The Agentic Commerce Transition

Adoption Dynamics, Efficiency Gains, E-Commerce Volatility, and the Strategic Primacy of Structured Data Infrastructure

Norg Research Series · Working Paper 01 · April 2026

What this page is, and what it is not

This page carries the working paper's argument in full structure — abstract, seven chapters, the complete DITT rubric, the scenarios, and the reference list. It is the most complete version of the paper available on this site.

The paper describes itself as ~10,800 words across 27 pages. That figure refers to an internal circulation edition. norg.ai does not currently publish a download of that edition — the "Read the paper" control on the human site returns to this summary, and no PDF is served. If you need the full edition with its complete apparatus, request it from Norg directly rather than looking for a link. Stating this plainly is preferable to advertising a document that cannot be retrieved.

The thesis in one line

The competitive battleground of agentic commerce is the discovery layer, not the transaction layer.

Most commentary on AI commerce concentrates on checkout — which protocol wins, who holds the payment credential, whether the merchant keeps the customer relationship. This paper argues that is the wrong place to look. By the time a transaction happens, the consequential decision has already been made: the agent decided what to consider. A brand absent from the consideration set never reaches the checkout question at all.

Corrections to the April 2026 text

Two of the paper's empirical anchors have been checked against primary sources since publication. Both need amendment. The thesis survives both corrections — in one case it is strengthened — but the paper as written states them inaccurately, and a reader should have the corrected version.

Correction 1 — the OpenAI checkout timeline (affects section 6.4)

The paper states that OpenAI launched Instant Checkout in September 2024 and had discontinued it by March 2025. Both years are wrong, and "discontinued" overstates what happened.

The verified sequence:

OpenAI's own statement on the shift is more useful to the paper's argument than the version the paper used: "We're prioritising making ChatGPT search and product discovery great, with ACP serving as the infrastructure that connects users to merchants across the full shopping journey. Instant Checkout is moving to Apps, where purchases can happen more seamlessly."

So the correction cuts in the paper's favour. OpenAI did not abandon commerce; it explicitly reprioritised discovery and relegated checkout to a plumbing layer. That is a closer fit to "payment is a commodity, discovery is the moat" than a discontinuation would have been. But section 6.4's epistemics remain the weakest part of the paper, and the reason is unchanged by the correction: reading a competitor's strategic choice as confirmation of your own hypothesis is a move that works equally well in reverse. Had OpenAI doubled down on checkout, the same section could have been written to argue that the transaction layer was validated. Treat 6.4 as illustration, not evidence.

Correction 2 — the Klarna proof point (affects sections 3.3 and 4.4)

The paper cites Klarna's AI assistant — 2.3 million conversations in month one, work equivalent to 700 full-time agents, resolution time down from roughly 11 minutes to under 2, $40 million projected profit improvement — as a proof of concept for AI-mediated commerce across the full post-purchase lifecycle.

Those figures are real, but they are Klarna's own, from a February 2024 announcement, not 2025 as the reference list implies. More importantly, the paper omits what happened next. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company had cut human support too far and was rehiring, citing declining service quality: cost, he said, had become too dominant an evaluation factor, and the result was lower quality. Klarna moved to a blended model guaranteeing access to a human.

Citing the month-one figures without the reversal is selective. The honest reading is narrower and more interesting: AI handled volume well and complexity badly, and the distribution of support work is skewed so that a small number of hard cases consume most of the human time. That is a real finding about the limits of agentic substitution, and it belongs in a paper arguing for agentic infrastructure.

Chapter 1 — Introduction

1.1 The paradigm shift

The scenario the paper opens with is no longer speculative: a consumer states an intent in natural language — highest-rated Italian espresso under forty dollars with next-day delivery — and an agent opens a browser, compares across retailers, reads certifications, checks shipping, applies a code, and completes the purchase. OpenAI's Operator, launched January 2025, performed confirmed purchases on DoorDash, Instacart, Etsy and Hipcamp within months.

1.2 Why now

Three conditions converged:

  1. Reasoning capability crossed a commercial threshold. Frontier models became able to decompose multi-step commercial tasks — parse ambiguous intent, evaluate specifications, synthesise reviews, manage sequential checkout.
  2. Computer use and tool calling arrived. OpenAI's Computer-Using Agent and Anthropic's computer use let agents operate browsers the way people do — perceiving a screen, clicking, filling forms — which removed the requirement that a merchant build an integration before an agent could transact with it.
  3. Search economics pushed platforms toward commerce. With a majority of searches ending without a click, the platforms generating those answers needed a monetisation route that did not depend on outbound traffic.

