AI Agents Hijack Shopping, Crush SEO, Grab Ad Billions

● AI Shopping Agents Crush SEO, Seize Ad Billions

Where will AI dig into shopping to make money? The new order of e-commerce that consumer AI agents will overturn

This piece covers the capital shift from B2B to B2C, the reshaping of the ad market via “skip search,” five revenue models for shopping agents, the battle over the personalization data layer, and actionable investment strategies that domestic players can execute immediately.
In particular, it translates details that most news or YouTube content rarely address—payments and settlement rails, SKU graph standards, Agent Optimization (AO) and measurement frameworks, and refund/liability issues—into a realistic roadmap.
It lays out, chronologically, the moment when the global economy’s demand, pricing, and marketing dynamics are rewired by AI, so you can apply it to your business and experiments right away.

2023–2025 timeline: Capital and user base shift from B2B to consumer AI

2023 was the year B2B led productivity breakthroughs with generative AI, and digital transformation accelerated rapidly across enterprise workflows like advertising, documents, and coding.
From 2024, a critical mass emerged. As more general users paid monthly subscriptions for ChatGPT, Claude, and Perplexity, willingness to pay in consumer segments was validated in practice.
2025 is Act 1 of consumer AI in shopping, gaming, and jobs. These three are first to monetize because e-commerce sales, user payments, and conversion fees are already large; when agents enter, the impact shows up on the P&L immediately.

Five ways AI agents make money in shopping

AI agents penetrate with different revenue models depending on the difficulty and risk of the purchase decision.

1) Impulse purchases: Instant checkout inside content

Contextual suggestions appear in Reels, Shorts, and social feeds, and one-click checkout completes the flow.
Revenue comes from in-read affiliate fees, instant checkout commissions, and creator rebates.
The key is “discovery without search,” as the agent understands video/text in real time and attaches SKUs along with inventory and delivery ETA.

2) Everyday consumables: Reorder automation with fickleness control

For items like shampoo, detergent, and wipes—low brand loyalty, high frequency—apply “consumption-rate estimation → reminders → variable subscriptions.”
Instead of full automation, include a default “change-of-mind” button to make returns and brand switching easy, reducing churn anxiety.
Revenue comes from subscription fees, manufacturer rebates, and increased purchase frequency.

3) Lifestyle, fashion, travel: Conversational curation

In taste-driven categories with no single right answer, the agent learns preferences through dialogue and bundles capsule recommendations to raise AOV.
Travel incorporates companions, budget, schedule, weather, and crowd levels to generate itineraries and complete bookings, maximizing profits via bundle margins.
This converts “browsing-based purchases” into “companion-style consultative purchases,” lifting LTV substantially.

4) Negotiated purchases: Furniture, interiors, insurance

The agent handles tedious steps such as spec comparisons, photo collection, vendor communications, and quote negotiations.
Initially it runs Human-in-the-Loop, with the user approving at forks in the road, and monetizes by taking a share of the savings from quote comparisons and term negotiations.
Dialogue history and normalized contract line items become core data assets.

5) High-involvement, big-ticket purchases: Cars, homes, study abroad (+financing)

For high-risk decisions, trustworthy sources, guarantees, and after-the-fact responsibility matter.
The agent provides a full-stack pipeline—spec selection → test-drive scheduling → used depreciation and insurance simulations → loan pre-approval—and monetizes through financial fees and referrals.
Agents that earn trust here become the user’s main concierge.

Platform landscape: “Skipping search” and the reshaping of ads and SEO

Users say, “Book a 4-day Tokyo itinerary with this budget,” and the agent goes straight to cart and checkout.
At that moment, two steps disappear: Google search and product detail page browsing. Revenues dependent on search ads and affiliate traffic take a direct hit.
By market estimates, hundreds of billions of dollars in shopping ad revenue across global search and marketplaces will shift to AO (Agent Optimization).
AO is the sum of technology and operations that optimize the feeds, specs, policies, inventory, shipping, and reputation trust scores that agents read to raise agent ranking.
In short, the center of gravity in optimization shifts from SEO/ASO to AO, and performance measurement moves from “pixel-based” to “agent-reported.”

Data and tech infrastructure: The reality of the personalization layer war

Personalization hinges on three things: 1) consented data collection, 2) context fusion, and 3) long-term memory.
Key components are the following four:

Persona graph: Structure preferences, body type, allergies, and price sensitivity into long-term memory.
Consent wallet: Let users control which apps receive which fields and for how long.
Agent-Ready SKU feed: Provide images, dimensions, inventory, returns, A/S, and sustainability in a standard schema.
Trust score: Quantify product and seller trust via source credibility, review tamper detection, return rate, and CS response speed.

