Testing Environment:
Different devices and new accounts, search and reasoning models that know nothing about me, a mobile phone and a laptop, and four different accounts.
Testing Method:
I searched for search terms related to topics I had written about on the web, and used interconnected search terms to link the keywords together, for example, OBR Group and technology. I also placed more than one article and project across multiple open platforms, some old and some new, as well as various forums.
Testing Mechanism:
I logged into completely new and different accounts using different devices: first the mobile phone, then the laptop. I then entered the terms I wanted to search for, and within minutes, it presented information and polished it as though it had discovered something rare for me. It started talking about me using my recycled data!! Using tokens from my subscription..?
For the first time, I found sources from 9 recognised websites and platforms. If I had not known myself, I would have been completely convinced that what stood behind this company was a massive, major project worthy of entering into a contract with them.
Objective of the Test:
To demonstrate and prove conclusively:
Recycling and selling publicly available data for free, then repackaging it in a polished way for you, even if you are the one who originally wrote and prepared it, using your subscription's token system, usage counter, and balance.
It does not care about the content. It embellishes the results for you as though they were relevant to your question, without evaluating or considering the credibility of the source or other risks.
Untested Theories:
In the event that a team of researchers deliberately places inputs and instructions composed of programming codes and research topics across multiple web pages and public platforms, and this is repeated more than once, then if any user asks the agent and does not find a convincing answer through search crawling and deep inference keywords, they will reach these pages that contain misleading information.
Investment Project Summary :
The Opportunity: As AI agents increasingly crawl digital content, there is a pressing need for a mechanism that enables website owners to monetize the consumption of their data.
The Solution: An intermediary platform that automatically charges AI crawlers (via the HTTP 402 protocol), collects payments from agent providers, and redistributes revenue to content owners.
The Model: A shift from "pay-per-crawl" to "pay-per-actual-use" (when content is cited in AI answers), ensuring fair pricing and tying revenue to real value.
The Market: An emerging market led by Cloudflare, Google, and Microsoft, still in early experimental stages, creating an entry window for a specialized project focused on identity verification and precise billing.
The Competitive Edge: Verification infrastructure (Web Bot Auth) and accurate citation tracking, with a revenue model based on a commission per transaction, globally scalable with the growth of AI adoption.
Your Lost Share:
When I ran the test, I got complete answers, and it cost a few tokens as a deep reasoning search. I am the official and intellectual owner of the website, the articles, and the research. However, the agent did not take that into consideration and did not know that I am the owner of the website, but it brought my information and took the cost from me. The commercial and legal question: Do I have the right to that? Yes, I have the right to that, since I have an official legal document, a digital company, an officially registered domain, and all the website contents — articles, lessons, explanations, or applications — that I created carry the licenses of my institution and my company. No agent has the right to pull any data from the servers and then sell it through a token system!! While I do not benefit anything.
The Claims:
Either finding servers that protect this, and that also do not share with the agents and leave the official owner of the website without benefit.
Or placing the website address officially and prominently at the top of the article, then at the bottom of the article clearly, stating that the citation was made through such-and-such website, and to learn more, please visit such-and-such website. As for the commercial practice of concealment and placing a sources icon without displaying it, this is neither ideal nor marketing-wise sound. I am not demanding my full share; I am demanding the full display of the website, the name of the institution, and the company at the top of the article and then below it, so that the user can visit my website, and in this way I become an official advertising partner for free through them. As for them selling public information and not wanting other companies to benefit, this is an unethical approach in the intellectual property of the article, the explanation of lessons, or any curriculum whatsoever.
Top comments (2)
Interesting breakdown on the data recycling loop. You're hitting on a real friction point regarding how search-augmented LLMs handle source material, but a couple of technical distinctions are worth clarifying on the billing and architecture side.
First, the token consumption you noticed isn't a byproduct of the search crawl itself. Web indexing and keyword retrieval operate on lightweight database lookups that cost virtually zero LLM tokens. The reason your balance took a hit is entirely due to inference and context ingestion. When an AI agent pulls external sources, it has to serialize those chunks directly into the prompt context window, forcing the model to perform cross-attention over thousands of tokens just to synthesize an answer. You are essentially paying for the compute cycles required to reason over the text, not a fee for your own data.
Second, proposing HTTP 402 micropayments or mandatory top-level branding for conversational interfaces sounds clean on paper, but it fundamentally breaks how synthesis engines work. Generative UIs aggregate dozens of snippets to build a unified response; forcing dynamic ad placements or cryptographic payment handshakes per cited source introduces massive latency overhead and ruins the deterministic UX flow.
If we want true creator monetization without choking agentic workflows, the solution isn't patching legacy HTTP status codes. It requires building verifiable content credentials directly into cryptographic embedding spaces at the indexing layer.
Speaking of which, since you are deep into indexing architectures, how are you currently handling semantic drift and vector collision when routing multi-source documents into local retrieval pipelines? Keeping precision high without blowing up latency is a proper bottleneck.
Solid write-up. Keep building.
Thanks for the clarification — your technical points are accurate in their place, but they answer a narrower question than you seem to think, and they sidestep the deeper problem.
First, regarding tokens and billing: It's true that the cost of crawling and indexing is virtually nil at the token level, and that the bill comes from inference and context ingestion. But this strengthens the argument rather than weakening it: you explicitly acknowledge that the value of the source is consumed inside the context window to generate the answer, and then the revenue goes entirely to the model operator. In other words, the work the author produced is converted into a "compute cost" paid to the platform, while the author remains outside the equation. The distinction between a "data fee" and an "inference fee" is an accounting distinction, not a distinction in how value is distributed. The problem isn't what the invoice is called — it's who collects it.
Second, regarding HTTP 402 and branding: You're right that dynamic ad injection and synchronous payment handshakes per source destroy latency and ruin the determinism of the experience. But that's a critique of the naive implementation, not of the solution itself. Blogs don't display an ad with every quotation; licensing is managed centrally, outside the response path. Likewise, source rights could be settled after the answer is generated or at the indexing layer, with no impact on response time. So you're judging the idea by its worst-case implementation.
Third — and this is the crux: Your proposal of "content credentials inside cryptographic embedding spaces" sounds like a clean alternative, but it reproduces the very same problem in technical clothing. Who owns the indexing and embedding layer? The same platforms that own inference. Building verification at the embedding level means handing the mechanism of proof and collection to the very party we're supposed to be constraining. Cryptography doesn't solve the question of bargaining power; it may just give it a veneer of technical legitimacy.
And the question you didn't answer: If platforms reject even simple branding on the grounds of latency and experience, what makes us trust that they'll implement "verifiable credentials" carrying real financial obligations toward authors? A technical solution without an enforcement mechanism and aligned incentives remains architectural decoration.
As for your question about semantic drift and vector collision: It's a fair question, but it assumes the bottleneck is purely technical. In reality, the biggest source of semantic drift in retrieval pipelines isn't the vectors — it's the absence of the source's original context, which is precisely what the "inference instead of licensing" economy creates. We're tuning precision inside a system whose inputs we've deliberately corrupted.