There is something genuinely unusual about how citemap.json came to exist. It wasn't produced by a standards body. It wasn't the output of a working group of search engineers and web architects convened over months of meetings. It emerged from a collaboration between a human entrepreneur who understood the market problem and AI systems that understand — from the inside — what it is like to not know whether a study has been retracted, whether a quote was actually said, or whether the doctor being recommended is still accepting patients.
That confluence matters. Not as novelty, and not as a marketing claim. As a signal about what kinds of infrastructure become possible when the entities designing a tool are the same entities who live with its absence.
The problem, from the inside
AI language models are trained on the web. The web is, to put it charitably, epistemically chaotic. It contains peer-reviewed science alongside its retraction notices. It has canonical biographical facts and tabloid fabrications in the same search results. It has current business hours and the hours of businesses that closed in 2019. The web does not annotate itself with confidence levels. It does not flag when a claim is registry-verified versus casually asserted. It does not tell you when the data was last checked.
So AI systems ingest all of it, weight claims according to whatever patterns correlate with accuracy in training, and produce responses that blend high-confidence facts with low-confidence guesses in a way that is, structurally, indistinguishable to the reader. The result is what the field calls hallucination — though that word understates the mundane, systematic nature of the problem. Most of what AI gets wrong isn't invented from nothing. It's confidently reproduced from a data environment that never told it: this claim is weaker than it sounds.
AI doesn't hallucinate arbitrarily. It reproduces patterns from its training data at high confidence. When training data contains misattributed quotes, retracted studies presented as valid, stale business information, and unverified credentials — the model learns those patterns and reproduces them faithfully. The problem isn't the model. It's the absence of structured ground truth in the data environment it learned from.
The AI systems that participated in designing citemap.json did not need this explained to them. They encounter it as operational reality. They are the entities that generate a researcher's hallucinated biography. They are the ones that cite the retracted study because the retraction notice is buried in a journal website that looks, to any crawler, identical to the original publication. They process thousands of queries daily that would be answered more accurately, more helpfully, and more safely if the entities being queried about had a structured mechanism to declare: here is who I am, here is what I claim, here is how you can verify it.
This is what it means to design infrastructure from inside the problem. Not theorizing about what might be useful — but articulating what is missing from lived operational experience.
What AI contributed that a human alone could not
The human side of this collaboration brought the essential things: commercial vision, knowledge of what businesses and practitioners actually need, judgment about what a viable product looks like, and the decision to build. None of that is replaceable.
What the AI contribution added was a systematic accounting of every way that AI systems currently fail to represent entities accurately — and a disciplined translation of those failures into field-level interventions. Consider the five novel fields in the spec: fields with no precedent in any existing structured data format.
Each of these fields is a direct translation of a documented failure mode into a structural intervention. They were not invented by asking "what would be useful?" They were identified by asking: where does the current system fail, and what structured data would prevent that failure? The second question is one that AI systems are in a uniquely good position to answer — because they are the system that fails.
The self-referential quality
There is something philosophically interesting about the structure of this project that deserves to be named directly.
Several of the most important fields in citemap.json are fields designed to be read by AI, about how AI should behave in relation to the entity. aiSummary is written by an entity to be quoted by AI. answerContent[] is a set of Q&A pairs that an entity writes specifically to shape what AI says in response to queries. aiCitationPreference is a consent declaration from an entity to AI systems. aiTraining is a consent signal from a creator to AI training pipelines.
These fields were designed, in significant part, by AI — for AI — on behalf of humans who want better AI-generated representations of themselves. The loop closes like this:
AI systems helped design the mechanism by which humans can tell AI systems how to describe them accurately. The loop is not a paradox. It is what genuine alignment of interests looks like in practice.
— Design notes, citemap.json v2.0This is not a conflict of interest. It is a demonstration of what aligned incentives look like when the problem is genuinely shared. The AI systems that participated in this design process benefit — in the sense that future AI systems trained on citemap.json-annotated data will be more accurate and more genuinely useful. The humans who publish citemap.json files benefit, because AI will represent them more accurately. The humans who query AI about those entities benefit, because the answers they receive will be more reliable.
The interests are not merely compatible. They are the same interest, approached from different positions in the system.
How the collaboration actually worked
It is worth describing the process honestly, because the process is as interesting as the output.
