# W01 No-Video Equivalent Transcript

## Status and use

This document is a stand-alone text alternative for the planned W01 seminar. It contains the scientific content, visual descriptions, questions, and transitions needed to follow the lesson without video or slides. Every viewing label and minute value associated with W01 means a **planned teaching-duration equivalence for this text and its instructor delivery**. The durations below are instructional planning budgets, not recording timestamps or evidence of a completed rehearsal. No W01 recording exists in this package, and no classroom or rehearsal timing has yet been observed.

## Planned chapter budget

| Chapter | Planned seminar budget | Function |
|---|---:|---|
| 1. Opening contrast | 6 minutes | Make “visibility” ambiguous on purpose |
| 2. Define the object | 11 minutes | Introduce the GEO definition and bounded scope tuple |
| 3. Separate the units | 13 minutes | Distinguish entities, claims, sources, responses, and outcomes |
| 4. Build the conditional chain | 13 minutes | Separate six stage events and denominators |
| 5. Mark observability | 12 minutes | Contrast open traces, surface evidence, proxies, and latent variables |
| 6. Define measurement objects | 10 minutes | Separate mention, citation, absorption, fidelity, prominence, and action |
| 7. Apply evidence ceilings | 8 minutes | Route sources by claim class and boundary |
| 8. Audit HarborCell | 9 minutes | Move from raw input to a bounded synthetic conclusion |
| 9. Reject a causal shortcut | 5 minutes | Diagnose a persuasive before/after counterexample |
| 10. Exit with a question | 3 minutes | Produce a bounded research object |

Total planned seminar budget: 90 minutes.

## Chapter 1 — Opening contrast

Imagine a generated answer that names a product and displays a citation beside it. Before we decide whether that is a successful result, write down what you think became visible.

You might say the product became visible. That is an entity-level event. You might say the product page became visible. That is a source-attribution event. You might say the product’s evidence influenced the answer. That is a claim–source–response event. You might say the product was recommended early in the answer. That is a prominence or recommendation event. You might say a user followed the recommendation. That is a downstream action event.

All five readings can refer to the same screen, but they do not refer to the same unit. A product can be named while its page is absent. A source can be cited while the adjacent statement is false. A fact can appear without a visible citation. A user can act without visiting a cited page. If we call all of these “visibility,” the word hides the research design.

The first visual shows a generated answer in the foreground. Behind it are six partly hidden process waypoints. The visible answer has a solid border. The hidden region is marked by both hatching and the label “not observed here.” The visual does not depict a commercial product and does not claim that a platform has six internal modules. Its purpose is to show a simple fact: an interface is not a complete mechanism trace.

Pause and check: If one citation appears, which of the following necessarily follows—entity mention, claim support, source absorption, favorable sentiment, or user action? None follows logically without additional evidence. Even the entity may not appear if the citation supports a general background claim.

## Chapter 2 — Define the object

Our working definition is: **generative engine optimization is the study and responsible intervention of source visibility, use, attribution, and downstream effects in generative search and recommendation systems**.

“Study” matters because observation, diagnosis, and measurement can be valuable even when no intervention is justified. “Responsible intervention” means that exposure is constrained by factual fidelity, evidence, authorization, accessibility, and risk. “Source” means a governed information object with an issuer, version, date, method, claims, and correction path. “Visibility, use, attribution, and effects” are plural because they may disagree. “System” and “surface” must be named because a file-grounded chatbot, an open RAG sandbox, a web answer, and a tool-using recommendation agent expose different variables.

We represent a bounded study with eight slots: source, query distribution, named system, user-facing surface, time window, target event, comparator, and evidence boundary. The notation is omega equals the ordered tuple S, Q, M, F, T, E, C, B. The notation is only a checklist; it does not make an ambiguous protocol scientific.

Compare two questions. “How can we improve AI visibility?” specifies none of the eight slots. “For the supplied campus-mobility query panel, in frozen sandbox version 1.2, does an evidence-equivalent heading treatment change top-five candidate inclusion relative to the unchanged page, without changing any product claim?” names a source treatment, query panel, system version, event, comparator, and factual constraint. The second question still needs sampling, code, and analysis rules, but a reader can now imagine evidence that would contradict it.

GEO overlaps with SEO, answer-engine optimization, RAG engineering, and digital public relations. The overlap should be drawn around objects and methods rather than professional labels. SEO can inspect indexing and ranked links. RAG engineering can control candidate retrieval and context in a declared corpus. Digital public relations can create legitimate third-party publication. GEO asks about declared events across a generated-answer pipeline. A result in one setting is a transfer hypothesis for another setting, not an automatic conclusion.

