Trust Wins: The Thesis at the Center of AI Optimization

The thesis, stated plainly

AIOFacts uses a working definition of AI Optimization: the practice of improving how clearly AI-powered systems can access, understand, verify, and represent an entity, its information, and its evidence. That definition is ours, not an official or industry-agreed one, and the terminology in this field remains unsettled. GEO, AEO, LLMO and AIO are all in active use, by different people, meaning overlapping but not identical things.

Inside that definition sits a compressible claim: trust wins. Stated more carefully, the AIOFacts position is that when an AI-powered system assembles an answer from retrieved material, the material it can locate, parse, corroborate, and attribute has a practical advantage over material it cannot, and that this advantage is structural rather than rhetorical. Persuasion is written for a reader who can be persuaded. Verifiability is written for a process that checks.

This is a position, and it is labelled as one throughout. It is not a description of how any model works internally. No one outside a model's operator can observe how that system weighs anything, and this article does not claim to.

Why verification became the visible constraint

The reason the thesis is worth stating now is a change in how consumer AI products answer questions. A system generating purely from model weights has no external record to check against. A system that retrieves documents at answer time does, and it exposes the result of that retrieval to the user as a citation. Anthropic's documentation for the Claude web search tool states that citations are always enabled for web search, and that each web search result location carries citation information. Google's Search Central documentation describes AI Overviews as a way for people to get to the gist of a topic quickly while providing a jumping off point to explore links.

Once an answer carries links, the correctness of those links becomes measurable by outsiders. That is a meaningful shift. It means the field acquired an evidence base that does not depend on any operator's cooperation, because anyone can read a citation and check whether it says what the answer claims it says.

What that evidence base shows is that attribution has been difficult in practice. Liu, Zhang and Liang audited four generative search engines and reported that on average only 51.5 percent of generated sentences were fully supported by their citations, and only 74.5 percent of citations supported their associated sentence. The Tow Center for Digital Journalism ran 1600 queries across eight AI search tools in early 2025 and reported that the tools failed to return the correct source more than 60 percent of the time, with the strongest performer in that test still incorrect on roughly 37 percent of queries. Both studies are snapshots of named products on specific dates. Platform behavior varies and these systems change frequently, so neither figure should be read as a current or permanent property of AI search.

The mechanisms that are publicly observable

The honest version of AI Optimization concerns itself with the parts of this system that are documented rather than the parts that are inferred. Several are documented in detail, by the operators themselves.

Access is governed by an actual standard. RFC 9309 specifies the Robots Exclusion Protocol, and operators map distinct user agents onto distinct functions. OpenAI documents three: GPTBot, which it describes as crawling content that may be used in training its foundation models; OAI-SearchBot, which it states is used to surface websites in search results in ChatGPT's search features; and ChatGPT-User, for user-initiated fetches. OpenAI's documentation notes that sites disallowing OAI-SearchBot will not appear in ChatGPT search answers, although they may still be shown as navigational links.

That separation matters more than it first appears. A publisher deciding whether to permit training and a publisher deciding whether to be retrievable are making two different decisions, and until an operator documents the split, those decisions are easy to conflate and make once by accident.

• Crawler access, governed by robots.txt under RFC 9309, with separate agents for training, retrieval, and user-triggered fetches
• Machine-readable identity, expressed through published vocabularies such as schema.org Organization, which supply names, identifiers, and relationships in a parseable form
• Citation surfaces, where retrieved sources are returned alongside the generated text and can be inspected by anyone
• Content provenance, addressed by the C2PA specification through cryptographically signed manifests that record the source and edit history of a media asset

What the operators say, including the parts that cut against the field

A reference has an obligation to record documentation that is inconvenient for the practice it documents. Google's guide to optimizing for generative AI features is the clearest current example. It states that best practices for SEO continue to be relevant because its generative AI features are rooted in its core Search ranking and quality systems. It states that structured data is not required for generative AI search and that there is no special schema.org markup a site needs to add. It advises against creating files such as llms.txt, against chunking content into small pieces for machine consumption, and against rewriting content specifically for AI systems.

Much of what circulates under the AIO, GEO and AEO labels asserts the opposite of each of those points. AIOFacts records the documentation rather than the assertion. Where a platform operator publishes guidance about its own products, that guidance is the strongest available evidence about those products.