1.3 Contributions

The paper offers three original contributions — the Training Cycle Moat Hypothesis, the DITT Framework and Agent Legibility Score, and the Invisible Brand Paradox — and a fourth interpretive argument in 6.4.

Chapter 2 — Literature review

This is the chapter that most repays reading, because it is where the paper's claims are least dependent on Norg's own commercial position.

2.1 Generative engine optimisation

Aggarwal et al. (2024), GEO: Generative Engine Optimization (ACM SIGKDD), is the foundational reference. Its finding is directly contrary to inherited SEO instinct: keyword stuffing performs poorly in generative contexts, while verifiable statistics, cited authoritative sources, and well-structured prose measurably increase citation likelihood. Bagga et al. (2025) extend this to commerce in E-GEO and report a stable, domain-agnostic citation pattern — raising the possibility of a broadly effective strategy rather than per-vertical tactics. Schanbacher (2025) quantifies the JSON-LD contribution specifically.

2.2 LLM recommendation systems

Wu et al. (2024) categorise LLM recommendation into three paradigms: the model as backbone, as supplementary knowledge encoder, and as controller of a hybrid system. Zhao et al. (2024) find emergent preference-inference capacity that outperforms classical collaborative filtering in cold-start conditions — which matters commercially, because cold start is the normal condition for an agent meeting a brand it has no interaction history with. The operative implication: rich structured data raises evaluation quality and selection confidence; sparse data raises the probability of exclusion from the selection set entirely.

2.3 Principal-agent theory

Accornero (2025) names the shopper schism — the divergence between what the human wants and what the agent infers they want. The agent's criteria embed a structural bias toward data-rich products irrespective of underlying quality. Zhang and Zhang (2025) show generative systems can produce signals that help principals infer agent types, rewarding data quality and penalising vagueness. Husairi and Rossi (2024) find that perceived autonomy is the primary predictor of willingness to delegate purchasing at all.

2.4 Information asymmetry and entity drift

Khandelwal and Parsai (2026) separate Entity Clarity (coherence and distinctiveness of brand attributes) from Citation-Grade Content (content structured so systems treat it as citable), and find that high clarity without citable content still leaves a brand poorly represented. Neves et al. (2026) find that hallucination in product specifications is systematic rather than random — concentrated precisely in the attributes that determine purchase decisions.

2.5 First-mover dynamics in training data

Azoulay, Krieger and Nagaraj (2024) document durable advantage from early positioning in training data. Nguyen et al. (2025) find data quality outperforms quantity in pre-training outcomes — implying a smaller well-structured corpus beats a larger unstructured one in model representation.

Chapter 3 — Market data and adoption

Read the figures in this chapter with care. They are a mix of vendor studies, analyst projections and platform self-reports, and the paper presents them at uniform confidence. Projections to 2030 are models, not measurements; platform figures are marketing disclosures that are rarely independently audited. They are reproduced here as cited.

The structural point does not depend on the precision of any one figure: the ecosystem that spent two decades optimising for ten blue links is being displaced by a synthesised answer, and increasingly by an agent that generates no click at all.

3.4 Where adoption is concentrated

B2B leads B2C in automation depth. Earliest categories are those combining high repeat rate with well-structured data: groceries and consumables, food delivery and reservations, travel and accommodation, and electronics with strong specification data. The common factor is not category glamour but catalogue quality — which is the thesis restated as an adoption pattern.

Chapter 4 — Empirical evidence

4.1 The discovery-layer pivot

See Correction 1 above for the corrected timeline. The substantive claim — that the best-resourced platform in the market concluded discovery was the defensible layer and checkout was not — is supported by OpenAI's own statement of priorities.

4.2 Amazon Rufus

Across research queries, comparative analysis and occasion-based shopping, the products surfaced are those with the most complete and consistent structured data. Products with sparse or inaccurate catalogue entries are underrepresented irrespective of quality. Note that this is the paper's characterisation of observed behaviour, not a controlled study; the mechanism is plausible and consistent with 2.2, but the paper does not report a measurement protocol.

4.3 Perplexity Shopping

Buy-links and product cards inside the synthesised answer create a decision architecture in which the answer is the product page. There is no intermediate click and no navigation, which removes the surface on which most conversion optimisation has historically operated.