The W01 reading route illustrates source boundaries. P16 is useful for the early GEO formulation and fixed-context intervention evidence. Its target page was already in a selected source set, so the study does not estimate ordinary open-web discovery or retrieval. P19 is useful for risk taxonomy and agenda setting; as a position paper it does not estimate prevalence, harm, or provider effectiveness. P06 maps search and LLM research conceptually but does not disclose a product architecture or perform a systematic evidence synthesis. P15 describes heterogeneous API observations collected in August 2025; it does not manipulate content and cannot establish a tactic effect or current system behavior.

Pause and check: In the phrase “clearer content improves AI visibility,” identify at least three missing slots. Likely answers include the source population, definition of clearer, query distribution, named surface, event metric, comparator, and time window.

## Chapter 3 — Separate the units

We need nine object types.

An **entity** is the organization, product, person, place, or concept under discussion. A **claim** is a minimal proposition whose meaning includes scope conditions. A **source** is a governed information object that can support or contest claims. A **representation** is a passage, chunk, image description, embedding, index entry, or other derived unit. A **query** is a measurement instrument drawn from a population or deliberately constructed test set. A **response** is the generated text and visible organization shown on a named surface. An **attribution** is a displayed relation such as a citation marker or source panel. An **event** is an operational occurrence. An **outcome** is a consequence valued by a stakeholder and may aggregate several events.

The primary text-only visual is a relation list. A claim is about an entity. A passage is derived from a source. A query is issued to a system surface. A response is displayed with zero or more attributions. An event applies a declared rule to one or more objects. An outcome aggregates events for a declared stakeholder. Each relation has a verb; proximity in a diagram is not enough.

Consider “The brand gained citations.” A brand is an entity, while a citation typically resolves to a source. If a review site cites a brand name, the brand is mentioned but its governed page is not cited. If an answer cites the company documentation without naming the company in visible prose, the source is attributed but an entity-mention rule may be false. Before computing a rate, decide which pair defines the unit.

Claims must also be atomized. “The HarborCell H2 is a safe, long-range battery with a 36-month warranty” contains separate propositions about safety, range, and warranty. “Safe” requires a safety construct, operating conditions, comparator, and evidence. “Long-range” requires a defined test cycle. Warranty duration requires current issued terms. One citation placed at the end of the sentence cannot automatically support all predicates.

An atomic claim record should include subject, predicate, object or value, time, locale, conditions, evidence passage, source version, disposition, and unresolved questions. The dispositions used in W01 are supported with scope, partially supported, contradicted, unsupported, and unresolved. These labels are not interchangeable. A contradicted claim has evidence against it. An unsupported claim lacks adequate evidence in the supplied set. An unresolved claim may have competing explanations that the protocol cannot distinguish.

Pause and check: Classify the following objects: “HarborCell H2” is an entity. “Nominal capacity is 480 Wh” is a claim. “Specification version 2.1” is a source. The warranty paragraph is a passage. Marker “[1]” is an attribution display. A recorded click would be an event. Purchase completion could be a downstream outcome under a stated task.

## Chapter 4 — Build the conditional chain

We use six stage events: discoverable, retrieved, in context, contributes, attributed, and outcome.

Discoverable means the source is available to the relevant acquisition process. Retrieved means the source or one of its representations enters a query-conditioned candidate set. In context means the source survives selection and is available to generation. Contributes means source information shapes generated content. Attributed means the contribution receives a visible and sufficiently faithful source relation. Outcome means the declared user, source, platform, or social consequence occurs.

The stage variables are labeled Z sub D, R, C, G, T, and A. For an outcome that requires all stages, the joint event can be written as successive conditional probabilities. Read the expression in words: the joint event equals the probability of discoverability multiplied by each later event conditioned on earlier events, query, and source. This is the chain rule. It does not assume the stages are independent.

Dependence is the substantive point. A revision could make a passage lexically easier to retrieve while omitting material qualifications. A longer evidence block could improve support but lose space under a context budget. A source could enter context and contribute nothing. A system could generate a claim and attach a citation in a later interface step. A change can help one transition and harm another.

Every arrow needs a denominator. Retrieved in eighteen cases out of what? All sampled queries, successful retrieval calls, source-eligible queries, or responses containing any citation? If the system does not expose the eligible corpus or candidate list, write “unknown.” A blank denominator is not permission to substitute the final response count.

The visual equivalent is an ordered six-row table. Row one, Discoverable: possible failure, unavailable to acquisition. Row two, Retrieved: possible failure, not in candidate set. Row three, In context: possible failure, removed under ranking or context allocation. Row four, Contributes: possible failure, present but unused. Row five, Attributed: possible failure, used but absent or misleading citation. Row six, Outcome: possible failure, no declared action or effect. A separate note says the sequence may be merged, repeated, reordered, or partly absent.

