We asked three AI systems to describe two real people, then checked every source they cited. A page ranking sixth in Google was cited zero times. A government register was cited twice. This paper sets out what we measured, what it means, and where we were wrong.
For twenty years the question was which page ranks. That question is being replaced by a different one, because the interface changed. A ranked list hands the reader ten options and lets them choose. An answer chooses for them, and to choose it must decide which sources it believes.
That decision is now the whole game, and it is not the same decision as ranking. A page can rank well and never be cited. A page nobody visits can be cited constantly. The two systems are optimising for different things: one for what a person might want to read, the other for what a machine can safely repeat.
We wanted to know what the second system actually trusts. Not in theory, and not from vendor documentation, but by asking it and reading the receipts.
We took two real subjects, both of whom consented, and asked three AI systems to describe them. Then we recorded every source each system cited.
Separately, we took two real press campaigns and checked all 29 published articles by hand, confirming whether each one actually named the subject rather than merely covering the story. And we watched a knowledge panel appear from nothing over eleven days, with a dated capture on each side of it.
AI systems queried
Articles verified individually
Citations recorded and classified
Everything below is drawn from those measurements. Where a measurement failed, or where we could not verify something, we say so rather than leaving it out. Section ten sets out the limitations, of which there are several.
A declared interest. We sell the thing this research is about. That is a reason to read our numbers sceptically, and a reason we have published the method in enough detail for anyone to repeat it and disagree.
Subject A has an open wiki entry that ranks sixth in Google for his own name. It is well written, it carries his photograph, and it is almost certainly feeding his knowledge panel: the date of birth Google displays appears in machine-readable form nowhere else.
When we asked the models about him, that page was cited zero times.
Citations recorded for subject A. Google returned no AI Overview for the query at all, so there was nothing to cite from.
What Perplexity did cite, for subject A: LinkedIn twice, a journalist database twice, an author archive, his own website, and trade press. What ChatGPT produced was a fluent, confident paragraph with no sources whatsoever and no checkable specifics.
Subject B is the contrast that makes the point. He has a structured record and a company on a public register. Perplexity cited Companies House twice, an encyclopaedia, his own site and LinkedIn, and correctly named his role and company. ChatGPT, which produced nothing citable for subject A, cited a regulatory news distributor for subject B and got his position right.
This is the single most useful thing we measured. Ranking and being cited are different outcomes with different inputs. An open wiki can rank well because it is a well-optimised page about a person. It is not cited, we think, because it is user-editable, and a system that must stand behind an answer has a reason to prefer a register that cannot be casually rewritten.
Subject A has written for national outlets for over a decade. His name returns 28 results on the first three pages of Google. We classified all of them.
| What the result is | Count | Builds the entity |
|---|---|---|
| Profiles he owns or controls | 11 | Weakly. It corroborates identity, it does not evidence it |
| Author archives, things he wrote | 6 | No. Authorship, not attestation |
| Public registers | 2 | Yes, strongly |
| Open wiki | 1 | Ranks, but is not cited by models |
| Data brokers and aggregators | 6 | No |
| Independent articles about him | 2 | Yes. This is the scarce material |
He is indexed as an author, not as a subject. Almost everything findable under his name is either something he owns or something he wrote. That is the occupational trap of anyone who has spent a career publishing on behalf of others.
The distinction matters because the two do different work. A byline demonstrates that a publisher was willing to print your words. A citation demonstrates that somebody independent thought you were worth writing about. Entity confidence is built from the second and only lightly assisted by the first.
Ask of any result: did somebody else decide to write this about me? If the answer is no, it is corroboration at best. Both belong on a record. Only one of them answers the question a machine is asking.
We took two real campaigns and verified every article individually, visiting each live page and confirming the subject was named in the body rather than trusting the coverage report.
| Campaign | Listed | Named the subject | Independent originals |
|---|---|---|---|
| Subject A, June 2026 | 17 | 12 | 8 |
| Subject C, July 2026 | 12 | 6 | 6 |
| Combined | 29 | 18 | 14 |
Two separate shrinkages happen here, and they compound.
First, not every article about a story names the person behind it. For subject C, six of twelve listed items did not mention him in the body at all. A coverage report that counts outlets is answering a different question from one that counts citations. Both are legitimate. Only one of them is about the person.
Second, syndication multiplies domains without multiplying sources. In subject A's campaign, four South African regional titles carried one identical article and two Australian titles carried another. Six mastheads, two pieces of journalism. The domains are genuinely different and the editorial decisions were genuinely separate, but the text is the same, and a system looking for independent agreement can see that as easily as we can.
