TL;DR
A source of truth is downstream, not upstream. Your page does not settle a disagreement about your company — it only gets believed once the disagreement has already been settled elsewhere.
Which means the page is worth exactly what the surrounding corpus agrees with. Build it around the claims that are contested, and spend the budget on the sources that disagree.
The instrument is the four states of an entity claim: uncontested, contested, stale-consensus and unwitnessed. The state, not the claim, decides what you do next.
The upgrade in 2027 is supersession. A resolver cannot choose between two conflicting values, but it can order two dated values in time — so publish the change, not just the current answer.
Corroboration count is not independent-source count. Twelve directories copying one stale database field are one source, and the correction has one address.
The standard advice for an entity home is a single sentence, repeated across every 2026 guide to the subject: state your facts clearly on your own page, mark them up consistently, and engines will default to your version. It is comforting, it is widely believed, and the measured outcomes do not support it.
In the largest audit published so far, Searchable put more than 13,000 queries about real companies to ChatGPT, Perplexity and Gemini and checked every answer against verified records. Ninety-three per cent of the companies had at least one basic fact hallucinated or missing. The fact types that failed most often were company size, contact details and founding year — the plainest, most stateable facts a business owns, of exactly the kind an entity home exists to publish. Worth noting the sample is UK-only, and worth noting something more uncomfortable: a July 2026 literature check by Stanislav Peev across OpenAlex and Semantic Scholar found no peer-reviewed study measuring AI accuracy on individual business facts at all. Industry research is the entire evidence base, and it should be read with the suspicion you would bring to any widely-repeated backlink statistic. Peev’s own audit excluded three circulating figures as untraceable or misquoted — including the claim that 45 per cent of customers use ChatGPT to find local services, when the underlying BrightLocal data says 45 per cent use any AI tool and the ChatGPT figure is 31.
Where do the wrong answers come from? Mergeflo traced the wrong summaries across 14 startups it onboarded in the first quarter of 2026: 71 per cent came from outdated bios on Crunchbase or Series A press sitting in the top five cited results, and another 19 per cent from misclassified industry fields on LinkedIn. Ninety per cent of the damage originated on somebody else’s property. None of those companies lacked an about page.
This is the thing the guides get backwards. An entity home is not an upstream declaration that engines consult before believing anything else. It is a downstream artefact whose credibility is set by how much of the surrounding corpus already agrees with it. On any fact where every other source concurs, the page is redundant. On any fact where sources conflict — the only occasion it is actually needed — it is the minority report.
That is not an argument for skipping the page. It is an argument for building a different page, and for understanding what it is for. The entity home is not the evidence. It is the address the evidence points at. Everything below follows from taking that seriously: which claims belong on it, how to publish a fact that has changed, where a correction has to land to take effect, and why the campaign that makes the page true looks far more like digital PR and outreach than like technical publishing.
What the page can do alone, and what it cannot
Split the problem in two, because the two halves behave completely differently and conflating them is why entity work so often disappoints.
The first half is identity: which real-world thing you are. When an engine merges you with a same-named company, it has failed at entity resolution — the process of deciding that two mentions refer to one thing — and it will then cross-contaminate the attributes of both, splitting your branded citations between two records. The cost is quiet, which is what makes it dangerous: nobody sees an error, just a confident answer about somebody else. It is also the failure that agentic browsers acting on a page propagate hardest, because an agent that has resolved you wrongly will keep acting on that resolution for the rest of the task. Here first-party material genuinely does most of the work. A stable canonical name, formal identifiers such as a company registration or LEI number, a specific registered address rather than a city, and a description narrow enough that a namesake could not paste it onto their own site unaltered: these are the handholds a resolver uses. If a rival with your name could copy your about page verbatim and have it read as true, you have not disambiguated anything.
The second half is attributes: what is true about the thing once identified. Founded when, based where, led by whom, owned by whom, does what for whom. Here first-party material does not win by default, because the resolver is not consulting a register — it is weighing documents that disagree, and yours is one of them.
What is an entity home page?
