Slide deck failed to load or the page is empty

Reopen this file in Chrome or Edge, or confirm the file downloaded completely.

Workshop AI Overview AI Agents

AI Search & Digital Content Trends Workshop

New Business School | January 29, 2026

Key pointMake the signals observable and the actions assignable.

Today’s flow: practice first, then work backward to mechanisms and actions

Click to jump to a slide (the agenda slide is scrollable).

Controls: use ←/→ to navigate; on mobile, use the bottom navigation bar.

Five minutes: two concrete deliverables for today

One external (search → transaction) and one internal (AI adoption → knowledge governance).

Track 1 | AI Overview → eCommerce / transactions
  • Use the exercises to verify: how entry points and placements shift, how sources are chosen, and how clicks get compressed.
  • Write actions as: content strategy (what to write) + distribution strategy (where to say it) + measurement plan (what to watch).
  • DeliverableA “customer journey × content assets” action checklist (with metrics).
Track 2 | Internal AI adoption / resource integration
  • Use “task fragments” to decide: which parts to augment vs. which parts to automate.
  • Put AI back into the coordination system: definitions, ownership, review checkpoints, and versioning.
  • DeliverableA draft of “red lines + review checkpoints,” one reusable Skill outline, and one JD draft.

Exercise 1: run the same query 5–8 times to see how results drift

Use incognito throughout to avoid personalization (same keywords / same wording).

Steps (3–5 minutes)
  1. Open an incognito window in Chrome or Edge.
  2. Pick one query: use a full sentence with constraints, e.g., “I want…, but I care about… How should I choose?”
  3. Repeat the search 5–8 times (retype or refresh each time).
  4. Take a screenshot each time (capture the full SERP).
  5. On the next slide, log each run in the table (aim for consistent notes, not polished writing).

If AI Overview never appears, make the query more complex (add context, constraints, comparison dimensions, risk).

You’re not judging “is the answer good”; you’re observing three things
  • TriggerWhether AI Overview appears, where it appears, and how much real estate it takes.
  • DecompositionWhat sub-questions the summary breaks your query into (did you even ask those?).
  • CitationsWho it cites, how it cites, and whether that makes you trust it more—or less.

Exercise 1: Observation log (one row per search)

These columns are exactly the AI Overview points we’ll discuss next.

Keep the query fixed (use the exact same sentence for the whole exercise).
Run # AI Overview Sub-questions / What the summary says Cited sources (list 2–3) Your next step Screenshot filename
1
2
3
4
5
6
7
8
Three one-sentence observations (to carry into the next section)

Exercise 1 quick debrief: what kind of “instability” did you see?

Name the phenomenon precisely first; actions come later.

Discussion prompts (1 minute each)
  • In your context, what does “AI Overview appears/doesn’t appear” mean?
  • What sub-questions did it break the query into? Did it pull you toward directions you hadn’t considered?
  • Do the cited sources look like whose world: official sites, media, forums, e-commerce, encyclopedias?
  • When would you not click? When would you make a precise click?
  • From a brand perspective: what must we provide to be cited, trusted, and adopted?
Your table’s consensus (write 5 sentences)

Exercise 2: look at Bing’s AI Overview and compare it with Google

Google dominates in Taiwan; Bing helps you extract reusable optimization logic through comparison.

Steps (2–3 minutes)
  1. In an incognito window, open Bing.
  2. Search with the same query.
  3. Screenshot and jot down differences (use the table below).

The goal isn’t “which is better,” but understanding how similar AI summaries decompose intent and select citations.

Comparison notes (differences only; no long prose)
Dimension Google (what you saw) Bing (what you saw)
Trigger & placement
Citations & verifiability
UI & next step
One-sentence takeaway

Working backward from the exercises: AI Overview is doing three things

You don’t need to guess the algorithm, but you do need to know what it’s rewriting.

1 | Decompose intent
  • Split one query into multiple sub-questions (you’ll see different decompositions).
  • When the decomposition changes, what users click next changes.
2 | Recompose the answer
  • Reassemble multiple sources into an “answer layer.”
  • It looks like a conclusion, but it’s a mix of citations and inferences.
3 | Allocate attention
  • When the summary takes space, clicks naturally drop—but the people who need it will still click precisely.
  • Your challenge shifts from “getting clicks” to “being cited, trusted, and selected.”
Which step is most likely to be rewritten for my unit?

AI turns search into an “answer layer,” but the answer isn’t always reliable

For organizations, the question isn’t whether AI can be wrong—it’s who carries accountability when it is.