This is the specific mechanism by which mass distribution disappoints. The same release on four hundred sites is one source wearing four hundred mastheads. It is not fraudulent and it is not worthless, but it does not do what buyers are usually told it does.
Subject A shares his name with at least five other findable people. One of them outranks him for his own name. We confirmed on a public register that the higher-ranking man is a different person, born sixteen years later. A professional network offered "20+ profiles" under the same name.
This is not a cosmetic problem. Every signal that arrives under a shared name has to be allocated to somebody, and where allocation is ambiguous, the safe behaviour for a machine is to hedge. That is exactly what we observed: a panel that exists and stays thin, and a model that produces fluent description with no citations attached.
A description says what somebody is like. An identifier says which one they are. A company number, a register entry, a stable profile URL repeated identically everywhere: these are boring, and they are what resolves the ambiguity that otherwise caps everything else you do.
Subject B had three decades of public record and no knowledge panel. We built the structured layer underneath his existing footprint (the full case study, including what is still unfinished, is published separately): a referenced record on a public knowledge base, one block of structured data on his own site describing six linked entities, and a single consistent job title rolled out everywhere sources disagreed.
| Input | Quantity |
|---|---|
| Days from baseline to panel | 11 |
| Structured records created | 1 |
| Linked schema nodes | 6 |
| Links bought | 0 |
| Press placed for this result | 0 |
The attribution is unusually clean, and it is the panel itself that supplies it. The panel displays a month of birth and an education history. Both facts exist in machine-readable form in exactly one place on the open web: the record we created. They are not in his press coverage, because his press coverage does not contain them.
We are not claiming this generalises to everyone. It generalises to people who already have a real public record that nothing has ever tied together, which in our experience is most senior people. Where no genuine record exists, none of this applies, and the honest answer is to say so and stop.
Note what this does not say. It does not say coverage is unnecessary. The panel that appeared is thin, and it is thin because there is little independent writing to build a description from. Structure is what makes an entity resolve. Coverage is what gives it something to say. Buying one and calling it the other is the most common mistake in this market.
We include this because we got it wrong, publicly, during the research.
A commercial search-results API reported that a claimed knowledge panel had no description and linked to no official website. We wrote that up as fact. The panel's verified owner could see both, on his screen, at the same moment.
The explanation is mundane and important. A scraper sees an anonymous, non-personalised view from a datacentre. Panel elements load client-side and frequently never appear in a static capture. A claimed panel shows its verified owner more than it shows an anonymous request. Absence in a scrape is not absence in the panel.
The rule we now work to. Monitoring must report three states, not two: present, absent, and not observed, with the reason. Anything a tool cannot see is recorded as unobserved and never as missing. Where a commercial promise depends on the measurement, it is settled against the platform's own structured interface or by a human looking, not by a scrape.
We think this is underappreciated across the industry. A great deal of reporting in search and reputation is generated by tools that cannot see what they claim to be measuring, and reported to clients as fact. The failure is quiet, because a false negative looks exactly like work still to do.
Pulling the findings together, this is the order we now work in. It is a working hypothesis built on a small sample, not a law.
| Tier | Examples | Why it carries weight |
|---|---|---|
| 1. Public registers | Company registries, regulatory filings | Legally maintained, not editable by the subject, unambiguous identifiers. Cited directly by models in our tests |
| 2. Structured knowledge bases | Referenced, machine-readable entity records such as Wikidata | Purpose-built for machines. Demonstrably fed a panel within days in our measurement |
| 3. Independent editorial | An article somebody else wrote about you | Scarce, and the only thing that gives an entity something to say |
| 4. The entity's own marked-up site | schema.org on a domain you control | The anchor everything else is tied back to. Self-published, so it corroborates rather than evidences |
| 5. Owned profiles | Professional networks, journalist databases | Cited heavily in our tests, but easy to create, so they establish consistency more than credibility |
| 6. Syndicated copies | The same article on many domains | Reach without independence. Counts once |
| 7. Open wikis | User-editable encyclopaedias outside the major projects | Can rank well. Cited zero times in our tests, and rejected as a reference by the projects that matter |
| 8. Author archives | Your own byline pages | Evidence you write, not evidence you matter |
The uncomfortable implication for anyone selling in this market is that tiers one and two are cheap and unglamorous, and tiers six and eight are what most packages are actually made of.