An entity home is the single canonical page on your own domain that states who your organisation is and the facts most likely to be misstated about it, at a stable URL that never moves. Its function is not to be believed on its own authority but to be the fixed reference that every other record about you can be aligned to.
The distinction tells you where effort pays. Identity work is cheap, one-off and under your control. Attribute work is continuous, mostly off-site, and looks like outreach. Teams that spend a quarter perfecting the page and none on the corpus have optimised the half that was already winning.
Key takeaway
First-party pages settle identity — which entity you are. They do not settle attributes — what is true about you. Attribute conflicts are resolved by weighing documents, and yours is one document among many.
The four states of an entity claim
Every fact you might publish about yourself is already in one of four states out in the corpus, and the state determines the action. This is the instrument. It is not a score, and there is nothing to weight: you look up the state, and the state tells you what to do.
The four states of an entity claim
Uncontested. Every source that mentions the fact agrees, and agrees with you. Your page adds no information. State it once, plainly, and spend nothing further on it.
Contested. Sources actively disagree. Your page is one vote among several and will not carry the decision. Action: trace the disagreement to its origin and correct there.
Stale-consensus. Every source agrees on something that used to be true. Corroboration is at its maximum and correctness is zero, and the moment you publish the new value you become the outlier. Action: publish the change with dates, not the answer.
Unwitnessed. Only you say it. There is nothing to resolve, because there is no second observation. Action: either create a witness or stop asserting it as a fact.
Run any about page through this and most of it sits in the first and fourth states — uncontested facts nobody was going to get wrong, and unwitnessed claims about being trusted, leading or innovative. Neither does any work. The two middle states carry the entire value of the page, and conventional about pages never address them, because they are written as though the reader arrives with no prior beliefs at all.
| State | What the corpus looks like | Where the page helps | The real move |
| Uncontested | All sources agree with you | Barely | State once; move on |
| Contested | Sources disagree with each other | As a reference | Correct the origin |
| Stale-consensus | All agree; all now wrong | As a dated record | Publish supersession |
| Unwitnessed | Only your page says it | Not at all | Get corroborated |
Why the state, not the fact, sets the budget
A contested founding year and an uncontested founding year look identical on the page. They cost radically different amounts to fix, and only one of them can be fixed by writing. Treating them the same is how entity projects end up as copywriting exercises with no measurable effect, which in turn is why the discipline has a reputation for being unfalsifiable — the measurement problem that dogs AI visibility work generally.
For example: a Sheffield firm whose site says Sheffield and whose every listing says Sheffield gains nothing from a bolder statement of Sheffield. The same firm, whose 2019 directory entries still say Manchester, gains nothing from a bolder statement either — because the sentence is not what is losing. Eleven documents are.
Supersession beats contradiction: the actual upgrade
Stale-consensus is the state worth building the page around, and it has a specific fix that almost nobody ships.
Consider what a resolver faces when a company moves its headquarters. Forty documents say Leeds. One document — your newly edited about page — says Manchester. There is no way to read that as anything but a minority claim contradicting a settled fact, and it will be treated accordingly, which is why teams so often report that the correct answer sits on their site for months while the engines carry on saying the old one.
Now change one thing. Instead of asserting the current value, publish the transition: the previous value, the new value, and the effective date of the change. The forty documents stop being contradictions and become consistent history. A resolver cannot arbitrate between two conflicting values, because nothing in the corpus tells it which is later. It can order two dated values in time trivially. You have converted an unwinnable disagreement into a chronology, and chronology is the one form of conflict a document-weighing system handles well.
This is the difference between an about page and a source-of-truth page, and it is the reason the entity home needs upgrading rather than merely writing. A source of truth carries its own history. It does not quietly overwrite last year’s answer; it records that last year’s answer was correct until a stated date and is not correct now. Handraise’s 2026 guidance on correcting brand answers arrives at the same place from the other direction, recommending that current state, previous state and effective date all be made explicit whenever timing matters — precisely because AI answers go wrong when sources blur the difference between what was true and what is.