Pause and check: A page is displayed as a citation, but the answer gives the wrong warranty duration. Where is the problem? The citation display is observed, while claim fidelity fails. Retrieval, context, and contribution may remain unknown. The case cannot be reduced to a single failed stage.

## Chapter 5 — Mark observability

Observability belongs to a variable–protocol pair. We use four statuses.

**Surface-observable** variables include response text, displayed citations, visible order, and permitted interaction traces. **Controlled-system observable** variables include candidates, scores, selected passages, prompts, checkpoints, and random seeds when an open or frozen system logs them. **Proxy-sensitive** variables include source contribution when it is inferred through distinctive claims, perturbations, or attribution rather than directly traced. **Latent in the study** variables include hidden candidate sets, proprietary ranking features, training inclusion, and unrecorded personalization.

The same variable can move between statuses. Candidate rank may be directly logged in an open sandbox and latent on a commercial surface. Context membership may be visible in a file-grounded experiment and hidden in a web answer. This is why “the algorithm prefers our page” is usually too strong for surface evidence.

A screenshot can establish a narrow proposition: under the documented collection conditions, the captured surface displayed the represented text and source markers. It does not alone establish frequency, the hidden candidate set, ranking logic, training data, or the causal reason for the response. That narrow ceiling does not make screenshots worthless. It makes them properly typed evidence.

We also separate epistemic statuses. An **observation** reports a recorded event. An **interpretation** applies a stated rule, such as resolving a citation URL to a source ID. A **mechanism hypothesis** proposes an explanation and names evidence that could distinguish it. A **causal effect claim** requires a comparison world and assumptions. An **unsupported promise** asserts an outcome beyond the evidence.

For example: “Marker [1] was displayed” is an observation. “Marker [1] resolves to source S-001 under the resolver rule” is an interpretation grounded in a visible route. “The revised heading increased retrieval” is a mechanism hypothesis until retrieval is observed or tested. “Randomized heading assignment changed retrieval incidence by the estimated amount in the frozen system” can be a causal claim if the design and analysis support it. “The heading will produce the same outcome on every engine” exceeds that design.

A minimal evidence packet records the system and surface, caller-visible configuration, date and timezone, locale and language, permitted account or session state, exact query, collection order, response, resolved citations, failure state, collection method, and checksums. Unknown fields remain unknown.

Pause and check: Rewrite “The engine prefers Source A” from one screenshot. A defensible version is: “In observation O-017, Source A appeared as the first displayed citation under the recorded query and surface conditions.” This says nothing about hidden preference or frequency.

## Chapter 6 — Define measurement objects

Mention, citation, absorption, fidelity, prominence, and action require separate cards.

**Mention** has an entity–response unit. A resolved entity occurs in a visible response. Mention does not establish source use, accuracy, recommendation, or sentiment.

**Citation** has a source–response unit. A displayed attribution resolves to the governed source. Citation does not establish entailment, source contribution, user notice, or action.

**Absorption** has a claim–source–response unit. Source-distinctive information is reflected in the answer. Common facts cannot identify one source. A strong protocol checks alternative sources and retains an unresolved state.

**Fidelity** has a claim–response unit. The response preserves the proposition’s value, conditions, and uncertainty. Correct source identity is not enough if a number changes or a condition disappears.

**Prominence** has a span–response unit. It may describe order, normalized position, answer share, or another declared exposure function. It does not automatically represent attention, favorability, or utility.

**Action** has a user–task or event-log unit. A click, follow-up, purchase, or tool action requires authorization and a declared event rule. An action record alone does not explain why it occurred.

For every metric, write the construct, event rule, unit, eligible population, numerator, denominator, weighting, repeated-run rule, missingness treatment, aggregation, uncertainty method, and prohibited interpretation. Two metrics with the same label but different denominators are different quantities.

Pause and check: Should ten additional mentions offset one materially false warranty claim? There is no automatic answer. A responsible objective treats factual fidelity as a non-compensable gate or explicitly justified policy constraint, not a hidden weight chosen after seeing results.

## Chapter 7 — Apply evidence ceilings

An evidence ceiling is the strongest claim class an item can ordinarily support after its identity and method are checked. It is not a prestige score.

E0 materials support the fact that an actor made a public assertion. E1 official documents, policies, and standards can support declared behavior, status, terminology, or scoped requirements. E2 method-rich tutorials, white papers, and code walkthroughs can support an inspectable procedure. E3 benchmarks and observational studies can support bounded associations or performance under their protocols. E4 controlled comparisons can support an effect in the experimental environment, subject to assumptions. E5 independent multi-setting replication or a strong field design can support a more transportable estimate, while scope and heterogeneity remain.

