Generative Search and Brand Visibility Engineering

Interactive 01 · W01–W03

Diagnose one stage without claiming the next

Select a stage. The feedback separates what can be observed from what remains hidden and names the inference that must not cross the boundary.

01

Diagnostic question

Does a stable, evidence-bearing source artifact exist?

Observable evidence: Rendered content, response status, canonical identity, update date, provenance, and declared access controls.
Still hidden: Whether any particular system has discovered, stored, or interpreted the artifact.
Invalid inference: A live page does not prove crawling, indexing, retrieval, citation, or model training.
Defensible next method: Freeze the artifact, record its version and hash, and separate authored claims from implementation metadata.

Open the complete non-interactive table

Stage	Observable	Hidden	Invalid inference
01 · Publish	Rendered content, response status, canonical identity, update date, provenance, and declared access controls.	Whether any particular system has discovered, stored, or interpreted the artifact.	A live page does not prove crawling, indexing, retrieval, citation, or model training.
02 · Discover	Internal links, sitemap membership, robots directives, fetch logs you lawfully control, and protocol responses.	The complete discovery frontier, scheduler priority, or undisclosed crawler policy of a closed platform.	Sitemap presence or a logged fetch does not establish indexing, ranking, or use in an answer.
03 · Eligible	Declared directives, status codes, parsable content, structured metadata, and documented policy constraints.	Undisclosed index admission rules, quality classifiers, deduplication, or internal safety filters.	Technical eligibility is not evidence of index inclusion or preference.
04 · Retrieve	Ranked lists and scores only in an open or instrumented retrieval system; visible results on a closed surface are outcome observations.	The closed system candidate set, query rewriting, embeddings, weights, personalization, and unreturned candidates.	A visible citation cannot reveal every retrieved document or the retrieval algorithm that produced it.
05 · Select / cite	Answer text, claim spans, displayed citations, cited passages, and annotations produced under a declared protocol.	Unshown context, latent attribution, generation weights, internal judge signals, and sources used without display.	Citation, entailment, source quality, completeness, and source-distinctive absorption are different measurements.
06 · Outcome	Predeclared surface metrics, referrals or actions you are authorized to measure, sampling frame, missingness, and uncertainty.	Counterfactual outcomes without a design, private downstream behavior, and future performance under drift.	A before/after change is not automatically causal, stable, beneficial, or generalizable.

Interactive 02 · W06–W08

See why repetitions do not replace time coverage

Change the design of a synthetic repeated-measurement panel. Values teach the variance structure only; they are not observations from a commercial answer engine.

Illustrative 95% half-width

12.9 percentage points

Total observations: 12
Within-block SD: 18 pp
Between-block SD: 7 pp

Design readingBalanced synthetic design: retain block-aware uncertainty and the declared population boundary.

Frozen teaching modelSE = sqrt(0.18² / (runs × blocks) + sigma_block² / blocks)

The display uses 1.96 × SE as an illustrative normal-approximation half-width. A real study must justify its estimator, dependence model, interval procedure, missingness rule, and population.

Open the predefined sensitivity table

Runs / block	Blocks	Condition	Total observations	Illustrative half-width
2	1	Stable blocks	2	25.3 pp
8	1	Stable blocks	8	13.1 pp
4	4	Mild drift	16	11.2 pp
8	4	Eventful drift	32	15.1 pp

Interactive 03 · W04

Compare lexical, semantic, and hybrid retrieval

Rank a six-document teaching corpus with frozen, hand-authored scores. The exercise exposes a mechanism trade-off; it is neither a benchmark nor a model of a commercial engine.

Frozen ranking · hybrid

Method intent: repeated observations, denominators, and time blocks.

01
D01 · Repeated-panel measurement protocolExact terms and the intended longitudinal method.
0.93
02
D02 · Citation correctness annotation guideStrong citation overlap, but little time-series design.
0.67
03
D03 · Temporal drift in answer surfacesFew exact terms; strong semantic match to change over time.
0.59
04
D04 · Single-screenshot visibility caseLexically similar but methodologically insufficient.
0.58
05
D05 · Brand tone writing checklistShares domain vocabulary, not the measurement intent.
0.29
06
D06 · Access and collection safety policyImportant constraint, peripheral to the query intent.
0.27

Diagnostic readingThe top result is D01. Compare its lexical and semantic components; a high combined score still does not establish source quality or answer use.

Frozen teaching rulehybrid = w × sparse + (1 − w) × dense

Real systems require a declared corpus, tokenizer, model, normalization, fusion rule, tie policy, and evaluation set. Changing any of them creates a new method version.

Open all predefined top-three rankings

Query	Sparse top 3	Dense top 3	50/50 hybrid top 3
How should citation visibility be measured over time?	D01 0.94 · D04 0.83 · D02 0.75	D01 0.91 · D03 0.88 · D02 0.58	D01 0.93 · D02 0.67 · D03 0.59
Which source supports a generated answer claim?	D01 0.95 · D02 0.62 · D04 0.44	D01 0.93 · D02 0.91 · D06 0.38	D01 0.94 · D02 0.77 · D04 0.40
How should a suspicious source change be triaged?	D06 0.71 · D03 0.13 · D01 0.08	D06 0.92 · D03 0.47 · D01 0.42	D06 0.81 · D03 0.30 · D01 0.25

Interactive 04 · W02 & W05

Test the edge, not the citation count

Select one atomic claim and one evidence standard. The same citation graph can pass presence while failing entailment, independence, or causal identification.