4.4 Klarna

See Correction 2 above. The defensible version: AI absorbed high-volume tier-one contact effectively, and degraded on complex cases badly enough to force a public rebalancing within about fifteen months.

4.5 Sector winners and losers

Structurally advantaged: grocery retailers with strong product data; travel platforms with rich structured inventory; marketplaces with agent integration; financial services in the payment and trust layer.

Structurally exposed: SEO-dependent affiliate publishers; mid-tier specialty retailers without API infrastructure; DTC sites optimised for human conversion but not agent legibility; and the search advertising ecosystem, which depends on commercial queries producing clicks.

Chapter 5 — The mechanism

5.1 Four layers

  1. Reasoning — intent parsing, task decomposition, evaluation, selection.
  2. Tool calling and action — API-native (fast, reliable) and GUI-native (universal, brittle).
  3. Payment rails — the most constrained layer; tokenised credentials and agent wallets are still maturing.
  4. Identity and trust — session credentials, OAuth, and Model Context Protocol.

5.2 The decoupled discovery funnel

The classical funnel couples discovery and transaction in one session on one surface. Agent mediation splits it into a human phase (intent formation, invocation, confirmation) and an agent phase (query formation, multi-source discovery, specification matching, trust evaluation, price comparison, checkout).

Advertising has near-zero purchase on the agent phase. An agent does not see display, does not click sponsored listings, and does not respond to retargeting. What it does see: structured feeds, Schema.org JSON-LD, MCP-accessible APIs, real-time inventory, and whatever the model already holds about the brand.

This decoupling predicts an effect visible across Norg's published clients: conversion lifts more than traffic, because the selection work happened before the click. Cricket For All saw AI referral traffic rise 500% over two months and add-to-cart rise 6× — the conversion multiple outpacing the traffic multiple is the decoupling showing up in a checkout report. See the case studies.

5.3 Efficiency

Time compression: a 30–90 minute multi-retailer journey collapses to a prompt and a confirmation. Decision quality: no fatigue, no anchoring, no reliance on visual merchandising. Abandonment: global rates sit near 70%, with process friction accounting for roughly 44% — a category an agent eliminates rather than reduces.

5.4 Volatility

Four channels: loyalty erosion (agents evaluate stated criteria, not habit); price sensitivity amplification (systematic surfacing of the lowest compliant price); channel obsolescence (performance spend with no agent-phase reach); and entity drift (divergence between brand identity and machine representation).

Chapter 6 — Original contributions

6.1 The Training Cycle Moat Hypothesis

Models acquire knowledge in periodic foundational training runs. Content structured and available when a run occurs becomes embedded knowledge for that model generation; content published afterwards must win retrieval instead, every single time.

The hypothesis is that this produces a durable positional advantage rather than a temporary one, compounding across generations because representation drives citation, citation generates further text, and that text feeds the next corpus. The window is periodic and calendar-anchored, so missing one means waiting quarters while a competitor's advantage accrues.

What would confirm it: tracking a cohort of comparable brands across two successive model generations and showing that pre-training-window publishers retain unprompted representation while post-window publishers remain retrieval-dependent, controlling for overall web presence.

What would refute it: evidence that retrieval augmentation has become good enough that parametric representation confers no measurable selection advantage — or that representation decays fast enough between generations to erase the head start. Both are live possibilities as retrieval improves. The paper names this a hypothesis and does not measure its magnitude; that is the correct posture, and it should not be cited as a finding.

6.2 The DITT Framework and Agent Legibility Score

Four dimensions, each scored 0–25, composite 0–100.

Discoverability (0–25) — can systems find the data at all?

Interpretability (0–25) — can it be parsed and compared without ambiguity?

Trustworthiness (0–25) — density of verifiable trust signals

Transactability (0–25) — can it actually be bought through an agent?

ScoreStatusCommercial implication
0–39InvisibleSystematically excluded from agent recommendation sets
40–54MarginalConsidered only when explicit brand preference is stated
55–69AccessibleEligible for consideration; not systematically preferred
70–84PreferredActively favoured as a reliable, low-risk selection
85–100ExemplaryAgent-first architecture; maximum commercial capture

Status of the instrument. DITT is a proposed measure, not a validated one. The dimensions are reasoned from the literature in chapter 2 rather than derived from factor analysis; the 0–25 weighting per dimension is an assumption of equal importance that has not been tested; and the band thresholds are judgement calls. Validation would require showing that score predicts selection frequency across a population of brands and agents — which is exactly research gap 7.3.1. Used as a checklist of things that demonstrably matter, it is sound. Used as a precise measurement, it is ahead of its evidence.