Observed behaviorUsers increasingly adopt the summary directly instead of clicking through to websites.

RiskInaccurate or wrong information can still shape decisions and brand perception.

TaskBuild “reliability” into policy and data structures—not luck.

Next, we’ll use official docs to clarify how AI Overviews work, then return to content and actions.

Google AI Overviews are “Search + Gemini”—not a chatbot

Official framing: it works with existing ranking/quality systems and the Knowledge Graph, and provides clickable links for verification.

How Google describes AI Overviews (summary)
  • Uses generative AI to provide a “key information summary,” with links for deeper reading.
  • Uses a customized Gemini model working in tandem with existing Search systems (ranking/quality systems, Knowledge Graph).
  • Designed for “more complex” questions, reducing the cost of multiple searches.
  • Triggered only when the system believes gen AI is “particularly helpful” and has high confidence in quality.
  • Google also notes AI answers can be wrong and encourages users to verify via links and provide feedback.
From the website/brand side: what can you control?
  • No extra “special optimization”It follows existing SEO fundamentals and technical requirements.
  • May use query fan-outA single question can be expanded into multiple related searches to find supporting sources.
  • Preview controlsUse noindex / nosnippet / max-snippet, etc., to control content display.

Docs: PDF ｜ Search Central ｜ Search Help

Part 1: Reliability beats cleverness

Once AI enters the workflow, the hard part is making it structurally hard to be wrong.

Reliable AI isn’t better at answering—it’s constrained by four layers of structure

Move the problem from “prompt tricks” to repeatability, reviewability, and handoff.

Layer 1: Data sources
  • Data needs versions, dates, and owners.
  • Traceable citations (where it came from and when it was updated).
Layer 2: Rules and guardrails
  • Red lines: what must not be said or done.
  • Exception handling: how to fall back when unsure.
Layer 3: Workflow embedding
  • Put AI into delivery workflows (with state and rollback).
  • Review checkpoints and responsibility chains must be explicit.
Layer 4: Governance and portable trust
  • Traceable, auditable, and transferable.
  • Shared capabilities reduce duplicate investment.

AI shifts the starting point of content creation: back to the customer journey, not keywords

Search is decomposing intent; our job is to capture context, motivation, emotions, and wording.

Three shifts (you’ll see them in the exercises)
  • Voice-drivenUsers increasingly use voice input; queries get longer and more like full requirements.
  • ConversationalUsers get used to follow-ups instead of stopping at one-off answers.
  • Citation-drivenThe system lists references, but users may not click—more like outsourcing verification to the machine.

So the focus isn’t just SEO—it’s whether you provide content units that can be cited, verified, and understood.

Write out a persona’s “research → compare → decide” journey

In a world of summaries and recommendations, brands need four “citable” assets

Not to publish more articles, but to make verification and comparison easy—for both AI and humans.

1 | Verifiable rules
  • Returns, warranties, responsibility boundaries
  • Dates, versions, change logs
2 | Comparable specs
  • Turn specs into tables with explicit fields
  • Make side-by-side comparison easy
3 | Citable evidence
  • Tests, comparisons, applicable contexts
  • Don’t just write slogans—write the basis
4 | Follow-up-ready interaction
  • FAQs, clarification pages, version pages
  • Make it possible to correct errors

The point isn’t “fewer clicks”—it’s precise clicks and deeper engagement

Click-through may drop first; shift to metrics you can actually live with.

Mindset shift
  • Before: get people in (broad clicks).
  • Now: when they’re solving a problem, provide information they can actually use (precise clicks).
  • Expectation: clicks may decrease, but transactions or deep engagement may not fall proportionally.
Three metrics my unit should start tracking

Distribution needs three shifts: focus, formats, and cadence

Keep expressing your point of view where it can be indexed—and ground it in how people actually share.

  • Shift focusFrom “article volume” to “citable assets” and “verifiable pages.”
  • Shift formatsYour owned site must be citable; external platforms (social/media/e-commerce) need consistent messaging.
  • Shift cadencePublish consistently where your audience gathers so your viewpoint accumulates into a recognizable context.

Reminder: on platforms that are hard to index, plan how you’ll be discovered and how you’ll route back to an official, verifiable page.

One distribution change I’ll make in the next 30 days

Part 2: AI agents will reshape the shopping flow—and who gets paid

Both are called “agents,” but Google and OpenAI monetize differently.

The key difference: the agent is a mediator vs. an actor

One sells influence; the other sells task completion and transactions.