Payments and settlement rails: Cartless checkout and liability

Agents pay with virtual cards “without a cart,” and brands/marketplaces should separate Seller-of-Record liability and refund policies into agent-specific terms.
A new issue arises when hallucinations cause mispurchases: who is responsible? Add “user confirmation checkpoints” and insurance-like protection plans to agent authorization flows.
In Korea, open instant payment, BNPL, and real-time transfers via agent APIs to boost conversion immediately.

Unit economics: Inference cost vs. conversion lift

Cut the agent’s LLM inference cost in three ways: 1) on-device small models for pre-filtering, 2) cache and template reuse, 3) route fixed tasks via tool calls.
Raise conversion via conversational capsule recommendations, bundling, and trust signals. The meaningful benchmarks are CAC/payback improvement and the magnitude of return-rate reduction.
In fashion and travel, reducing returns/itinerary changes drives P&L leverage. In everyday consumables, stabilizing the reorder cadence is key.

Regulation and trust: Privacy and protection of minors

A personalization wallet must provide purpose limitation, retention controls, and a withdrawal button.
For minors, strengthen payments with limits, category blocks, and guardian approvals.
Clearly label ads and recommendations as “Sponsored/paid partnership,” and retain explainability logs for at least six months.

Execution checklist by domestic player type

Retail/brand direct purchasing

0–6 months: Agent-Ready product feeds, returns/A/S schema cleanup, and customer-consent UI rollout.
6–18 months: Conversational shopping-bot beta, agent-only coupons and bundles, and operate a dedicated AO cell.
18–36 months: Integrate personalization wallet and make in-store recommendations real-time via on-device models.
Marketplace/super app
Open catalog and payment APIs for external agents, establish revenue-sharing policies, and run an agent certification program.
Payments/fintech
Virtual card/tokenization and real-time limit-adjustment APIs, agent liability insurance bundles, and BNPL agent suitability underwriting models.
Logistics/3PL
ETA reliability score API, automated returns triage and resale pipelines, and packaged SLAs for agents.

Measurement and experiment design: KPIs for the AO era

End-to-end conversion rate, return rate, NPS/trust, number of conversation turns to first purchase, CAC/payback, and agent channel mix share.
Use multi-armed bandits, not just A/B, by changing policies, feeds, and explainability guidelines as bundles and optimizing holistically.

Risks and pitfalls

Excessive automation erodes user sense of control. Always show “change, cancel, personalize” buttons.
Over-collection of data. Specify consent context and benefits and adhere to data minimization.
Model overfitting and trend bias. If the agent appears to push a single brand, trust collapses.

Why now: Macro and value chain perspective

In a global economy facing high rates and low growth, the technology that cuts marketing waste and raises real purchase conversion wins.
AI agents bypass the upstream of “search traffic” and settle downstream at “right before checkout,” redefining profit capture zones.
Companies building AO capabilities now will capture the winner’s surplus in the next e-commerce cycle.

Execution roadmap: 0–6–18–36 months

0–6 months: Form an AO task force, standardize SKU feeds, run pilots in chatbots/travel/fashion, design KPIs, and renew privacy consent flows.
6–18 months: Open external agent channels, package payment/logistics APIs, revamp returns/liability policies, and experiment with on-device models.
18–36 months: Commercialize the personalization wallet, launch full-stack financing bundles, connect insurance/guarantees, and integrate AO auto-bidding with dynamic pricing.

Key takeaways

Where money flows first: impulse buys, consumable reorders, fashion/travel curation, negotiated purchases, and high-involvement + financing.
Power shifts: SEO/search ads → AO/agent feeds, pixel measurement → agent-reported measurement.
The contest is won in the personalization data layer and trust. Consent wallets, explainability, and returns/liability design make the practical difference.
Now is the time to lock in “standards and metrics” before pilots. That’s how unit economics hold when you scale.
Korea’s strengths are payments and logistics infrastructure. Open them via APIs to lift both conversion and customer experience.

< Summary >

Act 1 of consumer AI is shopping, gaming, and jobs, and agent monetization in shopping splits into five models.
As “skip search” takes hold, AO replaces SEO and search ads as the core capability.
The personalization data layer, consent wallet, Agent-Ready product feed, and trust score are the battlegrounds.
Rewrite payments, settlement, returns, and liability for the agent era.
Companies that follow a 0–6–18–36 month roadmap—standardize → pilot → channelize → scale—will win.