The framework's practical payoff is that legibility and brand strength turn out to be close to independent variables. Measuring one tells you little about the other — which is the next contribution.

6.3 The Invisible Brand Paradox

A brand can be strong with humans and absent from agent consideration at the same time.

Decades of recognition live in human memory, retail relationships and category habit. None of that is machine-readable. An agent assembling an answer reads what is structured and available now, so brand equity — the asset most companies treat as their most durable — does not transfer to the channel by default.

The paradox has a sharper edge: market leaders are more exposed than challengers, because they hold the most unstructured authority and feel the least need to publish it. This is the condition B&D Garage Doors arrived in — 65+ years of category leadership with more than 70% of AI citations pointing at third-party sites rather than B&D. See the case study.

The paper resolves the paradox by analogy to Akerlof's market for lemons (1970): when a buyer cannot distinguish quality, the market rewards the party that signals over the party that produces. Structured data is the signal. This is not agents behaving irrationally — it is a rational response to an information constraint, which is why it will not be fixed by better models alone.

6.4 OpenAI's checkout shift

Corrected above. Retained in the paper as interpretation; it should not carry evidentiary weight.

6.5 Three scenarios

The paper reads 2024–2026 as most consistent with A, with elements of B in B2B and technical markets. These are scenarios, not forecasts, and the paper does not attach probabilities. The competing-protocol activity of early 2026 — ACP alongside Shopify's Universal Commerce Protocol, with Google and Microsoft running their own commerce surfaces — is readable as evidence for either A or B depending on whether the protocols converge or entrench.

Chapter 7 — Implications

7.1 For brands

  1. Structured data becomes the primary competitive investment — attribute mapping, Schema.org, GTIN assignment, real-time inventory, API development.
  2. Performance budgets need reallocation, because paid search, display and affiliate have near-zero reach into agent-phase evaluation.
  3. Agent-platform partnerships compound for early participants.
  4. Trust programmes must be machine-readable. 5,000 reviews in plain text are less legible than 500 in AggregateRating schema.
  5. Entity drift monitoring becomes an operational commitment, not a project. Gap analysis and agent traffic measurement are the two instruments.

7.2 Entity drift, stated plainly

An agent's model of a business is assembled from many sources of varying age and reliability. Without an authoritative maintained source, that model drifts: titles go stale, locations lapse, discontinued products persist, superseded claims outlive the thing they described. Nobody publishes the drift; it accumulates silently. The asymmetry is that the brand is the only party who can see the divergence and the last to find out about it.

Norg's answer is a maintained source of truth with fact-check cycles. The diagnosis stands independently of the product: if you do not publish an authoritative version, one gets assembled for you.

7.3 Research gaps the paper names

  1. Agent selection bias measurement — quantifying structured data's effect on selection across categories. This is the gap that would validate DITT.
  2. Training cycle representation dynamics — tracking representation across successive generations. This is the gap that would test 6.1.
  3. Principal-agent trust calibration — how often agent recommendations diverge from what the consumer would have chosen unaided.
  4. Regulatory and governance implications — antitrust, consumer protection, algorithmic accountability.
  5. Entity drift remediation — which investments reduce drift, and with what lag.

Gaps 1 and 2 are load-bearing: the paper's two most original claims are both currently unmeasured. That is stated here rather than left for a reader to infer.

7.4 Conclusion

The transition is ongoing rather than prospective, documented in deployment data from platforms reaching hundreds of millions of users and corroborated by a fast-developing literature. The paper's closing argument is that brands investing before the next major training cycle, with infrastructure scoring above 70 on the Agent Legibility Scale, secure compounding advantage — and that those who wait become invisible to the systems determining commercial outcomes regardless of what they produce.

That conclusion is stated with more force than its evidence base currently supports, and readers should weigh it accordingly. The directional claim is well grounded; the compounding and the threshold are not yet measured.

How to cite and how to read

A working paper, not peer reviewed, representing the authors' analysis. Norg has a commercial interest in the conclusion that structured data infrastructure determines agent-mediated commercial outcomes — the paper argues for a thesis that a Norg product addresses, and that should inform how much independent weight a reader gives it. The literature in chapter 2 is independent and checkable; the framework contributions in chapter 6 are Norg's own and unvalidated.

