Information retrieval
Documents, queries, indexes, ranked lists, relevance, Recall@k, MRR, and NDCG. A bridge note and supplied run files support refresh.
Course contract
Generative Search and Brand Visibility Engineering
A 16-week, research-intensive course in advanced studies in generative engine optimization, measurement, evidence engineering, intervention, and responsible practice. The course is English-first, local-first, and account-free on its core path.
- Format
- 16 weeks · 90-minute seminar + 90-minute studio/clinic
- Weekly load
- 5–7 hours
- Status
- Course packages v0.3 · Notes v0.6
Positioning
Who this course is for
The course is designed for graduate students, research engineers, technical strategists, information-retrieval practitioners, and advanced content/evidence leaders who need to evaluate generative-search claims rather than repeat them. It assumes comfort reading empirical papers and inspecting structured data.
It is not a beginner marketing course and not a platform-specific certification. A non-coding learner can complete conceptual and supplied-output routes, but the graded technical route expects basic Python, shell, statistics, and version-control literacy.
Prerequisites and bridges
Prepare before Week 01
Empirical reasoning
Units, populations, samples, estimands, uncertainty intervals, missing data, and association versus intervention effects.
Reproducible work
CSV/JSON inspection, Python 3.10+, command line, checksums, environment files, and a minimal Git workflow.
Learning outcomes
By the end, a learner can
- 01
Define a bounded generative-search research object and its observable events.
- 02
Route atomic claims to fit-for-purpose, versioned evidence.
- 03
Decompose discovery, retrieval, ranking, generation, attribution, and outcome stages.
- 04
Implement or audit sparse, dense, hybrid, reranking, and context-selection baselines.
- 05
Measure mention, citation, entailment, absorption, and referral without conflation.
- 06
Construct query samples and repeated observations with explicit uncertainty.
- 07
Design reversible interventions and defensible causal comparisons.
- 08
Model manipulation, privacy, accessibility, and governance risks.
- 09
Package a traceable and reproducible evidence release.
- 10
Defend, narrow, or withdraw a claim when the evidence requires it.
Instructional rhythm
Prepare, explain, practice, revise
Orient and diagnose
Essential question, outcomes, bounded reading/viewing route, readiness check, and a preparation artifact.
Explain and interrogate
Mechanism or evidence model, worked micro-example, assumptions, counterexample, and claim-boundary discussion.
Produce an artifact
A 70–80 minute core artifact route with frozen inputs and a stop condition, plus 10–20 minutes reserved for setup, accessible pacing, feedback, recovery, and packaging.
Complete and revise
Lab checkpoint, five retrieval questions, methods reflection, peer/instructor feedback, and a versioned revision.
Curriculum spine
Four phases, four gate artifacts
Observe
Objects, evidence, and the information pipeline
Define what can be observed before making claims about a closed system. Build the query, source, claim, response, and event objects that every later analysis reuses.
Gate artifactAuditable surface map + Claim–Evidence–Source ledger
Explain
Attribution, measurement, and identification
Separate retrieval from use, citation from absorption, repeated-sampling variance from drift, and an observed association from an intervention effect.
Gate artifactVisibility protocol + metric specification + pre-analysis plan
Intervene
Evidence engineering across source ecosystems
Design bounded interventions in content, entities, metadata, and multimodal assets while preserving factual integrity and a reversible change record.
Gate artifactVersioned intervention package + ablation design
Validate
Experiments, safety, incentives, and synthesis
Estimate effects under drift, test failure modes, audit manipulation risk, and defend an evidence-backed result without overstating external validity.
Gate artifactReproducibility bundle + governance memo + oral defense
Assessment
The grade follows the evidence trail
20%Methods memosTwo concise critiques of a paper, protocol, or industry claim.
30%Eight lab artifactsLineage, reproducibility, and decision quality; 3.75% each.
10%Replication noteReproduce or stress-test one bounded result and document deviations.
40%CapstoneStudy, evidence release, oral defense, and evidence-led revision.
Common rubric: evidence integrity 25%, methods 20%, reproducibility 20%, inference and uncertainty 15%, safety/governance 10%, communication/accessibility 10%. Fabricated evidence, unauthorized collection, or a broken primary claim path blocks release regardless of average score.
Calendar
Sixteen essential questions
W01What GEO is—and is notWhat is the actual object of optimization?W02Evidence before expressionWhich source can support which claim?W03Discovery, indexing, and retrievalCan the system obtain the source at all?W04Ranking and context selectionWhy does an eligible source enter the context?W05Generation, citation, and absorptionWas a source merely displayed, or did it shape the answer?W06Queries as a measurement instrumentWhat distribution do the prompts represent?W07Metrics, uncertainty, and repeated observationWhen is a visibility change repeatable?W08From observation to identificationCan the change be attributed to the intervention?W09Evidence-rich content architectureHow can structure improve use without distorting facts?W10Entities, authority, and source ecosystemsHow does a fact become corroborated beyond one owned page?W11Multilingual, multimodal, and agentic surfacesWhat changes when the source is an image or the system can act?W12Optimization under competing queriesWhich gain survives a portfolio of intents?W13Field experiments and monitoringHow do we learn under a changing platform?W14Manipulation, poisoning, and defenseWhen does optimization become system abuse?W15Fair attribution and mechanism designCan visibility incentives reward useful evidence?W16Capstone synthesis and defenseWhat would make the result worth trusting?
Revision and integrity
Evidence-led revision is required
Submission states
Draft → validated artifact → review finding → revision → release. The final file must preserve deviations, negative results, version, and checksum. Late technical failure is not hidden; it becomes a documented incident and recovery decision.
AI-use disclosure
Declare model/tool, version or date, task, retained output, human verification, and material effect. Generated prose, labels, code, or sources are not evidence until independently checked.
Safety boundary
No live poisoning, evasion, unauthorized load, credential use, deceptive attribution, personal-query collection, or fabricated authority. Use the supplied synthetic route whenever authorization is uncertain.
Accessibility
Core tasks cannot depend on pointer, color, audio, proprietary accounts, or high-compute models. Visuals require captions and text equivalents; recorded material requires transcripts.