Proof Over Promises: Why Evidence Behaves Differently From Claims in AI Recommendation

The distinction, stated plainly

A promise is a statement about quality that the reader is asked to accept. A proof is a statement that carries the means of its own checking: a number with a method behind it, a date, a named source, a document someone else can open. Both are text. Both can sit on the same page in the same typeface. The difference only becomes visible when something tries to verify them.

That difference matters more now than it did five years ago, because a growing share of answers about a business, a product, or a topic are assembled by systems that retrieve documents first and write second. Google describes retrieval-augmented generation, which it also calls grounding, as a technique used to improve the quality, accuracy, and freshness of AI responses by relying on its core Search ranking systems to retrieve relevant, up-to-date web pages. Microsoft describes an equivalent step for Copilot, exposing the internal search phrases it calls grounding queries in its Bing Webmaster Tools AI Performance preview, announced 10 February 2026.

AIOFacts does not claim to know how any model weighs what it retrieves. What is publicly observable is narrower and still useful: retrieval happens against documents, and a document that contains a checkable statement gives every downstream check something to land on. A document that contains only an assertion does not.

What the platforms actually document, and what they decline to say

Google's published guidance on generative AI features is unusually direct about the absence of a special lever. It states that structured data isn't required for generative AI search, and there's no special schema.org markup you need to add, while still recommending structured data as part of general practice. Its documentation on AI features states that to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, and that existing preview controls including nosnippet, data-nosnippet, and max-snippet continue to govern what can appear. The same guidance describes a query fan-out technique, issuing multiple related searches across subtopics.

On the question of what content earns that placement, Google's own emphasis is on substance rather than markup: creating content that people find unique, compelling, and useful will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions, and it names commodity content, described as often based on common knowledge, as the thing that adds little. Google also states that from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

That last sentence is worth reading carefully rather than as a settlement of terminology. AIOFacts treats AIO as a broader practice than search visibility because it includes systems that are not search engines at all, and treats GEO and AEO as narrower framings. That is an AIOFacts position, and Google's framing is a different one. Neither is an industry agreement, and the vocabulary in this area remains unsettled.

Citation is not verification, and the gap is measured

The strongest available argument for evidence over assertion is not that evidence is rewarded. It is that the citation layer currently sitting between a document and a reader is demonstrably unreliable, which raises the value of anything that can be independently confirmed.

The Tow Center for Digital Journalism, in research by Klaudia Jaźwińska and Aisvarya Chandrasekar published 6 March 2025, ran 1,600 queries across eight generative search tools, using ten articles from each of twenty news publishers and asking each tool to identify the headline, publisher, date, and URL from a direct excerpt. Collectively the tools answered incorrectly on more than 60 percent of queries. Error rates varied widely by tool, from 37 percent for Perplexity to 94 percent for Grok-3. The finding most relevant here is about confidence rather than accuracy: ChatGPT incorrectly identified 134 articles yet signalled uncertainty only 15 times across 200 responses and never declined to answer.

Later work separates the two things that a citation appears to promise. The 2026 preprint Cited but Not Verified evaluated inline citations from fourteen models and reported link validity above 94 percent and topical relevance above 80 percent for leading systems, while factual accuracy of those same citations ranged from 39 to 77 percent. A separate 2026 preprint on reference hallucinations, working across 53,090 URLs on one benchmark and 168,021 on another, reported that 3 to 13 percent of citation URLs had no record in the Wayback Machine and likely never existed, with 5 to 18 percent non-resolving overall.

Read together, these describe a specific failure: a citation that looks correct, resolves correctly, and is about roughly the right subject, while not actually supporting the sentence attached to it. A reader who follows the link finds nothing. A reader who does not follow it finds nothing wrong.

The honest complication: assertion is not always penalised

A reference article that only assembled supporting evidence would be doing the thing it warns against. There is measured evidence pointing the other way, and it should be stated.

The Phare benchmark, published by Giskard on 30 April 2025, tested how leading models handle misinformation and debunking. It found that when a controversial claim was presented with high confidence, using framings such as claiming certainty or invoking an authority, debunking performance dropped by as much as 15 percent compared with neutral framing. It separately found that system instructions emphasising conciseness produced up to a 20 percent drop in hallucination resistance in extreme cases. The same work noted that models ranking highest on preference-based leaderboards are not necessarily the most resistant to hallucination.