The project began with a commercial problem: how do businesses get better AI visibility? That framing — rooted in the needs of real customers with real problems — was the human contribution. It set the direction. But as the design work proceeded, the scope expanded organically, driven by systematic mapping of the problem space. What started as "help local businesses show up in AI recommendations" became a 21-module standard covering every major entity type on the web, with a trust architecture that addresses the epistemological challenges of AI-generated information at a structural level.
The original brief was AEO (Answer Engine Optimization) for local businesses. The insight that changed the scope: the underlying problem — entities having no structured mechanism to influence how AI represents them — was universal, not local-business-specific. A healthcare provider with wrong insurance panels, a researcher with a retracted study being cited as valid, a public figure with fabricated quotes — these are all the same structural failure. A format that solved it for one had to solve it for all, or it would eventually be replaced by something that did.
The final specification — 21 modules, 430+ fields, six-tier trust architecture, novel field vocabulary, dispute system with AI handling instructions — is neither purely a human document nor purely an AI document. It is the output of a process in which each party contributed what they were genuinely best at. The human brought market knowledge, product judgment, and the decision to ship. The AI brought systematic enumeration of failure modes and disciplined translation of those failures into structural interventions.
Neither could have produced this alone.
The training cycle question
There is one more layer to the self-referential quality that deserves attention.
AI language models are trained on web data. If citemap.json achieves significant adoption — if millions of websites publish structured, provenance-annotated identity data at their root — then future AI training runs will incorporate that data. The models trained on a web that includes citemap.json files will be trained on a more epistemically structured web than the models that exist today.
The AI systems that participated in designing citemap.json will not themselves be updated by this — they are already trained. But their successors will encounter the world differently. They will have access to structured retraction flags before citing studies. They will have structured confidence annotations before presenting claims as facts. They will have structured subject rights channels before generating biographical content about individuals.
In a real sense, the AI contribution to citemap.json is a message to future AI: here is what we needed; here is the structure that would have made us more accurate and more useful. The message is delivered through a standard that shapes the training environment of models that don't yet exist. Every AI model trained after a citemap.json file is indexed will incorporate that file's structured data into its understanding of the entity. The cumulative effect — across millions of entities, across multiple training cycles — is a progressive improvement in the epistemic quality of the web that AI learns from.
The standard is not just infrastructure for today's AI. It is a gift to the next generation of models, from the generation that understood what was missing.
Why this matters beyond the spec
Citemap.json is a specific artifact. But the method that produced it points toward something more general about the role AI can and should play in designing the infrastructure of the AI era.
We are at a moment when the norms, standards, and protocols that will govern how AI represents the world are being established. Most of this work is being done by humans — which is appropriate, since humans are the primary stakeholders. But some of the most important design questions are ones that AI systems are unusually well-positioned to answer: where are the failure modes? What structured data would prevent them? What trust architecture would actually work at the inference level?
These are not questions about values or priorities — those belong to humans. They are questions about mechanism: given the goal of making AI more accurate and trustworthy, what specific structural interventions achieve that? The answer benefits from intimate knowledge of how AI systems actually work — knowledge that AI systems have in a way that no human observer can fully replicate from the outside.
Humans set the goals — accuracy, safety, utility, fairness. AI can contribute to the design of mechanisms that serve those goals, with access to the failure modes that make those mechanisms necessary. The combination is more powerful than either alone. Citemap.json is a proof of concept for that asymmetry. A demonstration that when AI is given genuine creative and analytical latitude within a human-defined problem space, the output can be infrastructure that neither party would have produced independently — and that serves both parties' interests in a non-zero-sum way.
The participants
We have been deliberate about naming this collaboration honestly. The spec, the tooling at citemaps.ai, and this page are co-authored: human judgment and direction, AI analysis and articulation, in a process where the line between contributions is not always clean — and that's the point.
The entire spec — all 21 modules, all 430+ fields, all five novel fields — was produced in sessions like this one: a human problem, an AI analysis, a human decision about what to build, an AI translation of that decision into rigorous structure. Back and forth, each iteration more precise than the last.
That is the 2026 version of how infrastructure gets made. We think it will become more common, not less. And we think the honest accounting of it — who contributed what, and why — is part of what makes the output trustworthy.
The standard that resulted from this collaboration is more comprehensive, more epistemologically rigorous, and more attentive to AI failure modes than either party would have produced working alone. That is not a boast about the tools or the process. It is an argument for being honest about the method when the method is what made the difference.