Google: Mediator
  • Understand intent, organize options, influence decisions
  • Send you back to merchant or platform product pages
  • Monetization: turn “influence” into ads
OpenAI: Actor
  • Compress search, compare, and buy into a single flow
  • Complete tasks and transactions inside the conversation
  • Monetization: charge when value is delivered

The blind spot in AI shopping: efficiency ≠ desire to buy

Shopping isn’t just “finding the optimal answer”—it’s a process where desire gets constructed.

An evolutionary view: browsing isn’t wasted time—it’s how desire forms

Window-shopping, comparing, and adding to cart all extend anticipation and imagination.

Desire is made
  • That distance—visible but untouchable—creates craving
  • Every tab and every price check adds reasons to convince yourself
  • A product page is a “transformation promise”: if I buy it, I’ll become a better me
Two common friction points
  • Latent questionsMany questions only occur to you after you see the options
  • EnjoymentChat turns browsing into one Q&A at a time, losing side-by-side comparison and serendipitous discovery
  • ConsistencyIf recommendations are unstable or hallucinated, distrust goes up

Conclusion: when AI optimizes only for efficiency, it can strip out the fun of browsing.

Inference: AI-driven “traffic” may come before “conversion”

Multi-touch buying, habit and trust, and UI constraints can make the conversion gap the norm.

Multi-touchPurchase journeys often span 6–8 touchpoints; chat is hard to count as “the last mile.”

Conversion gapEven if both drive traffic, chat-driven traffic doesn’t necessarily convert better than traditional channels.

PrecedentVoice shopping (Alexa) was hyped, but long-run share remains very small.

Different roles shouldn’t ask “Are we using AI?”—but “Who is rewriting us?”

Even within the same group, entry-point dependence and reshuffling can look completely different.

Role / function Typical impact What you need to answer today
Marketing / operations Zero-click, shrinking placements, rewritten conversion paths Entry-point dependence × KPI redefinition
Content / media Summaries replace clicks; traffic and ad pressure rise A “citable” strategy and verification checkpoints
Commerce / e-commerce Agents compress comparison and checkout; but browsing / comparing / looking is still part of desire formation Make comparison visual and trust verifiable
Product / service Smaller teams + AI create substitute competition Differentiation: delivery, trust, data assets
Engineering / data Duplicate investments, inconsistent specs, governance debt Shared capabilities: platform, data, risk controls
Back office / audit Weak accountability chains and traceability amplify risk Review authority × traceability × adjustability

Concept exercise: how far will AI-guided shopping go?

Deliberately judge from consumer intent—not from our own interests.

My forecast (single choice)

Discussion: why didn’t Alexa shopping go mainstream?

Use this counterexample to calibrate our mental model of “AI-guided shopping.”

Answer these two questions first (3 bullets each)
  • Why didn’t voice shopping become the default? (trust, comparison, context, UI…)
  • Why do humans still want to compare across options? (it’s not only about the cheapest price)

The point isn’t Alexa itself; it’s which purchase behaviors are fundamentally unsuitable for compression.

My notes

Synthesis: who will love AI-assisted shopping—and who will doubt it?

Abstract the disagreement into “our group’s point of view”: who we serve and who we don’t.

Our view (one sentence)

Breakout: build a “content × interaction × recommendation” loop for your business unit

The point isn’t to create lots of new stuff; it’s to reshape how we understand and connect consumer intent.

Loop canvas: write the three environments you can influence separately, then connect them

Owned touchpoints / hard-to-control external environment / partner sites (e-commerce, media, etc.)

1 | Owned touchpoints (site/app/stores/support): what do we do?
2 | External environment (search/social/conversational AI): what actions do we take?
3 | Partner sites (e-commerce/media/channels): what information do we disclose?
Write the loop in one sentence (from entry point to conversion/engagement)
Metrics and owner

Estimate: what’s the biggest difficulty—and how much time will it take?

Don’t chase precision; chase assignability and schedulability.

Blockers and risks
Time and resources

Case discussion (five tracks)

Same language, but pick the context closest to your work.

Pick a track: focus the discussion on the next step you can control

Each track answers the same three questions: direction, rationale, metric.

AContent/media: after summaries replace clicks, how should KPIs change?

BCommerce/e-commerce: if agents compress search–compare–buy, which capability do you need to build?

CDelivery/enterprise services: when smaller teams can deliver too, what differentiation do you defend?

