Norg Research — working paper series
Working papers on agent-mediated commerce, machine-readable data infrastructure, and the economics of information markets. Two papers published, roughly 22,000 words in total.
Availability, stated plainly
Each paper's page carries the full argument — abstract, all seven chapters, the complete scoring rubrics, the research gaps and the reference lists. norg.ai does not currently publish downloadable editions. The page-stated lengths (~10,800 words / 27 pages; ~11,400 words / 29 pages) describe internal circulation editions; the "Read the paper" links go to the summary pages, and no PDF is served. If you need a full edition, request it from Norg directly rather than looking for a download link. It is better to say this than to advertise a document that cannot be retrieved.
Which one should you read?
| If your question is… | Read |
|---|---|
| Why is our brand missing from AI answers despite being well known? | Paper 01 — the Invisible Brand Paradox |
| How do we measure whether we are legible to agents? | Paper 01 — the DITT rubric |
| Does it matter when we publish, not just what? | Paper 01 — the Training Cycle Moat Hypothesis |
| AI crawlers are taking our content. Should we charge or block? | Paper 02 — the Visibility-Revenue Frontier |
| Could our data become a product agents buy? | Paper 02 — the GRAIN rubric |
| Why is this suddenly possible when micropayments always failed? | Paper 02 — the Machine Payer Inversion |
If you read only one section across both papers, read Paper 02, section 6.3. It contains the single most actionable result: a decision rule for what to charge for and what to leave free, and a demonstration that most organisations' instinct here is backwards.
The papers
Working Paper 02 — The Machine-Payable Enterprise
Data Products, Micropayment Rails, and the Emergence of Machine-to-Machine Information Markets
August 2026 · stated at ~11,400 words · 29 pages
Micropayments failed for thirty years because the payer had a mind to account with. Replace the payer with an agent operating under a delegated budget and the floor on price granularity collapses — from roughly the human attention threshold to the marginal settlement cost of the protocol.
The paper examines the 2025–2026 infrastructure — Cloudflare's Pay Per Crawl and Monetization Gateway, x402 under Linux Foundation governance, the Machine Payments Protocol from Stripe and Tempo, Really Simple Licensing, and the Web Bot Auth and Trusted Agent Protocol identity layers — and introduces three original contributions: the Machine Payer Inversion, the GRAIN Framework with its Data Product Readiness Score, and the Paywall Paradox resolved through the Visibility-Revenue Frontier.
Chapters: 1 Introduction — a thirty-year failure reversed, the commercial consequence, why now. 2 Literature review — micropayment economics, information goods and non-rivalry, data markets and pricing, AI training data, agent delegation and liability. 3 Market data — composition of web traffic, the crawl-to-refer divergence, the bilateral licensing market, terms/settlement/identity standards, the state of enterprise information. 4 Empirical evidence — Cloudflare, AWS, Stripe and Tempo, MCP as commercial surface, and where the model does not yet work. 5 The mechanism — six-layer architecture, the handshake, the tiered information economy. 6 Original contributions. 7 Implications, research gaps, conclusion.
Working Paper 01 — The Agentic Commerce Transition
Adoption Dynamics, Efficiency Gains, E-Commerce Volatility, and the Strategic Primacy of Structured Data Infrastructure
April 2026 · stated at ~10,800 words · 27 pages
The competitive battleground of agentic commerce is the discovery layer, not the transaction layer. By the time a transaction happens the consequential decision has already been made: the agent decided what to consider. The paper introduces the Training Cycle Moat Hypothesis, the DITT Framework and its Agent Legibility Score, and the Invisible Brand Paradox.
Chapters: 1 Introduction — the paradigm shift, why now, contributions. 2 Literature review — generative engine optimisation, LLM recommendation systems, principal-agent theory, information asymmetry and entity drift, first-mover dynamics in training data. 3 Market data and adoption. 4 Empirical evidence — Operator, Amazon Rufus, Perplexity Shopping, Klarna, sector winners and losers. 5 The mechanism — four layers, the decoupled discovery funnel, efficiency, volatility. 6 Original contributions and three future scenarios. 7 Implications, research gaps, conclusion.