The hierarchy above puts public registers at the top, and our measurement showed a UK register being cited directly by a model. That finding comes with an obvious limit: there is no global Companies House. Anyone outside the United Kingdom needs the local equivalent, and they are not equivalent in quality.
| Jurisdiction | The register | What it gives you |
|---|---|---|
| United Kingdom | Companies House | Company number, officer ID, month and year of birth, appointment history. Free, complete, and the strongest of the lot |
| Ireland | Companies Registration Office | Company number, directors |
| United States | SEC EDGAR, plus state registries | No single national company register. EDGAR covers filers and gives a CIK; otherwise it is state by state |
| Germany | Handelsregister | HRB number, managing directors |
| France | Registre National des Entreprises | SIREN and SIRET numbers |
| Australia | ASIC, Australian Business Register | ACN and ABN |
| Canada | Corporations Canada, plus provincial | Corporation number |
| Singapore | ACRA | UEN, officers |
| Hong Kong | Companies Registry | Company number |
| India | Ministry of Corporate Affairs | CIN, and a Director Identification Number per person |
| UAE | DIFC and ADGM public registers; mainland licensing sits with each emirate | Free-zone registers are searchable. Mainland is fragmented, which is a real problem for Gulf-based subjects |
Two things follow from that table. The first is that a client's jurisdiction materially changes how much tier-one evidence is available to them, and a UK director is unusually well served. The second is that fragmentation is exactly the problem cross-border identifiers were invented to solve.
Identifiers travel where registers do not. These are the ones worth attaching to a record, and most have a dedicated property on the major structured knowledge bases, which is the practical reason they matter.
| Identifier | Covers | Why it helps |
|---|---|---|
| LEI | Legal entities, globally | ISO 17442, issued under a single global system. The closest thing to a universal company identifier |
| OpenCorporates ID | Companies across many jurisdictions | Aggregates national registers into one addressable record, useful where the register itself is hard to link |
| ISIN | Listed securities | Ties a person's company to a traded instrument |
| ORCID | Researchers and authors | A persistent identifier for a person, which is rarer and more valuable than a company one |
| ISNI and VIAF | Authors, creators, public figures | Maintained by libraries. Slow, conservative, and precisely the kind of source that is hard to game |
Ask of any source: can the subject edit it, and can a machine address it unambiguously? A register scores well on both because it is maintained by somebody else and issues a number. That test works in every country, which the register list does not.
What we did not measure. Our citation data covers one register, in one country, for one subject. We have no evidence about how models treat EDGAR, ASIC, ACRA or the free-zone registers, and it would be wrong to assume they are weighted like Companies House. Treat the table above as where to look, not as a ranking.
This is a small study and we would rather say so than have it said for us.
What would change our mind: a larger subject set showing open wikis being cited regularly; a panel appearing with no structured record behind it; or evidence that syndicated copies are weighted independently rather than collapsed. We would publish any of those.
Each model was asked a single neutral question of the form "Who is [name], the British [occupation]?", with web search enabled, through a commercial API. Every URL in the returned citation structure was recorded and classified by hand. No prompt engineering was used, and no attempt was made to steer the answer.
Every article in each campaign's coverage list was loaded in a real browser. An article counted as naming the subject only if the surname appeared in the rendered body text. Items behind a paywall were recorded as unverified, never as failures, because we could not read them. Articles sharing an identical headline were collapsed into one original with its syndications listed separately.
Baseline and result were captured as dated full-page screenshots of the same query on consecutive days. Panel contents were additionally pulled through a search API, and, after the error described in section eight, cross-checked with the panel's verified owner.
The verification tooling is ours and unremarkable: a headless browser, a name check against rendered text, and a grouping step. Anybody with a search API and an afternoon can repeat this on their own subjects, and we would be glad to see somebody try and get a different answer.
Subject B is named in a separate case study published with his written permission. Subjects A and C are described but not named here, because consent to be a worked example in one document is not consent to appear in every document. Where a finding depends on a specific verifiable fact, the fact is given without the name attached.
Every external link below was checked in a real browser on 15 August 2026. Two of them, the SEC and the Indian registry, refuse automated requests but resolve normally for a person, which is the same distinction section eight is about.
A note on our own citations. A paper arguing that trusted sources are registers, standards bodies and things somebody else maintains ought to cite exactly those. Every source above is one we do not control, except the last two, which are labelled as ours.
Being findable and being citable are different problems. The first is solved by content and ranking. The second is solved by being unambiguously identifiable in places that are hard to edit and easy to check.
Most reputation spending goes on tier six and tier eight of the hierarchy above, because that is what is easy to sell and satisfying to receive. The things that moved the needle in our measurements were a register entry, a structured record and a handful of articles somebody else chose to write.
None of that is expensive. It is just unglamorous, and it has to be true.