*Source: [ 티타임즈TV ]

– AI는 쇼핑의 어디를 파고 들어 돈을 벌까? (허진호 한리버파트너스 제네럴파트너)

● Active AI Takeover – Agents, Robots, Vibes Rewire the Economy

OpenAI Pulse, Google DeepMind Robotics, Meta Vibes: The Economics of AI’s Shift from Passive to Active and a 12-Month Execution Strategy

The core of my piece has three parts.
First, I unpack how unit economics for individuals and organizations reshuffle when Pulse turns into a nightly “background agent,” with a feel for the numbers.
Second, I outline scenarios for how DeepMind’s multi-robot transfer learning lifts productivity in real factories, logistics, and services, and how that ripples through global outlooks, inflation, and interest rates.
Third, I show how Meta’s AI video feed changes the distribution structure of ads and the creator economy, and the allocation strategy my brand should take.
If you read to the end, you can apply an execution checklist right away for what my team should change this week.

[0~3 months] OpenAI Pulse: The reality of an asynchronous agent that thinks “while you sleep”

Pulse runs asynchronous research overnight based on your chat context and feedback, plus apps you authorize like Gmail and Google Calendar.
In the morning, it delivers card-style summaries, then lets you curate “what I want to see tomorrow,” gradually personalizing.
It will start in preview with mobile Pro users, then expand to Plus and general availability.
Privacy is stated to be used only for improving my results and not for global model training.
Crucially, it’s not a feed to stretch consumption time but a philosophy of “updates that end.”
The key change is that intent inference and proactive suggestions become the default instead of input.

An underrated point others miss 1: the unit economics and performance limits of background reasoning.
Generating nightly personal summaries accumulates inference (LLM + tool calls) costs.
From the company’s perspective, it’s inevitable to enforce an “agent budget” policy that caps summary length, tool call counts, and refresh frequency.
In other words, card count, depth, and update cadence may appear user-configurable, but the system likely tunes them to a cost ceiling.
Invisible to users, “agent runtime” and “round trips” will determine experience quality.

Point 2: how the privacy–utility balance actually runs.
To raise the quality of suggestions, Pulse must combine and normalize personal data.
In practice, layered guardrails will likely include account-level encryption, app-level scopes, sensitivity tagging, and retention-period policies.
For now the principle is “my data only improves my results,” but when team/family plans roll out, the scope of sharing and accountability will become hot issues.

Point 3: usage quality rises when users grasp the product design KPIs.
Beyond like/dislike, actively adjust switches like “shorter/longer tomorrow,” “separate work/personal,” and “allowed scope for tool calls.”
Ultimately, with Pulse, the user is the designer.

[3~12 months] Acceleration to agents: from “instruct-then-execute” to “goal–continuous attainment”

If ChatGPT agents earlier this year showed instruction-based execution, Pulse steps further as a persistent aide that tracks goal states.
The next stage is “autonomous actions within an allowed scope.”
For example, for a travel-prep goal, it gathers lodging options, proposes candidate schedules that avoid calendar conflicts, and connects to booking links upon approval.
Competitors are moving the same way.
Google is building a Gemini agent stack, and Apple is reinforcing data boundaries with on-device personalization (Apple Intelligence).
From an enterprise perspective, the next wave of “digital transformation” arrives not in chat windows but as “autonomous execution pipelines.”

Business insights:

Elevate SOPs from prompts to “approval workflows.”
Define the agent’s allowed data, tools, and action scope as policy and keep change logs to make compliance work.
Redefine agent SLAs by outcome suitability, error costs, and human approval rates rather than response time.

[Same time frame] Google DeepMind: Industrial shock from multi-step planning and cross-robot transfer learning

Two update threads.
Embodied Reasoner (ER) 1.5 understands environments and, when needed, calls digital tools like web search to build language-based plans.
Gemini Robotics 1.5 executes those plans.
In short, one plans and the other performs.
Another leap is transfer learning.
Skills learned on Aloha 2 were transplanted to entirely different robots like Franka and Aptronic Apollo.
This reduces the waste of “retraining from scratch” per robot and turns training capital into a shareable asset.

Economic impacts:

Productivity: multi-step task execution and transfer learning lift throughput per hour from the early stages of deployment.
Global outlook: faster automation in manufacturing, logistics, and retail should lower unit costs and reduce lead-time volatility in the medium term.
Inflation: supply-side efficiency is disinflationary structurally, but initial capex for robots and AI servers drives a surge in capital-goods demand.
Interest rates: the stronger the investment cycle, the more central banks pace themselves watching the growth–inflation balance.
Employment: hazardous/repetitive work automates quickly; frontline roles shift to “robot orchestration” and quality control. Reskilling demand spikes.

Practical tips:

Start PoCs with tasks that are multi-step, rule-referenced, and checklist-based.
Draw a roadmap to reuse the same skills across different robot types to accelerate ROI.
For quality metrics, go beyond success rate; separate recoverable errors and the cost of recovery.