Two limits on that evidence should be stated with it. First, it describes Google's features only. It says nothing about ChatGPT search, Claude, Perplexity, or Copilot, which are separate systems built by separate organizations with separate retrieval stacks. Second, guidance describes intent and current design, and neither is the same as measured behavior. The correct reading is narrow: on Google's surfaces, Google states that no distinct optimization discipline is required. That is a real and load-bearing fact, and it is not a claim about the whole field.

Trust as a property of the record, not of the audience

The thesis becomes more useful when trust is treated as an attribute of a body of evidence rather than a feeling held by a person. In that reading, an entity is trustworthy to a machine process to the degree that its claims are checkable: stated in one place, consistent with the same claims stated elsewhere, dated, attributed, and reachable.

Two independent standards efforts point the same direction. The C2PA specification builds provenance into the asset itself, encoding origin and edit history into a signed manifest so that authenticity can be established without trusting the party presenting it. The NIST AI Risk Management Framework, published as NIST AI 100-1 in January 2023, lists the characteristics of trustworthy AI systems as valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Neither document is about marketing. Both treat trust as something demonstrated through documentation and verifiable structure.

The practical consequence for an entity is unglamorous. An organization whose basic facts are stated inconsistently across the properties it controls is expensive to represent correctly, and the available evidence indicates that systems assembling answers under time and token constraints do not reliably resolve that kind of ambiguity in the entity's favor. The Tow Center findings show what happens when resolution fails: a confident answer, produced anyway, frequently attached to the wrong source.

What follows in practice, in bounded terms

If the thesis holds, the implied work is mostly editorial and structural rather than technical, and none of it is exotic. What follows is offered as reasoning from documented mechanisms, not as a set of tactics with measured effects.

• State the entity's core facts identically across every property under the entity's control, because inconsistency is the ambiguity a retrieval process cannot resolve for you
• Attach a source and a date to every factual claim, so that a claim can be corroborated by something other than its own confidence
• Publish identity in a machine-readable form using an established vocabulary such as schema.org Organization, while noting that Google states this is not required for its generative features
• Decide crawler access deliberately, distinguishing training access from retrieval access, since RFC 9309 and the operators' own documentation treat them separately
• Maintain a correction path, so that an incorrect claim can be withdrawn from the record rather than left to be retrieved indefinitely
• Prefer content that could not be assembled from other sources, which Google's guidance describes as non-commodity content carrying a unique perspective

How this thesis could turn out to be wrong

A position that cannot fail is not a position. Several plausible developments would weaken or invalidate the claim that verifiability confers durable advantage, and AIOFacts would rather name them than wait for them.

Retrieval could recede. If model operators shift weight back toward parametric recall, or toward a small set of licensed corpora acquired through commercial agreements, then the checkability of a public document would matter less than whether its publisher signed a contract. Licensing arrangements between AI companies and publishers already exist and are not visible to outside measurement, so this cannot be ruled out from the outside.

Verification could become cheap enough to stop being a differentiator. If retrieval and cross-checking improve to the point that a system reliably reconstructs a correct account of an entity from a poorly organized record, the advantage described here would compress toward zero. The 2023 and 2025 citation studies describe a gap; a gap that closes stops being leverage.

And familiarity could simply dominate. It is possible that entity recognition in these systems tracks scale of mention more than quality of evidence, in which case the honest advice would be about volume, not verifiability. AIOFacts cannot distinguish those two explanations from outside a model, and does not claim to.

Measurement, and why the numbers are hard to hold

Anyone attempting to test this thesis rigorously runs into the same four obstacles, and they are worth recording because they bound every number published in this field, including the ones cited above.

Generation is not deterministic, so the same prompt can produce different sources on different runs. Answers may be personalized or regionalized, so two observers see different results and neither is wrong. Retrieval indexes and model versions change without notice, which means a measurement has a shelf life that is rarely stated alongside it. And no operator publishes a ranking or selection API for these surfaces, so there is no ground truth to compare an observation against.

The consequence is that credible measurement in AI Optimization looks like repeated sampling with dates attached and variance reported, not a score. Any single figure describing how an AI system treats an entity is a reading taken at a moment, and should be published as one. This applies to the AIOFacts working framework as much as to anyone else's.