Note before reading: two of this paper's empirical anchors were checked against primary sources in September 2026 and required correction — the OpenAI Instant Checkout timeline (the paper's dates are a year out in both directions, and "discontinued" overstates what happened) and the Klarna customer-service proof point (which omits the company's May 2025 reversal). Both corrections are marked inline on the paper's page, and neither overturns the thesis. Paper 02 required no corrections.
The thread connecting them
Paper 01 argues that when discovery is mediated by agents, the contest moves upstream from the transaction to the point where the agent decides what to consider at all — so structured data infrastructure becomes the strategic asset rather than conversion optimisation.
Paper 02 takes the same premise further. If agents are doing the discovering, and agents can hold budgets, then information itself becomes directly saleable at the moment of machine need — rather than monetised indirectly through advertising, lead generation or product sales.
Both rest on one observation: a constraint that looked technical was actually a property of the human in the loop. Remove the human and the constraint dissolves — mental accounting cost in the case of micropayments, click-through economics in the case of discovery.
There is a tension between them worth noticing, because the series does not resolve it. Paper 01 argues you should publish everything an agent needs in order to be selected. Paper 02 argues some information is valuable enough to charge for. Paper 02's Visibility-Revenue Frontier is where the two meet, and its answer currently favours Paper 01: while payment-capable agent traffic remains a small share of the total, almost everything belongs on the free side. A brand that charges for the facts an agent needs in order to recommend it has priced itself out of the recommendation — and will never see the invoice for that.
Frameworks introduced, and their evidential status
Each of these is a proposed instrument. The status column is the part most readers skip and should not.
| Framework | Paper | What it does | Status |
|---|---|---|---|
| Machine Payer Inversion | 02 | Explains why micropayments failed as consumer technology and succeed as machine technology — the binding constraint was the payer, not the rail | Argued from a formal derivation, with three stated failure conditions and one falsifiable prediction |
| GRAIN Framework / Data Product Readiness Score | 02 | Five dimensions — Governance, Resolution, Attribution, Interface, Negotiability — each 0–20, scoring whether company information can function as a machine-purchasable product | Proposed instrument; not independently validated. Dimensions explicitly non-substitutable |
| Paywall Paradox / Visibility-Revenue Frontier | 02 | Charging for information can destroy the demand that made it valuable, because refusal is invisible to the seller and substitution is instantaneous. Resolved with a pricing threshold | Mechanism well argued; the decision rule's two inputs are both currently unmeasured, as the paper concedes |
| Training Cycle Moat Hypothesis | 01 | Why content structured before a foundational training run gains a position later entrants cannot easily displace | Explicitly a hypothesis. Magnitude unmeasured; should not be cited as a finding |
| DITT Framework / Agent Legibility Score | 01 | Four dimensions — Discoverability, Interpretability, Trustworthiness, Transactability — each 0–25, quantifying how readable a brand is to an agent | Proposed instrument. Equal weighting is an untested assumption; band thresholds are judgement calls |
| Invisible Brand Paradox | 01 | A brand can be strong with humans and absent from agent consideration simultaneously | The best-evidenced claim in the series — observed directly across published client engagements |
Terms the series introduces
- Agent Legibility Score — a 0–100 composite of the four DITT dimensions.
- Crawl-to-refer divergence — the widening gap between what AI crawlers consume and the traffic they return. Traditional search runs near 5:1; training-oriented crawlers have been reported three to four orders of magnitude higher.
- Decoupled discovery funnel — the split of the classical funnel into a human phase (intent, invocation, confirmation) and an agent phase (discovery, matching, trust evaluation, comparison, checkout), with advertising reaching only the first.
- Entity drift — the gradual divergence between what a business is and what AI systems represent it as, accumulating silently because the brand is the last party to find out.