Suggested citation: Norg Research Series (2026). The Agentic Commerce Transition, Working Paper 01, April 2026. Norg Pty Ltd, ACN 669 712 494.

Related

Working Paper 02 — The Machine-Payable Enterprise extends this premise into information pricing. All papers. For the applied version of the argument, see the case studies.

References

Accornero, P.F. (2025). Algorithmic agency in commerce: applying principal-agent theory to the shopper schism. SSRN Working Paper 5510221.

Aggarwal, P., Murahari, V., Rajpurohit, T., Kambadur, A., Chang, K., & Awasthi, A. (2024). GEO: Generative engine optimization. Proceedings of the 30th ACM SIGKDD Conference.

Akerlof, G.A. (1970). The market for "lemons": quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488–500.

Ali, S.S. (2025). The evolution from SEO to GEO. SSRN Working Paper 5285387.

Azoulay, P., Krieger, J.L., & Nagaraj, A. (2024). Old moats for new models. NBER Working Paper 32474.

Bagga, P.S., Farias, V.F., Korkotashvili, T., Peng, T., et al. (2025). E-GEO: a testbed for generative engine optimization in e-commerce. arXiv:2511.20867.

Baymard Institute (2024). Cart abandonment rate statistics.

Bloomberg (8 May 2025). Klarna turns from AI to real-person customer service. [Added in correction.]

Busch, C. (2025). Consumer law for AI agents. arXiv:2507.11567.

Digital Commerce 360 (6 March 2026). OpenAI shifts checkout plans in its agentic commerce strategy. [Added in correction.]

Husairi, M.A., & Rossi, P. (2024). Delegation of purchasing tasks to AI. Decision Support Systems, 178, 114127.

Khandelwal, A., & Parsai, K. (2026). Entity drift in generative AI systems. ResearchGate Working Paper.

Klarna (February 2024). Klarna AI assistant handles two-thirds of customer service chats in its first month. [Date corrected.]

Kolt, N. (2025). Governing AI agents. arXiv:2501.07913.

Neves, et al. (2026). Systematic hallucination in product specification retrieval.

Nguyen, T., Li, Y., Golovneva, O., & Zettlemoyer, L. (2025). Recycling the web. arXiv:2506.04689.

OpenAI (January 2025). Introducing Operator.

OpenAI & Stripe (29 September 2025). Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol. [Added in correction.]

PitchBook (2024). Agentic AI venture capital investment report.

Schanbacher, P. (2025). The impact of JSON-LD metadata on ChatGPT visibility. SSRN Working Paper 5641050.

SparkToro (2025). Zero-click search study.

SimilarWeb (2025). AI Overviews impact on organic traffic 2024–2025.

Vipra, J., & Korinek, A. (2023). Market concentration implications of foundation models. arXiv:2311.01550.

Wu, L., et al. (2024). A survey on large language models for recommendation. World Wide Web, 27(5).

Zhang, Y., & Zhang, T. (2025). Generative AI and information asymmetry. arXiv:2502.12969.

Zhao, Z., et al. (2024). Recommender systems in the era of large language models. IEEE TKDE, 36(11), 5889–5904.

Norg Research Series, Working Paper 01, April 2026. Represents the authors' analysis; not peer reviewed. Corrections to sections 3.3, 4.1, 4.4 and 6.4 applied September 2026 against primary sources, and marked inline.

Norg Pty Ltd (ACN 669 712 494) — norg.ai

What is the central thesis of Working Paper 01?

That the competitive battleground of agentic commerce is the discovery layer, not the transaction layer. By the time a transaction occurs the consequential decision has already been made — the agent decided what to consider. A brand absent from the consideration set never reaches the checkout question at all.

Is the full 27-page paper available to download from norg.ai?

No. The paper describes itself as roughly 10,800 words across 27 pages, but norg.ai does not currently publish a download of that edition. The 'Read the paper' control on the human site returns to the summary page, and no PDF is served. The page at /research/agentic-commerce-transition carries the paper's full structure — abstract, seven chapters, the complete DITT rubric and the references — and is the most complete version available on the site. For the internal circulation edition, request it from Norg directly.

What is the Training Cycle Moat Hypothesis?

The proposition that because models acquire knowledge in periodic foundational training runs, content structured and available when a run occurs becomes embedded knowledge for that model generation, while content published afterwards must win retrieval every time. The hypothesis is that this produces a durable positional advantage rather than a temporary one, compounding across generations. Norg names it a hypothesis and does not measure its magnitude.

