The Web Has More Sources Than It Has Witnesses

Five different websites say the same thing. A search engine counts five sources. An AI answer engine cites five domains. A trust system logs five corroborating signals. But what if all five got their information from the same press release?

Then you have five publications and one witness. Publisher multiplicity is not evidentiary multiplicity. The web is very good at multiplying pages. That does not mean it is multiplying evidence.

The Source-Counting Problem

When multiple domains assert the same fact, the natural reading is that consensus exists. The more domains, the stronger the apparent consensus. This is how humans interpret corroboration in everyday experience: several independent people saying the same thing is stronger evidence than one person saying it.

The web imports that intuition and breaks it. A single corporate announcement can propagate through a wire service to dozens of publishers in hours. A government regulatory decision can appear on multiple official domains, several news sites, and a handful of legal analysis newsletters by end of day. Each propagation step adds a URL. None of them necessarily adds an independent witness.

Systems that count domains, citations, or URLs as proxies for independent corroboration can substantially overstate the breadth of consensus behind a factual assertion.

What AATL Investigated

AATL investigated factual assertions that appeared to be corroborated across multiple web publishers. The research question was: when an assertion has multi-source corroboration on the web, how much of that apparent corroboration survives when the provenance of those sources is resolved?

The unit of analysis was not a webpage or a domain. It was an entity paired with a specific factual assertion. The research separated two questions that are often conflated: is the assertion supported, and how many genuinely independent evidentiary origins support it? Those are not the same question. An assertion can be extremely well supported while having only one evidentiary origin.

Provenance Compression

The phenomenon we observed and named is provenance compression: when multiple apparent sources supporting an assertion resolve, after tracing their provenance, to fewer underlying evidentiary origins.

The user sees ten sources. A naive consensus system sees ten sources. Evidentially, there may be only one witness. This does not mean the underlying assertion is false. Collapsed consensus is not equivalent to false information. It means the apparent breadth of independent corroboration did not survive provenance analysis.

Case One — Henry Schein

The assertion appeared across multiple publishing domains. Provenance tracing showed the material was distributed through Business Wire. One publisher explicitly identified the content as originating from Henry Schein via Business Wire. Multiple domains did not establish multiple independent evidentiary origins. They established distribution.

Domain count measured propagation, not corroboration. One statement traveled through the web and acquired the visual appearance of consensus without gaining another independent witness.

Case Two — Naspers / Vastu Housing Finance

The assertion appeared across numerous domains including government publications, financial news publishers, and legal analysis sites. The underlying fact was exceptionally well documented. The Competition Commission of India published both its announcement and the detailed regulatory order.

But those are two documents from the same authoritative regulatory origin. The Press Information Bureau distributed the CCI action rather than independently witnessing it. Financial news coverage traced back to the regulator and wire reporting. Legal newsletters independently analyzed the decision — but independent analysis of an authoritative decision is not an independent evidentiary origin for the underlying regulatory act.

This case illustrates a specific pattern: government-domain multiplicity does not necessarily mean governmental-source independence. Two government websites can still represent one underlying governmental origin. There was little reason to doubt the CCI approval occurred. The collapse concerned consensus breadth, not truth confidence.

Case Three — Sekisui House / M.D.C. Holdings

Numerous publications carried the transaction announcement. Some collapsed into the Sekisui House announcement and distribution chain. But M.D.C. Holdings also recorded completion in its own regulatory filing. Apparent source multiplicity compressed substantially — but it did not compress to a single evidentiary origin. At least two origins remained: Sekisui House and M.D.C. Holdings independently.

Provenance resolution does not automatically destroy consensus. Sometimes ten apparent sources become two genuine origins. And two independent origins can constitute real corroboration. Provenance analysis is not a mechanism designed to make consensus disappear. It is a mechanism for measuring consensus more accurately.

Hosting Is Not Provenance

One of the clearest lessons from this research: the domain hosting a document does not necessarily identify the evidentiary origin of that document.

A company filing hosted on SEC.gov remains fundamentally a company-origin disclosure. An issuer press release appearing on an exchange website does not become exchange-origin evidence. A government publication redistributing another regulator's decision does not constitute another governmental witness. Therefore: URL domain is not publisher, and publisher is not evidentiary origin. Those concepts need to be separated before any consensus measurement can be trusted.

Depth vs. Independence

Several documents from the same authoritative organization can make evidence much stronger without making it more independent. A regulator's press release, regulatory order, and database record may provide extremely deep confirmation of a single fact. They still constitute one evidentiary origin. Evidence can become deeper without becoming broader. Depth of support and independence of support are separate dimensions and should be modeled as such.

A system should not collapse all four dimensions into a count of URLs.

Why This Matters for AI and AEO

AI systems frequently present multiple citations as a visible signal of confidence. Users naturally interpret several citations from different websites as independent corroboration. A citation architecture that does not resolve provenance confuses publication diversity with evidentiary independence.

An AI system could cite five different domains while all five derive the claim from one corporate announcement. The answer might still be correct. But describing that evidence as five-source consensus would overstate what the evidence actually establishes. The exact proposition matters too. A regulator approving an acquisition is not the same assertion as the companies completing it. A construction permit is not an operating license. Consensus belongs to a precise proposition, not merely to an entity name.

Research Limitations

What This Research Establishes

Multiple publishing domains do not necessarily represent multiple independent evidentiary origins. Apparent consensus can compress when source provenance is resolved. Provenance compression can occur even when the underlying assertion is true and strongly supported. Independent corroboration can also survive provenance resolution — as the Sekisui case demonstrates. Support strength and consensus breadth are different dimensions. Domain count alone is insufficient for measuring evidentiary independence. Hosting location does not reliably establish provenance.

The next generation of AI trust measurement cannot stop at asking how many sources say something. It has to ask the harder question: how many sources independently know it?