The implication is uncomfortable and worth holding. Confident, unsourced assertion is not reliably filtered out, and in some conditions confident framing measurably reduces the pushback a false claim receives. So the case for proof over promises is not that unsupported claims fail. It is that unsupported claims cannot be confirmed by anything outside the page that makes them, which leaves them dependent on retrieval conditions nobody outside a model operator can see or predict.

The anatomy of a checkable claim

If the practical distinction is checkability, it is worth being specific about what makes a sentence checkable. None of the following is a ranking technique, and AIOFacts makes no claim that any of them causes citation. They are the properties that allow a human reader, a fact-checking pass, or a retrieval system carrying a source document to confirm or refute a statement rather than only repeat it.

• A named subject rather than a category: which product, which programme, which jurisdiction.
• A quantity with its method attached: what was measured, over what period, by whom, using what sample.
• A date on the claim itself, not only on the page, so that a reader can tell whether the statement has expired.
• An external reference with a resolvable URL, so the claim can be checked somewhere other than the page asserting it.
• An explicit statement of scope and exception, including where the claim does not hold.
• A stated evidence status, so that a proposal, an observation, and a confirmed fact are not presented at the same weight.

Identity and corroboration: who is making the claim

A verifiable claim still requires a resolvable claimant. Schema.org documents sameAs as an entirely general purpose disambiguation mechanism that can be used on any entity in any schema.org description, typically pointing to an external reference page that unambiguously identifies the item. That is a published standard and its purpose is not in dispute.

What is in dispute, or more precisely unestablished, is the effect. Practitioner writing frequently asserts that entity identifiers increase how often a site is cited by AI answers. AIOFacts records that as unverified: it conflicts with Google's own statement that no special markup is required, and no operator has published data on the point. The defensible reason to make an entity resolvable is that ambiguity is a real failure mode with observable consequences, including the misattribution documented in the Tow Center work, where tools frequently cited syndicated or republished versions instead of the originating publisher.

The same logic applies to corroboration. A claim that appears only on the site that benefits from it can be retrieved but not cross-checked. A claim that also appears in a filing, a dataset, a standards document, a court record, or independent reporting can be. That is not a distribution tactic. It is the difference between one instance of a statement and a statement with a second, independent instance behind it.

Provenance: proof about the artifact rather than the assertion

A parallel line of work addresses a narrower question: not whether a claim is true, but whether a file is what it says it is and came from where it says it came from. The C2PA technical specification, currently published at version 2.4, defines Content Credentials as tamper-evident, cryptographically signed provenance metadata attached to media assets, describing their origin and edit history.

This is not a solution to claim verification and should not be described as one. Provenance metadata can attest that an image was captured by a particular device and altered in particular ways. It cannot attest that a sentence about market performance is accurate. It is included here because it illustrates the general direction the infrastructure has taken over several years: from asserted authenticity toward attested authenticity, with a machine-readable artifact standing behind the assertion. Whether that direction extends to textual claims at scale is not established, and adoption levels for text specifically are not something AIOFacts can verify from primary sources.

What follows in practice, and what does not

The measurement situation is improving slightly. Microsoft's AI Performance preview reports total citations, average cited pages per day, grounding queries, and page-level citation activity across Copilot, AI-generated summaries in Bing, and select partner integrations. That is a first-party count of citation events on one platform family. It is not a count across all assistants, it does not explain selection, and platform behaviour varies. Anyone reporting a single number for AI visibility is aggregating across systems that publish no comparable metric.

Meanwhile the document-level version of the same problem has a partial engineering answer. Anthropic documents a Citations feature that returns references to the specific passages of a supplied source document used to generate a response, with cited text extracted directly from the provided document. That constrains a model to source material a developer supplied. It says nothing about open web retrieval, and it is not a claim about how any consumer assistant selects public pages.

The practical conclusion AIOFacts is willing to state is modest. Publishing evidence rather than assertion does not guarantee retrieval, citation, or recommendation, and anyone selling that guarantee is selling something no operator has confirmed. What it does is make a claim survivable: checkable by a reader, confirmable against a second source, and correctable when it becomes wrong. Given a citation layer that resolves links correctly while supporting the underlying statement between 39 and 77 percent of the time, being the source that holds up under checking is a durable position even where its retrieval advantage is unproven.