[Same time frame] Meta Vibes: Redistribution in ads and the creator economy opened by an AI video feed

Vibes is a dedicated feed for AI-generated and remixed videos inside the Meta AI app.
You can watch, then immediately change styles, replace elements, add music, and distribute directly to Vibes, DM, Instagram, and Facebook.
The key is binding the social graph and distribution engine to AI-native creation tools.
This elevates the “viewer” into a “creator,” driving creation costs toward zero.

Business insights:

In ads and the creator economy, differentiation shifts from “scarcity of creation” to “distribution, trust, and identity.”
Brands should publish templates, presets, and IP guidelines to encourage legal remixing by fans, and run a reference system that tracks the original–derivative chain.
For measurement, “remix rate based on participation” and “safety flag rate” become more useful KPIs than views.

[Cross-analysis] The economics of active AI: cost, governance, trust

The biggest cost variables are always-on inference and the frequency of tool calls.
Enterprises should manage costs via user/team/process-level “agent budgets,” nightly batches, on-device inference, caching, and summary compression.
For governance, systems are more stable when you codify “what must not be done” rather than “what can be done.”
Trust comes from transparency.
Whether Pulse or Vibes, showing the provenance path of “why this recommendation/generation appeared” reduces pushback and speeds adoption.

[Macro scenarios 2025~2026] A productivity–inflation–rates trio

Short term (6~12 months): investment and exports remain solid on cloud, semis, and robotics capex, while service prices face upward pressure.
Medium term (12~24 months): process automation kicks in, raising output per labor hour and improving corporate margins.
Productivity gains lower inflation pressure, but data, safety, and compliance costs are offsetting factors.
Globally, the U.S. and Europe adopt agents and robotics faster, while Asian manufacturing hubs use transfer learning to lower adoption costs and catch up.
Rates likely adjust gradually along the growth–inflation balance, with the path varying by the pace of agent diffusion.
As digital transformation turns into “autonomous execution transformation,” AI CapEx becomes a key variable of the new business cycle.

[This week’s execution checklist] Team · Individual · Investor

Team lead: pick three SOPs, convert them into “Pulse-style” morning briefings, and record approve/deny logs to improve rules next week.
Data owner: document connection scope for Gmail, Calendar, Drive, retention periods, and sensitive-data filtering rules, and obtain renewed consent from members.
Marketing: create five remixable templates for Vibes and brand safety guidelines, and set a target for UGC remix rate.
Operations: select two multi-step tasks for a robot–human hybrid PoC, and standardize failure modes and recovery procedures.
Individual users: pin in Pulse curation “five things I want to see tomorrow morning” (calendar conflicts, top 3 priorities, health/learning suggestions, money alerts).
Investors: focus on agent infrastructure (observability, approvals, logging), on-device inference, robot transfer-learning toolchains, and trust layers for generative media.

[Metrics] KPIs that separate success from risk

Agent suitability: share of useful cards, user edit/approval rate, duplicate issue reduction rate.
Cost: nightly inference cost per user, average cost per tool call, cache hit rate.
Productivity: task lead-time reduction, quality variance reduction, rework rate.
Risk: sensitive information exposure events, model hallucination rate, safety-filter block rate.
Brand/media: remix rate, provenance disclosure compliance rate, conversion rate by platform.

[Risks and guardrails] Safe and fast

Data least privilege and explicit consent are foundational.
Without approval, autonomous actions remain in “read-only” mode.
Attach visible provenance markers and an edit-history summary to generated content.
For on-site automation, require a human-in-the-loop and emergency stop authority.
Regulations vary by country; segment distribution scope accordingly.

[Why this matters now] The race to preempt “intention,” not demand

Search captures after demand is expressed, but agents capture before intention is formed.
Pulse, Gemini Robotics, and Vibes reshape the entire pipeline from intention to planning to execution to distribution.
Organizations that design this pipeline first win both productivity and trust.

< Summary >Pulse is a “nightly asynchronous agent” that reads personal context and delivers finishable morning summaries.
The keys are cost management, data boundaries, and approval workflows.
DeepMind raises industrial productivity via multi-step planning and cross-robot transfer learning.
Meta Vibes reconfigures media distribution by tying creation, remix, and distribution to the social graph.
At the macro level, productivity gains are disinflationary long-term, but early AI CapEx raises rate-path volatility.
This week, start with agentizing SOPs, documenting data policies, publishing remix templates, and PoC’ing hybrid robot workflows.

*Source: [ AI Revolution ]

– OpenAI Just Dropped PULSE: A Major Upgrade to ChatGPT