Content Trust, Ten Years Later

On 19 December 2016 I published an essay asking whether fake-news classification would create a new ranking factor: content trust. Facebook was partnering with Snopes. Sites like The Onion made the parody case obvious. Propaganda farms made the opposite case obvious. The hard case was the publisher that ran some careful work and amplified some junk, or the writer who was wrong without trying to mislead anyone.

I thought search would eventually put a number on that. Site reputation was already on every ranking-factor list. Social shares were being treated as a quality signal even though, as I wrote then, people share what they have not read. A content-trust score looked like the next simple answer. “When all things are equal,” I wrote, “search engines need to assume that people are seeking true information first and foremost. Truth over satire.”

Ten years later the function arrived. The score did not. What arrived is not a way of knowing whether a claim is true. It is a way of retrieving a claim, attaching a source, and folding it into an answer. The problem I was pointing at got more urgent, not more solved. Search no longer only decides which sources we encounter. It increasingly decides which information becomes part of an answer.

The 2016 argument still holds as a diagnosis of its moment. Ranking by links and shares could not tell a hoax from a satire, or a satire from an opinion, or an opinion from a fabrication written to look like reporting. PageRank was a genuine advance because it used other people’s citations instead of a site’s self-description. It still could not tell you whether the cited page was true. A visible trust score seemed like the fix that would fit products that had to stay simple, fast, and predictable.

Parts of that work did get built. Quality systems, spam filters, and publisher inclusion rules all do some of what I was pointing at. Experience, Expertise, Authoritativeness, and Trustworthiness is a bundle of judgments, not a ranking factor, and rater scores do not determine where a page sits. None of it is a public mark on a document, and none of it fixes the unit. A careful publisher can run a bad assertion. An unknown site can hold the one primary document. A parody page can be honest about what it is and still outrank a dry correction. Domain-level ratings treat trust as a property of a website. A single number on a domain cannot resolve those conflicts. It can only hide them.

The unit moved

The more useful question now is where the trust boundary sits. For a long time the convenient unit was the site, then the page. In February 2021 Google launched Passage Ranking. Martin Splitt described it as scoring different parts of a page independently: a page might be about gardening in general and still have one passage that is the best match for a specific tomato. That was not a new index of free-floating claims. It was ranking catching up to a fact writers already knew. Most pages are not one claim. The later move into answer engines is a conceptual continuation of that pressure, not the same mechanism running on a smaller object.

Google’s generative features in Search use retrieval-augmented generation, which Google also calls grounding: core ranking systems retrieve pages from the index, then “review the specific information from those retrieved pages” to produce a response, with links back to the supporting pages. Microsoft’s Bing team, in a May 2026 post on the evolving role of the index, named the same split from the other side. Traditional search asks which pages a user should visit. Grounding asks what information an AI system can responsibly use to construct a response. The unit of value, they wrote, shifts from documents to “groundable information”: discrete, supportable facts with clear provenance.

I would not let a vendor own the thesis. Google and Microsoft are describing their own machinery. Google’s May 2026 guide treats answer-engine optimization as still SEO, rooted in core ranking and quality systems, and it still asks for unique, non-commodity content. Bing writes more directly about the index as support for answers. Those are different stresses on a shared transition, not two schools of search.

Search engines were never neutral directories. Ranking always influenced which voices we encountered. What changed is the distance between selection and assertion. The system no longer merely chooses what we might read. It increasingly composes what we are told. If ten sources exist and the retriever keeps three, those three largely define what the model is allowed to say.

Grounding solves a problem of traceability. It does not, by itself, settle truth. A passage can be relevant to the question, accurately attributed to its source, and still be wrong. Even when the claim is supported, another question remains: who stands behind it once the model has made it part of an answer? Microsoft’s word “supportable” hides a decision: who decides, by what standard, and what happens when the most often-repeated claim is wrong. A source cited repeatedly can be evidence of distribution rather than independent corroboration. Secondary pages that share the query’s vocabulary can outrank the drier primary document. Semantic consensus can end up standing in for independent corroboration.

The cost of a mistake changed with that. Bing’s post is blunt. An imperfect ranking is often recoverable: the user can open another result. An error in a synthesized answer can compound across reasoning steps, and the user may never see the conflict. Abstention, they argue, is a valid outcome when the evidence is missing, stale, or contradictory. That is an editorial standard, whether or not anyone at the company would use the word.

In May 2024, Melanie Mitchell searched “How many Muslim presidents has the US had?” Google’s AI Overview answered that the United States had had one: Barack Hussein Obama. The system cited a chapter from an academic book whose title asked the question and then argued against it. The source was real. The citation was visible. The passage matched the query. The answer was still wrong, because the model treated a title’s provocation as a fact. A 2026 Washington University study found that roughly one verifiable claim in nine lacked support in the cited material or contradicted it, although the researchers cautioned that incomplete source capture could inflate that estimate.

That is why provenance still matters and why it is not enough. In Mona Lisas and Readymades I argued that knowing how a text was manufactured is not the same as someone taking responsibility for what it claims. A perfectly authenticated passage can still be the wrong ingredient. The receipts, when they exist, are citations. The story is still the model’s.

From visibility to evidence

If you write for the open web, the practical target has moved, and it is smaller than a playbook. Search engine optimization asked how a page earns a position in a list. Answer-mediated discovery asks how a claim becomes eligible to be retrieved, attributed, and shown inside a synthesized response. Bing’s webmaster tools report citations and grounding queries, not only ranked positions. OpenAI’s help page on searching the web now tells users that search citations can be incomplete, outdated, or incorrect, and that they should open a cited source to check that it supports the answer. Google’s May 2026 guide says you do not need a special file, a chunking scheme, or a rewrite aimed at the model. That is right as a refusal of a new religion. It can still hide the change in object. The work is still to be crawlable, specific, and worth citing. The thing being selected is often a passage or a claim, not a homepage. A writer who wants a claim used as evidence can make it extractable, name the source, and leave a path back. That is older than generative search. It now matters in a new way: the claim may become part of an answer the reader never traces back to its source.

I imagined technology discovering trust and encoding it. Platforms also govern it. They decide who counts as an authority, which sources are eligible for retrieval, how citations appear, and when a safety rule overrides a relevant page. Those are editorial decisions implemented in software. Ranking still exists. Grounding, as Bing was careful to say, does not replace it. Grounding adds a second job: decide not only what to show, but whether the evidence is good enough to answer at all.

A publisher can make a claim checkable. A ranking system can prefer sources with a history of care. A retrieval model can choose the fragments that enter the context. A generating model can refuse to answer. A reader can still believe a fluent sentence without opening a citation. Each of those is a real place for trust to fail. None of them is a score you can publish next to a domain name. Identifying everyone who contributed is not the same as identifying who is accountable for what the answer asserts. When that job is split across a publisher, a retriever, a generator, and a reader, responsibility can become harder to locate, even as the answer appears more authoritative. The citations document where the information came from. They do not establish who stands behind the finished claim.

Madam I’m Adam

This continues from Mona Lisas and Readymades, where provenance was not authorship, and The Merge Tax, where a clean story can erase the path. This one follows what ranking became once answers replaced lists.