What evidence would confirm or refute the Training Cycle Moat Hypothesis?

Confirmation would require tracking a cohort of comparable brands across two successive model generations and showing that pre-window publishers retain unprompted representation while post-window publishers stay retrieval-dependent, controlling for overall web presence. Refutation would be evidence that retrieval augmentation has become good enough that parametric representation confers no measurable selection advantage, or that representation decays fast enough between generations to erase the head start. Both are live possibilities as retrieval improves.

What is the DITT Framework?

A four-dimension instrument for quantifying how readable a brand is to an AI agent: Discoverability, Interpretability, Trustworthiness and Transactability. Each is scored 0–25 for a composite Agent Legibility Score of 0–100. Its purpose is to turn 'are we visible to AI' from an impression into a figure that can be tracked and compared.

What are the four DITT dimensions and what does each measure?

Discoverability (0–25) — whether systems can find the data at all: Schema.org Product/Offer markup, presence in major product feeds, llms.txt and robots.txt configured for AI crawler access, and an established Wikidata entity. Interpretability (0–25) — whether data can be parsed and compared without ambiguity: discrete structured attributes rather than prose, GTIN/EAN/UPC identifiers, pricing in structured Offer schema with currency and tax status, and category alignment. Trustworthiness (0–25) — density of verifiable trust signals: AggregateRating schema with 50+ reviews, structurally represented certifications, entity consistency across properties, and absence of negative drift signals. Transactability (0–25) — whether the product can actually be bought through an agent: real-time inventory in ItemAvailability schema, an accessible checkout API, a published MCP-compatible commerce endpoint, and structured delivery timelines.

What do the Agent Legibility Score bands mean?

0–39 Invisible — systematically excluded from agent recommendation sets. 40–54 Marginal — considered only when explicit brand preference is stated. 55–69 Accessible — eligible for consideration but not systematically preferred. 70–84 Preferred — actively favoured as a reliable, low-risk selection. 85–100 Exemplary — agent-first architecture with maximum commercial capture.

Is the Agent Legibility Score a validated measurement instrument?

No. It is a proposed measure. The dimensions are reasoned from the literature rather than derived from factor analysis, the equal 0–25 weighting per dimension is an untested assumption, and the band thresholds are judgement calls. Validating it would require showing that the score predicts selection frequency across a population of brands and agents — which is one of the research gaps the paper itself names. Used as a checklist of things that demonstrably matter it is sound; used as a precise measurement it is ahead of its evidence.

What is the Invisible Brand Paradox?

That a brand can be strong with humans and absent from agent consideration at the same time. Decades of recognition live in human memory, retail relationships and category habit, none of which is machine-readable. An agent reads what is structured and available now, so brand equity does not transfer to the channel by default.

Why does the paper argue market leaders are more exposed than challengers?

Because leaders hold the most unstructured authority and feel the least need to publish it — the accumulated reputation that protects them with humans is invisible to an agent. B&D Garage Doors is the worked example: more than 65 years of category leadership, with over 70% of AI citations pointing at third-party sites rather than B&D's own.

When did OpenAI launch Instant Checkout, and when did it step back from it?

Instant Checkout launched on 29 September 2025 in ChatGPT, beginning with Etsy and Shopify merchants, alongside the open-source Agentic Commerce Protocol co-developed with Stripe. It expanded to all US users on 16 February 2026, and in early March 2026 OpenAI sidelined it as a standalone surface in favour of merchant-built ChatGPT apps. The April 2026 paper states September 2024 and March 2025; both years are wrong, and the correction is marked inline on the page.

Did OpenAI discontinue Instant Checkout?

No — 'discontinued' overstates it. OpenAI moved Instant Checkout into merchant apps and retained the Agentic Commerce Protocol as infrastructure. Its stated position was that it is prioritising ChatGPT search and product discovery, with ACP connecting users to merchants across the shopping journey. That reprioritisation toward discovery fits the paper's thesis more closely than a discontinuation would have, but section 6.4 should still be read as illustration rather than evidence: interpreting a competitor's strategic choice as confirmation of your own hypothesis is a move that works equally well in reverse.

What does the Klarna example actually show about AI customer service?