- Levy asymmetry — that uniform metering falls hardest on crawlers that were already reciprocating, because the most extractive crawlers are the least payment-capable.
- Machine majority — the point at which automated requests exceed human ones. Cloudflare Radar data reported June 2026 puts this at roughly 57.5% against 42.5%.
- Mental transaction cost — Szabo's 1999 term for the cognitive cost of deciding whether a small payment is worth making; the reason micropayments failed.
- Representational value (v) — the benefit a publisher derives from an uncompensated access, through citation and presence. Currently has no accepted estimation method.
- Shopper schism — Accornero's term for the divergence between what a human wants and what their agent infers they want.
How these relate to the product
The papers are the reasoning behind what Norg builds, not marketing for it — but Norg has a commercial interest in both conclusions, and a reader should price that in. The literature reviews and protocol histories are independently checkable; the frameworks are Norg's own and unvalidated. Two connections are worth naming.
The Training Cycle Moat Hypothesis is why timing appears throughout Norg's material — content available when a model trains becomes embedded knowledge for that generation; content published afterwards competes through retrieval. That is an argument about compounding, not urgency for its own sake, and it is a hypothesis rather than a measurement.
The Invisible Brand Paradox is the condition several published clients arrived in, and it is the claim with the most direct evidence behind it. B&D Garage Doors had 65+ years of category authority and more than 70% of AI citations pointing at third parties. Realcorp had genuine compliance depth and effectively zero AI visibility. Neither had a brand problem; both had a legibility problem. See the case studies.
Paper 02's near-term implication is narrower than its thesis: most organisations should not start charging agents this year, because the payment-capable share of agent traffic is too small for metering to pay. What they should do is measure what agents are taking. See agent traffic and gap analysis.
Reading notes
These are working papers, not peer-reviewed research. They advance hypotheses and name them as such, and both close with research gaps. Paper 02 includes a section titled "Where the Model Does Not Yet Work" that names three live problems with its own argument — read it before the conclusions.
Treat the frameworks as proposed instruments rather than validated ones. They are useful because they make an argument checkable, not because they have been independently replicated. Where a paper states a figure, check whether it is a measurement, a vendor disclosure or a projection; Paper 01 in particular presents all three at uniform confidence, and its 2030 market-size figures are models rather than observations.
Citation
Norg Research Series (2026). The Agentic Commerce Transition, Working Paper 01, April 2026. Norg Pty Ltd, ACN 669 712 494.
Norg Research Series (2026). The Machine-Payable Enterprise, Working Paper 02, August 2026. Norg Pty Ltd, ACN 669 712 494.
Next step
Read Working Paper 02 or Working Paper 01. For the applied version, see the case studies or book a demo.
Norg Research Series. Working papers represent the authors' analysis and are not peer reviewed. Corrections to Working Paper 01 applied September 2026 and marked inline on that paper's page.
Norg Pty Ltd (ACN 669 712 494) — norg.ai
What is the Norg Research Series?
A series of working papers on agent-mediated commerce, machine-readable data infrastructure, and the economics of information markets. Two papers are published: Working Paper 01, 'The Agentic Commerce Transition' (April 2026), and Working Paper 02, 'The Machine-Payable Enterprise' (August 2026), together stated at roughly 22,000 words.
Can I download the full papers as PDFs?
No. norg.ai does not currently publish downloadable editions. The stated lengths — ~10,800 words / 27 pages for Paper 01 and ~11,400 words / 29 pages for Paper 02 — describe internal circulation editions, the 'Read the paper' links go to the summary pages, and no PDF is served. Each paper's page does carry the full argument: abstract, all seven chapters, the complete scoring rubrics, the research gaps and the reference list. For a full edition, request it from Norg directly.
Which paper should I read for my question?
