Four gold corner tabs label the constraints. Outside the frame, grey phrases such as “magic wording” and “one universal score” are crossed out. **Speaker notes** Unpack each phrase. “Study” allows observation and explanation even without intervention. “Responsible intervention” requires factual fidelity and authorization. “Source” means a versioned information object, not only a string. “Plural outcomes” resists a single visibility score. “Named surface” prevents platform-free claims. **Pause/check** Ask learners which word in the definition creates the largest burden of proof. Accept multiple defended answers. **Accessibility** Read the full definition verbatim. Crossed-out phrases are also listed in text: “not magic wording; not one universal score.” ## Slide 05 — The bounded scope tuple **On-screen text** \[ \Omega=\langle S,Q,M,F,T,E,C,B\rangle \] Source · queries · system · surface · time · event · comparator · boundary **Visual** Eight labeled cells arranged as a scope card. Empty cells are visibly marked “unknown” rather than filled. A poor question, “Improve AI visibility,” appears above; a bounded synthetic question appears below. **Speaker notes** Define each slot. Emphasize that a field may be unknown, but the unknown must be explicit. Read the HarborCell example: supplied 40-query panel, frozen open sandbox, named passage event, unchanged comparator, versioned fixture, factual-equivalence constraint. Explain that the tuple does not replace a protocol; it establishes whether the proposed object is even inspectable. **Pause/check** Learners identify the three most important missing slots in “Does clearer content improve AI visibility?” **Accessibility** The formula is followed by the full ordered list. The example is available as plain prose, not only spatial cells. ## Slide 06 — Adjacent fields, overlapping objects **On-screen text** | Practice | Primary visible object | |---|---| | SEO | indexed/ranked link, visit | | AEO | extracted/direct answer | | RAG engineering | controlled retrieval and grounded response | | Digital PR | legitimate third-party publication | | GEO | declared events across a generative pipeline | **Visual** Five overlapping translucent ellipses, each anchored to its primary visible object. GEO crosses the others but does not contain them. No area size implies market size or importance. **Speaker notes** Explain that these are problem formulations, not fixed professional territories. GEO inherits methods from information retrieval, web publishing, measurement, and governance. Transfer is conditional: a high-ranked page may not enter a generated context; a controlled RAG result may not transfer to a web surface. **Pause/check** Ask for one research question that belongs to both RAG engineering and GEO, then name the transfer boundary. **Accessibility** Provide the table as the primary equivalent. State explicitly that overlap signals shared methods, not hierarchy. ## Slide 07 — Nine object types **On-screen text** Entity · Claim · Source · Representation · Query · Response · Attribution · Event · Outcome Rule: “Do not change the unit mid-sentence.” **Visual** A three-row object map: information objects, system/measurement objects, and evaluative objects. Arrows are labeled “about,” “derived from,” “issued to,” “displayed with,” and “aggregates.” **Speaker notes** Define the nine objects with one HarborCell example each. Point out that a citation normally resolves to a source while a mention resolves to an entity. A passage is a representation derived from a source, not the source identity itself. An outcome is valued by a stakeholder; an event is the operational occurrence used to measure it. **Pause/check** Classify “HarborCell,” “480 Wh,” the product page, a chunk, and a displayed footnote. **Accessibility** The relation graph is followed by a text sequence in speaker notes. Every arrow is labeled, so color is redundant. ## Slide 08 — Atomize before you audit **On-screen text** Compound sentence: > “H2 is a safe, long-range battery with a 36-month warranty.” Atomic claims: safety · range · warranty duration **Visual** The sentence is split with vertical cut marks into three claim cards. Each card has blank fields for scope, source, evidence passage, status, and unresolved condition. **Speaker notes** Show why one terminal citation cannot automatically support all predicates. Safety needs a defined property, test, population, and comparator. Range needs a test cycle and operating conditions. Warranty duration needs the current issued terms. Atomization is not a writing trick; it changes the unit of audit. **Pause/check** Ask learners to find a fourth implied claim in “long-range campus option.” Possible answers include suitability for campus use or comparative advantage. **Accessibility** Read the sentence and each resulting claim. Cut marks are reinforced by numbered cards. ## Slide 09 — The conditional visibility chain **On-screen text** Discoverable → Retrieved → In context → Contributes → Attributed → Outcome Footer: “Conditional events, not one conversion funnel.” **Visual** Six rounded nodes with a denominator box beneath each transition. A red-hatched background marks potentially hidden stages; a purple outline marks visible surface events. Each transition has one failure label. **Speaker notes** Introduce the chain as a diagnostic decomposition. Give a failure example at each transition: inaccessible source, not retrieved, removed by context budget, present but unused, used but not attributed, attributed but no declared action. Explain that individual systems may merge, repeat, or omit stages. **Pause/check** Learners choose where “page was cited but the answer misstated its warranty” belongs. The answer spans attribution and fidelity, not simply retrieval. **Accessibility** Supply an ordered table with stage, event, and failure example. Hatching is described in the caption. ## Slide 10 — Discoverability is not retrieval **On-screen text** Discoverable: available to the relevant acquisition process Retrieved: enters the query-conditioned candidate set **Visual** Two nested but non-identical sets. The discoverable corpus contains many documents; a small query-specific candidate set intersects it. Unknown corpus boundaries use dashed outlines. **Speaker notes** Explain that crawlability, fetch success, indexing eligibility, and retrieval can be separate. A robots or sitemap statement can define protocol behavior within its scope but does not establish that a named answer engine fetched, indexed, retrieved, or cited the page. In a closed surface, both denominators may be unknown. **Pause/check** Ask: “What additional log would convert retrieval from a hypothesis into an observation in an open sandbox?” Expected: a query-linked candidate/run record. **Accessibility** Set membership is duplicated in text: “many eligible documents; fewer query-conditioned candidates; boundary may be unknown.” ## Slide 11 — Context is not contribution **On-screen text** In context: available to generation Contributes: shapes a response claim **Visual** A fixed context tray holds four passages. Two connect to generated claims with solid labeled edges; one has no edge; one connects with a dashed “possible/common fact” edge. **Speaker notes** A selected passage may be ignored. A generated claim may come from common knowledge or another source. Distinctive evidence, controlled perturbation, or traceable open-system behavior can strengthen a contribution inference, but the protocol must retain alternative sources and uncertainty. Avoid using token overlap alone as proof of source use. **Pause/check** Ask whether the number “480 Wh” is source-distinctive. Learners should request an alternative-source check. **Accessibility** The four passages and their edge statuses appear in an adjacent text table. Solid and dashed edges have labels. ## Slide 12 — Attribution is not outcome **On-screen text** Attribution: a visible source relation Outcome: a declared consequence for a stakeholder **Visual** A citation marker branches into four possible downstream paths: noticed, clicked, acted without click, and no action. All paths terminate in separate logged or unknown states. **Speaker notes** Explain that attribution correctness, visibility, and user response are separate. A citation may resolve correctly yet fail to entail the adjacent claim. A user may act without clicking, and a click may not reflect persuasion. Downstream effects require consent, event definitions, and an identification design. **Pause/check** Prompt: “Which event log proves that a citation caused a purchase?” Correct response: no single event log; a causal design and assumptions are required. **Accessibility** List all four branches in reading order. Arrow direction indicates sequence only, not causation. ## Slide 13 — The chain rule, without independence **On-screen text** \[ P(Z_D,\ldots,Z_A\mid q,s)=\prod_j P(Z_j\mid Z_{