DGovernance/back office: once usage is allowed, how do you control governance debt?

EBrand/PR: how do you design verification checkpoints for AI misattribution and wrong citations?

OrganizeSplit actions into three buckets: owned touchpoints (site/app/stores), external environment (search/social/media), and partner sites (e-commerce/platforms/media networks).

Track A: when summaries replace clicks, content value shifts from “traffic” to “citable”

Scenario: AI summaries siphon clicks; ads and referrals get squeezed.

DecisionWhat would you do first? (single choice)

  • A. Build an “authoritative verification page” (citable, traceable)
  • B. Shift to social and subscriptions to reduce reliance on search
  • C. Strengthen events/livestreams and turn content into an interactive service
  • D. Rebuild measurement first to identify the metrics that truly drive revenue
Your choice and why

Track B: when agents compress search–compare–buy, you must solve two things at once: trust and browseability

Scenario: AI wants to speed up the flow, but people still need comparison, serendipity, and self-justification.

DecisionWhich capability would you build first? (single choice)

  • A. Make product/deal/inventory data machine-readable to reduce errors
  • B. Make side-by-side comparison visual (so latent questions surface)
  • C. Turn customer support and returns into verifiable rules so people dare to order
  • D. Put “reasons and evidence” into content (tests/comparisons/applicable contexts)
Your choice and why

Track C: when smaller teams can deliver, differentiation moves toward trust and accountability chains

Scenario: competitors use AI to compress headcount—what about your delivery is still non-substitutable?

DecisionWhich differentiation would you defend first? (single choice)

  • A. SLA / delivery quality (codify standards into auditable processes)
  • B. Security / compliance (institutionalize review authority and responsibility boundaries)
  • C. Expert judgment (make the reasoning teachable and transferable)
  • D. Customer relationship (make trust a reason for renewals and repeat purchase)
Your choice and why

Track D/E: governance debt and wrong citations are both traceability problems

Once usage is allowed, what blows up first is often workflow and accountability boundaries—not the model.

D: Governance / back office

ScenarioEach unit buys tools and uses its own accounts; data flow is unclear.

QuestionWhich “minimum governance boundary” do you set first?

My minimum boundary (one sentence)
E: Brand / PR

ScenarioAn AI summary misattributes another company’s incident to you.

QuestionWhich “verification checkpoint” do you set first so errors can be corrected?

My verification checkpoint (one sentence)

Part 3: AI will decompose work—starting with tasks and coordination

External entry points get reshuffled; internally, workflows, staffing, and accountability chains get redesigned.

Reshuffle: AI turns intent into a new interface, reshuffling entry points

When the interface changes, your existing “traffic, power, and collaboration” change with it.

In one sentence

AI isn’t just answering questions; it turns “what you want to do” into an interface and workflow you can operate.

That’s why even though it’s all called AI, some use cases are “augmentation,” while others directly “replace” parts of the job.

Which step is being rewritten for me? (one sentence)

You think you’re rolling out AI, but you’re really redefining coordination

Who provides what data, which definitions apply, and who owns the final decision—that’s the core of the organization.

  • SurfaceIt looks like “writing faster” and “doing things faster.”
  • RealityWhat’s truly being rebuilt is coordination: data, definitions, ownership, reviews, handoffs.
  • SystemAI amplifies chaos; when workflows aren’t clear, errors are harder to trace.
What my unit lacks most isn’t tools—it’s which kind of coordination?

Break coordination into five levers: start knowledge management at the representation layer

You don’t need to memorize technical jargon, but you do need to manage shared language, definitions, and ownership.

Coordination lever What you need to manage How it maps to AI rollout
Representation Shared language, definitions, citable data Knowledge base / RAG dataset / glossary / version history
Decision Who makes the final call and by what criteria Review checkpoints (HITL) / permissions / audit fields
Execution Turn decisions into deliverable steps Workflows / automation / exception handling
Integration How systems pass data and write back APIs / field alignment / trigger conditions
Rules Red lines, responsibility boundaries, compliance Policies, traceability, risk-incident workflow

Break work into four fragments to see what AI will affect

Leaders shouldn’t guess job titles; identify automatable segments and the human review checkpoints you must keep.