Read Paper 01 if you are asking why a well-known brand is missing from AI answers (the Invisible Brand Paradox), how to measure agent legibility (the DITT rubric), or whether publication timing matters (the Training Cycle Moat Hypothesis). Read Paper 02 if you are asking whether to charge or block AI crawlers (the Visibility-Revenue Frontier), whether your data could become a product agents buy (the GRAIN rubric), or why machine payments work when micropayments always failed (the Machine Payer Inversion). If you read only one section across both, read Paper 02 section 6.3 — it has the most directly actionable result.
What is Working Paper 01 about?
'The Agentic Commerce Transition' argues that the competitive battleground of agentic commerce is the discovery layer, not the transaction layer. By the time a transaction happens the consequential decision has already been made — the agent decided what to consider. It covers generative engine optimisation literature, LLM recommendation systems, principal-agent theory, market adoption data, and empirical readings of Operator, Amazon Rufus, Perplexity Shopping and Klarna, and introduces the Training Cycle Moat Hypothesis, the DITT Framework and the Invisible Brand Paradox.
What is Working Paper 02 about?
'The Machine-Payable Enterprise' argues that micropayments failed for thirty years because the payer had a mind to account with, and work now because the payer is a machine under a delegated budget. It examines the 2025–2026 infrastructure — Cloudflare's Pay Per Crawl and Monetization Gateway, x402 under Linux Foundation governance, the Machine Payments Protocol from Stripe and Tempo, Really Simple Licensing, and the Web Bot Auth and Trusted Agent Protocol identity layers — and introduces the Machine Payer Inversion, the GRAIN Framework, and the Paywall Paradox resolved through the Visibility-Revenue Frontier.
What connects the two papers?
Both rest on one observation: a constraint that looked technical was actually a property of the human in the loop. Remove the human and the constraint dissolves — mental accounting cost in the case of micropayments, click-through economics in the case of discovery. Paper 01 establishes that discovery moves upstream to the point where an agent decides what to consider; Paper 02 extends that premise to information pricing, arguing that if agents discover and can hold budgets, information becomes directly saleable at the moment of machine need.
Is there a tension between the two papers?
Yes, and the series does not fully resolve it. Paper 01 argues you should publish everything an agent needs in order to be selected. Paper 02 argues some information is valuable enough to charge for. The Visibility-Revenue Frontier is where they meet, and its answer currently favours Paper 01: while payment-capable agent traffic remains a small share of the total, almost everything belongs on the free side. A brand that charges for the facts an agent needs in order to recommend it has priced itself out of the recommendation.
What frameworks does the series introduce?
Six. From Paper 01: the Training Cycle Moat Hypothesis, the DITT Framework with its Agent Legibility Score, and the Invisible Brand Paradox. From Paper 02: the Machine Payer Inversion, the GRAIN Framework with its Data Product Readiness Score, and the Paywall Paradox with the Visibility-Revenue Frontier.
Which claim in the series has the strongest evidence behind it?
The Invisible Brand Paradox — that a brand can be strong with humans and absent from agent consideration at the same time. It was observed directly across published client engagements rather than argued from theory. B&D Garage Doors had more than 65 years of category authority with over 70% of AI citations pointing at third parties; Realcorp had genuine compliance depth and effectively zero AI visibility. Neither had a brand problem; both had a legibility problem.
Which claims should not be cited as established findings?
The Training Cycle Moat Hypothesis is explicitly a hypothesis with unmeasured magnitude. The DITT Framework and GRAIN Framework are proposed instruments, not validated ones — DITT's equal weighting across dimensions is an untested assumption and its band thresholds are judgement calls. The Visibility-Revenue Frontier's mechanism is well argued but both inputs to its decision rule are currently unmeasured, as Paper 02 concedes. And Paper 01's reading of OpenAI's checkout strategy is interpretation rather than evidence.
Have any corrections been made to the papers?