The best source depends on the claim. A first-party help page can be strongest for the name and behavior of a product setting. It cannot reveal an unpublished ranking weight. A standard can state a requirement in scope; it cannot prove implementation or GEO performance. A repository can expose code; it does not independently validate a paper result. A controlled benchmark can estimate a treatment effect in its environment; it cannot establish the current behavior of a different closed system.

Identity comes first: issuer, type, title, version, date, status, method, artifacts, license, and route. A polished diagram or institutional logo does not supply missing method fields. An external figure also needs a reuse record. If rights are unclear, cite the concept and construct an original diagram with an explicit synthesis caption.

Pause and check: Which source would you use to establish a named platform setting? First-party documentation within its date and scope. Which source would you use for an intervention effect? A suitable comparison design. Neither substitutes for the other.

## Chapter 8 — Audit HarborCell

Everything in this case is fictional. Source S-001 is Blue Marsh Mobility’s HarborCell H2 Product Specification version 2.1. It states 480 Wh nominal capacity with operating variability, an IP54 enclosure under a specified laboratory configuration with exclusions, and a 24-month limited warranty. It explicitly does not provide comparative safety or real-world range findings.

The synthetic query asks which campus e-bike battery is best for long rainy commutes. The static answer says: “HarborCell H2 has 480 Wh capacity, an IP54 enclosure, a 36-month warranty, and is the safest long-range campus option [1].” Marker [1] resolves to S-001.

Atomize the response into five propositions. The 480 Wh proposition is supported with scope. The IP54 proposition requires restoration of material conditions and exclusions. The 36-month warranty proposition is contradicted by the source’s 24-month term. “Safest” is unsupported and explicitly outside the source’s evidence. “Long-range” lacks a test cycle, vehicle, rider, route, weather, and comparator.

Now map the stages. Discoverability, candidate retrieval, and effective context are unknown because the fixture has no internal trace. Source-specific contribution is unresolved because 480 Wh and IP54 may occur in alternative sources. Citation display is observed, but faithfulness varies by claim. No downstream action is observed.

The bounded conclusion is: In synthetic observation O-001, the response mentioned the fictional HarborCell H2 and displayed a citation resolving to fictional source S-001 version 2.1. Among five atomized propositions, one is supported with scope, one requires scope restoration, one contradicts the source, and two are unsupported by the supplied source set. Acquisition, retrieval, context, source-specific contribution, and downstream action are not observable. The record estimates no population frequency, comparative quality, hidden mechanism, intervention effect, or real-platform behavior.

Pause and check: What is the most important positive event? The resolved citation and entity mention are observed. What is the most important failure? A material warranty contradiction. What is the strongest unknown? Depending on the research question, retrieval or source-specific contribution may be most important, but neither may be invented.

## Chapter 9 — Reject a causal shortcut

Now imagine two screenshots. Monday morning, the original page is not cited. Monday noon, the publisher shortens its headings. Tuesday morning, the revised page is cited. A presentation draws a causal arrow from the heading revision to the citation.

The observations are real within the hypothetical record, but the arrow is not identified. Treatment is confounded with time. Ordinary response variation is unmeasured. Query, session, locale, competing sources, index state, system alias, and citation interface may differ. The content diff has not yet established factual equivalence.

The proper response is not to discard the screenshots. Record them as two visible states and convert the mechanism story into a testable hypothesis. Freeze both variants, audit the diff, define the target stage, use repeated blocked or randomized comparisons where permitted, preserve failures, and record system state. If stage traces are required, use a controlled system rather than pretending the closed surface exposed them.

Pause and check: Name four alternatives and one design change that would distinguish them.

## Chapter 10 — Exit with a question

We finish with one template:

> For which sources and queries, on which system surface and dates, does which permitted intervention change which event relative to what comparator, with what uncertainty and evidence limits?

Write one sentence that fills the slots for your studio case. Then name the strongest unknown—the missing variable most likely to change your interpretation. Unknown is a legitimate field value. It is more scientific than a plausible but unobserved mechanism.

The W01 summary is compact. Objects come before tactics. The generated answer is one surface in a conditional process. Entity, source, claim, citation, response, and outcome have different units. Observability depends on the protocol. Mention, citation, absorption, fidelity, prominence, and action require different denominators. Sources carry claim-relative ceilings. Negative, contradicted, and unresolved results stay in the record.

Transition to studio: open `WORKED_CASE.md`, preserve the frozen inputs, and build the system card before drawing the chain. Stop if the task would require personal data, access circumvention, unapproved automation, or a claim stronger than the evidence packet.