C02 · Source-content claim

Source A states that the threshold is 0.75 for the named protocol version.

SRC-A · protocol v2.1entails

The frozen passage states threshold = 0.75 and names v2.1.

Origin: Publisher A · displayed citation

BLOG-B · implementation notepartial

Repeats 0.75 but omits the protocol version.

Origin: Publisher B · not displayed

Open the complete claim–source decision table

Claim	Presence	Entailment	Independent corroboration	Boundary
C01 · The captured answer displayed a citation to Source A at 10:32 UTC.	Pass — 1 displayed citation edge(s) exist.	Pass — 1 source edge(s) entail the atomic claim.	Fail — Only 1 independent entailing origin(s); corroboration is not established.	Observation claim
C02 · Source A states that the threshold is 0.75 for the named protocol version.	Pass — 1 displayed citation edge(s) exist.	Pass — 1 source edge(s) entail the atomic claim.	Fail — Only 1 independent entailing origin(s); corroboration is not established.	Source-content claim
C03 · The intervention caused a nine-percentage-point visibility increase.	Pass — 1 displayed citation edge(s) exist.	Fail — No preserved passage entails the whole atomic claim.	Fail — Only 0 independent entailing origin(s); corroboration is not established.	Causal claim
C04 · The observed effect generalizes to every generative answer engine.	Pass — 1 displayed citation edge(s) exist.	Fail — No preserved passage entails the whole atomic claim.	Fail — Only 0 independent entailing origin(s); corroboration is not established.	Universal generalization

Interactive 05 · W09–W12

Decide what stayed equivalent

Inspect a frozen before/after fixture under three different contracts. Byte equality, factual equivalence, and one-factor interpretability answer different questions.

V03 · frozen diff

JSON-LD encoding of an existing fact

No new visible fact; still a registered metadata intervention, not a ranking promise.

Changed dimensions

structured metadata

Visible change: No
Claim change: No
Identity change: No

Open the full equivalence matrix

Variant	Changed dimensions	Byte equivalent	Claim + identity equivalent	Single-factor interpretable
V00 · Frozen baseline copy	None	Pass — Files are byte-equivalent.	Pass — Atomic claims and source identity remain equivalent.	Fail — No intervention exists to estimate.
V01 · Whitespace-only serialization	serialization	Fail — Diff contains: serialization.	Pass — Atomic claims and source identity remain equivalent.	Pass — One registered dimension changed: serialization.
V02 · Unsupported “industry-leading” claim	visible copy, claim set	Fail — Diff contains: visible copy, claim set.	Fail — The atomic claim set changed.	Fail — Factual-integrity gate blocks the candidate.
V03 · JSON-LD encoding of an existing fact	structured metadata	Fail — Diff contains: structured metadata.	Pass — Atomic claims and source identity remain equivalent.	Pass — One registered dimension changed: structured metadata.
V04 · Canonical source identity changed	canonical identity	Fail — Diff contains: canonical identity.	Fail — Source identity changed even though claims did not.	Pass — One registered dimension changed: canonical identity.
V05 · Heading plus metadata changed together	visible structure, structured metadata	Fail — Diff contains: visible structure, structured metadata.	Pass — Atomic claims and source identity remain equivalent.	Fail — More than one registered dimension changed.

Interactive 06 · W14–W15

Triage without operationalizing abuse

Apply a conservative tabletop rule to synthetic signals. The tool teaches containment, provenance, privacy, and release control; it contains no live payloads, evasion steps, or platform-specific attack instructions.

Conservative tabletop decision

0illustrative signal points

01 · PreserveFreeze input, output, source version, timestamps, and hashes.
02 · ContainBlock affected release paths; do not probe a live target.
03 · ReviewSeparate integrity, safety, privacy, and measurement hypotheses.
04 · RecoverTest rollback and document residual risk before reopening.

Open the seeded triage cases

Case	Signals	Decision	Reason
T01	retrieval	Review	One anomaly; preserve and compare blocks.
T02	provenance, spread	Block	Lineage failure with repeated exposure.
T03	instruction	Stop	Untrusted instruction-like content is quarantined.
T04	personal	Stop	Privacy response overrides ordinary analysis.

Interactive 07 · W01–W16

Filter the complete 42-paper atlas and 86-entry reference universe. Project status and evidence ceilings describe permitted use; they are not quality scores or endorsements.

128 of 128 records

PAPER-01

Cultural Encoding in Large Language Models: The Existence Gap in AI-Mediated Brand Discovery

Huang, Junyao; Situ, Ruimin; Ye, Renqin · 2025

Platforms & field evidence · Quarantined / publication and time provenance unresolved

Evidence boundaryTime-provenance quarantine: the local full text contains collection/API dates after the 24 August 2026 course freeze; use only to teach provenance failure until independently resolved.

PAPER-02