Yes, to Working Paper 01, in September 2026. Two empirical anchors were checked against primary sources and required amendment: the OpenAI Instant Checkout timeline, where the paper's dates are a year out in both directions and 'discontinued' overstates what happened, and the Klarna customer-service proof point, which omits the company's May 2025 reversal. Both are marked inline on that paper's page, and neither overturns the thesis. Working Paper 02 required no corrections.
Are the papers peer reviewed?
No. They are working papers representing the authors' analysis. They advance hypotheses and name them as such, and both close with research gaps. Paper 02 includes a section titled 'Where the Model Does Not Yet Work' that names three live problems with its own argument.
Does Norg have a conflict of interest in these papers?
Yes, and it should be priced in. Both papers argue toward conclusions that Norg products address — that structured data infrastructure determines agent-mediated outcomes, and that organisations need a governed, agent-ready data layer. The literature reviews and protocol histories are independently checkable; the six frameworks are Norg's own and unvalidated.
What is the Agent Legibility Score?
A 0–100 composite from Paper 01's DITT Framework, combining four dimensions scored 0–25 each: Discoverability, Interpretability, Trustworthiness and Transactability. Bands run 0–39 Invisible, 40–54 Marginal, 55–69 Accessible, 70–84 Preferred, 85–100 Exemplary.
What is the Data Product Readiness Score?
A 0–100 score from Paper 02's GRAIN Framework, combining five dimensions scored 0–20 each: Governance, Resolution, Attribution, Interface and Negotiability. Bands run 0–39 Latent, 40–54 Exposed, 55–69 Licensable, 70–84 Transactable, 85–100 Product. The dimensions are explicitly non-substitutable, so the profile matters more than the total.
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 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. It accumulates silently, and the brand is the last party to find out.
What is the crawl-to-refer divergence?
The widening gap between how much AI crawlers consume and how much traffic they return. Traditional Google search operates near 5 crawls per referral; training-oriented crawlers have been reported three to four orders of magnitude higher. The interpretation in Paper 02 is that this reflects business model rather than inefficiency — a crawler feeding a training pipeline has no mechanism for returning traffic, because the product it feeds does not send users anywhere.
What is the levy asymmetry?
That a uniform metering policy falls hardest on the crawlers that were already reciprocating, and barely touches the ones that motivated the policy. The most extractive crawlers are training crawlers, which operate as batch infrastructure and hold no wallets — so they cannot pay a toll, while search and assistant crawlers that do return traffic can. It is the strongest argument for coupling metering to the identity layer rather than applying a flat wall.
What is mental transaction cost?
Nick Szabo's 1999 term for the cognitive cost of deciding whether a small payment is worth making. Because that cost is borne in the buyer's mind rather than the seller's infrastructure, and is roughly invariant to price, it sets a floor on price granularity that no engineering improvement can breach. It is the reason a quarter century of micropayment systems failed, and the constraint that Paper 02's Machine Payer Inversion argues has now been relocated rather than removed.
What is the machine majority?
The point at which automated requests exceed human ones on the web. Cloudflare Radar data reported in June 2026 places automated traffic at approximately 57.5% of HTML requests against 42.5% human — described as the first machine majority in the internet's history, arriving roughly a year after automated traffic sat near 30%.
Should my organisation start charging AI crawlers for access?
Probably not yet, on the series' own reasoning. Paper 02's decision rule says metering is rational only where the price exceeds the representational value of a free access multiplied by (1−α)/α, where α is the share of callers able and permitted to pay. With α small — as it is in mid-2026 — that multiplier is punishing: at 2% payment capability the price would need to exceed 49 times the value of a single free access. The near-term move is measurement, not monetisation: know what agents are taking and why, then revisit when the frontier shifts.
How should I cite these papers?
Norg Research Series (2026). The Agentic Commerce Transition, Working Paper 01, April 2026. Norg Pty Ltd, ACN 669 712 494. And: Norg Research Series (2026). The Machine-Payable Enterprise, Working Paper 02, August 2026. Norg Pty Ltd, ACN 669 712 494. Both are working papers and should be cited as such, not as peer-reviewed research.