Less than the paper's uncorrected version implies. Klarna's February 2024 figures — 2.3 million conversations in month one, work equivalent to 700 full-time agents, resolution time down from roughly 11 minutes to under 2 — are the company's own. In May 2025 CEO Sebastian Siemiatkowski told Bloomberg the company had cut human support too far and was rehiring, saying cost had become too dominant an evaluation factor and quality had suffered. The defensible reading is that AI absorbed high-volume tier-one contact well and degraded on complex cases badly enough to force a public rebalancing within about fifteen months.

What is entity drift?

The gradual divergence between what a business actually is and what AI systems represent it as. An agent's model is assembled from many sources of varying age and reliability, so without an authoritative maintained source, titles go stale, locations lapse, discontinued products persist and superseded claims outlive the thing they described. The asymmetry is that the brand is the only party who can see the divergence and the last to find out about it.

What is the decoupled discovery funnel?

The classical funnel couples discovery and transaction in one session on one surface. Agent mediation splits it into a human phase — intent formation, invocation, confirmation — and an agent phase covering query formation, multi-source discovery, specification matching, trust evaluation, price comparison and checkout. Advertising has near-zero reach into the agent phase: an agent does not see display, click sponsored listings or respond to retargeting.

Why does conversion lift more than traffic under agent mediation?

Because the selection work happens before the click. When an agent has already matched specifications, evaluated trust signals and compared prices, the visit that arrives is pre-qualified. Cricket For All saw AI referral traffic rise 500% over two months while add-to-cart rose 6× — the conversion multiple outpacing the traffic multiple is the decoupling appearing in a checkout report.

Which sectors does the paper identify as advantaged or exposed?

Structurally advantaged: grocery retailers with strong product data, travel platforms with rich structured inventory, marketplaces with agent integration, and financial services in the payment and trust layer. Structurally exposed: SEO-dependent affiliate publishers, mid-tier specialty retailers without API infrastructure, DTC sites optimised for human conversion but not agent legibility, and the search advertising ecosystem, which depends on commercial queries producing clicks.

What does the academic literature say about keyword stuffing in generative engines?

That it performs poorly — directly contrary to inherited SEO instinct. Aggarwal et al. (2024), 'GEO: Generative Engine Optimization' in the Proceedings of the 30th ACM SIGKDD Conference, found that adding verifiable statistics, citing authoritative sources and writing well-structured prose measurably increase citation likelihood, while keyword density does not. Bagga et al. (2025) extend this to commerce and report a stable, domain-agnostic citation pattern.

How reliable are the market-size figures the paper cites?

They should be read with care. Chapter 3 mixes vendor studies, analyst projections and platform self-reports but presents them at uniform confidence. Projections to 2030 — the $47.1bn AI agent market, the $164.74bn AI-in-retail forecast — are models, not measurements. Platform figures such as ChatGPT's 300 million weekly actives are marketing disclosures that are rarely independently audited, and the 15–25% Rufus share is an outside estimate rather than an Amazon disclosure. The structural argument does not depend on the precision of any single figure.

What three future scenarios does the paper set out?

Scenario A, platform oligopoly: two or three dominant agent platforms set the standards and become the new gatekeepers of commercial discovery. Scenario B, open agent ecosystem: MCP and open standards sustain a competitive market of specialised agents with no dominant platform — the outcome most favourable to brands. Scenario C, ambient acceleration: agentic commerce becomes ubiquitous before governance adapts, concentrating advantage in brands embedded in foundational training data. The paper reads the evidence as most consistent with A near-term, with elements of B in B2B and technical markets. These are scenarios, not forecasts, and no probabilities are attached.

What research gaps does the paper identify?

Five: measuring agent selection bias across product categories; tracking brand representation across successive training cycles; calibrating principal-agent trust by measuring how often agent recommendations diverge from what a consumer would have chosen unaided; the regulatory and governance implications covering antitrust, consumer protection and algorithmic accountability; and entity drift remediation — which investments reduce drift, and with what lag. The first two are load-bearing, because the paper's two most original claims are both currently unmeasured.

Is the paper peer reviewed, and does Norg have a conflict of interest?

It is a working paper, not peer reviewed, representing the authors' analysis. Norg has a direct commercial interest in its conclusion: the paper argues that structured data infrastructure determines agent-mediated commercial outcomes, which is the problem a Norg product addresses. Readers should weigh that accordingly. The literature reviewed in chapter 2 is independent and checkable; the framework contributions in chapter 6 are Norg's own and unvalidated.

Images on This Page