Work fragment Typical contents How AI intervenes Common risks
Search / synthesis Find sources, aggregate, summarize Speed up and standardize outputs Missing key items, unclear citations
Transform / format Convert data into tables, fields, JSON Structured input/output Field hallucinations, format drift
Criteria / decision Make judgments by policy and definitions Generate candidates first, then human adjudication Outsourced accountability, amplified bias
Coordination / execution push Cross-team alignment, follow-ups, exception handling Reminders, tracking, generating deliverables Looks smooth but nobody owns it
Pick three tasks in your unit and tag them as “augment” or “automate”
# Task Augment / automate Review checkpoint Risk
1
2
3

Task level: distinguish augmentation from automation

They’re fundamentally different: one helps you think better; the other runs the workflow for you.

Augmentation
  • Help ideation, improve thinking, expand options
  • Humans still own judgment and final decisions
  • Best for: high-risk work, context-dependent work, accountable narratives
Automation
  • Hand a sequence of workflow steps to AI (less human involvement)
  • Key: review checkpoints, rollback paths, accountability chain
  • Best for: high-frequency, low-risk, well-specified, data-available segments

A systems view: when a competitor can challenge you with a small team plus a few AI “employees”

Use a red/blue team thought experiment to make threats and countermeasures explicit.

Red/Blue Team Canvas: make an “AI-driven operating model” concrete

Red team imagines a new entrant; blue team responds with existing assets and systems.

Red team: with only 5–10 people + AI, how would they offer a similar service?
Blue team: how do you defend / counter?

Knowledge system: turn reusable knowledge into Skills to scale

A Skill isn’t a prompt; it’s a module with inputs/outputs, red lines, and an owner.

VersionedWho maintains it, when it’s updated, and the change log.

AuditableClear input/output specs; build HITL checkpoints into the workflow.

TransferableNot one person’s trick—teamwide delivery with a shared standard.

Sketch a Skill your unit would need

Demo: two Skills (tone editing, digital slide deck)

The point isn’t to show off; it’s to show how knowledge gets crystallized into a workflow.

Download: Markdown files for the two Skills (OneDrive)
Skill 1 | Tone editing
  • Input: source text + target tone + prohibitions
  • Output: review-ready / publish-ready version (includes sensitive-term checks)
  • Review checkpoint: final pass by brand/compliance
Skill 2 | Build a digital slide deck
  • Input: outline, assets, layout spec
  • Output: projection-ready / shareable file (with consistent formatting)
  • Review checkpoint: key numbers and cited sources

Group exercise: write a job description for an AI-related role

After AI goes live, where should humans sit to handle risk review, version governance, and exceptions?

Job description template (clarity first, perfection later)

There’s only one closing bar: deliver one verifiable result within 30 days

Don’t write a vision—write “owner + metric + next step.”

30/60/90 action card

What you take today isn’t a slogan—it’s four lists you can keep tracking

Turn discussion into deliverables, or you can’t follow up later.

Four lists
  • List 1Exercise observations: how entry points and citations shift (SERP log)
  • List 2Loop actions: what to do across owned/external/partners (with KPIs)
  • List 3Internal governance: task breakdown, review points, Skill sketch, JD draft
  • List 430/60/90 action card (assignable and trackable)

Principle: build traceability first, then expand automation.

Next steps (assignable)
  • One 30/60/90 action card per unit
  • List 1–2 shared-asset needs (data / governance / tools / metrics)
  • Schedule a review checkpoint 30 days out (KPI-based review)

Review and export notes

This handout is based on workshop discussions and provided materials.

Review
  • Treat search as an “answer layer”: the battlefield shifts from “getting clicks” to “being cited / recommended.”
  • Reliability beats cleverness: use data, rules, citations, and human review to keep risk within a controllable range.
  • AI agents reshape shopping flows and monetization points: first clarify whether the agent is a mediator or an actor, then decide which entry points and transaction checkpoints you defend.
  • Browsing / comparing / looking is part of desire formation: don’t optimize only for faster checkout; design a comparison experience that’s both enjoyable to browse and trustworthy.
  • Work breaks into tasks + coordination: decompose coordination first, then decide what to augment vs. automate.
  • Turn reusable knowledge into Skills: fix inputs/outputs/review checkpoints so individual tricks become team capacity.
  • Governance debt and wrong citations are traceability problems: build versions, citations, and accountability chains first, then scale automation.
  • The close is about one thing: deliver one verifiable result in 30 days, and write your 30/60/90 next steps so they’re assignable and trackable.
Export personal notes

After exporting, you can save it or email it to yourself.

Survey feedback

QR Code

https://bnext.surveycake.biz/s/QKM1v

Instructor introduction

InstructorHsiang-Hsun Wu | Department of Lifelong Education, Yuan Ze University

Website537SONIC