What a supersession record looks like in practice
One line per changed fact, in visible body text rather than tucked into markup, each carrying four elements: the attribute, the superseded value, the current value, and the date the change took effect. Registered office: Leeds until 14 March 2024; Manchester from 15 March 2024. It reads like a corrections column, which is the point — corrections columns are a two-century-old solution to exactly this problem, and they work because they make the old value findable rather than pretending it never existed. The instinct is the same one behind provenance records that travel with an asset: a claim you can place in time is worth more than a claim asserted flatly, because the reader can reconcile it against everything else they hold.
Two disciplines make the difference between a supersession record and a changelog nobody can use. Keep it to attributes a stranger might state about you, not internal events; a funding round belongs here, a rebranded internal team does not. And never delete an entry once published, because the value of the record is that a document from 2023 can be reconciled against it. A record that only ever shows the present is just an about page with dates on it.
Key takeaway
A resolver cannot choose between two conflicting values, but it can order two dated values in time. Publishing the transition — old value, new value, effective date — turns every outdated document from a contradiction into corroborating history.
The resolver map: a correction aimed at the wrong corpus does nothing
Not every fact about you is settled in the same place. This is the finding that reorganises the work, and it comes out of direct observation rather than theory.
In August 2026 Vishnu Harshan published a log of the same company question put to four engines, with every cited URL recorded. The result was not that owned sources beat third-party ones — the opposite. The database-heavy answer was the more accurate one on founding year, headquarters and a recent acquisition, while being thin on what the company actually teaches. His conclusion is the useful one: source class determines which part of the question an engine can answer. Investor databases track firmographics as a matter of course and carry nothing about capability; editorial carries capability and is unreliable on dates.
So a correction campaign has to be aimed. Rewriting your about page to be clearer about your founding year targets a corpus that was never going to decide your founding year. The corollary is that you should read your own citation logs by source class rather than by domain, since the diversity of sources an answer draws on tells you which corpus is holding the pen on each part of the question.
| Fact class | Examples | Settled by | Correction lead time |
| Firmographic | Founded, HQ, funding, owner | Company and investor databases, registries | Two to six weeks |
| Operational | Hours, contact, service area | Listings and platform profiles | Days to weeks |
| Capability | What you do, for whom, category | Editorial, reviews, owned long-form | Months |
| Relational | Acquisitions, partners, departures | Press and public filings | Fast, then permanent |
Read the right-hand column carefully, because it sets expectations that otherwise get set by hope. DeepSmith’s 2026 account of the same work puts high-authority third-party corrections showing up in retrieval answers within two to six weeks — slower than editing your own site, and worth the wait. Capability claims are slowest of all, because they are settled by accumulated editorial rather than by a field somebody can amend, which is why positioning is fixed by publishing and being covered rather than by declaring.
A second timing layer sits underneath. A retrieval-grounded answer changes as soon as the underlying documents do; a fact memorised in model weights shifts only at the next training cut. That is a real ceiling, and a reason to treat recovery from a bad AI answer as a programme rather than a fix.
The origin trace: twelve copies is one source
The instinct on finding a wrong fact in eleven places is to correct eleven places. Usually there are not eleven sources. There is one source and ten copies.
Data-integration research has worked on this for well over a decade. The truth-discovery literature — Dong, Berti-Equille and Srivastava’s work on source dependence and copying detection, later folded into Google’s Knowledge Vault approach to probabilistic knowledge fusion — starts from the observation that sources are not independent, and that treating them as independent lets a copied error masquerade as consensus. The detection principle is elegant: sources that make many of the same mistakes are unlikely to be independent of one another. The published surveys are candid about where it breaks, too. Copy detection struggles when sources copy from a good source, because shared correct values look identical whether they were reasoned or lifted.
That cuts two ways, and both are useful. While a wrong value is propagating you cannot count on any resolver noticing that the eleven agreeing documents are one document repeated — the discount applies only once the shared error is detectable, and by then the answer has been served thousands of times. But the machinery is on your side after a correction: fix the origin and the copying that worked against you starts working for you, free.
Finding the origin without a research budget
Three signals locate an origin quickly. Identical phrasing across supposedly independent records means one of them wrote the sentence and the rest pasted it. Identical error patterns — the same transposed digit, the same misspelled surname — mean the same thing more strongly. And publication dates give you direction of travel: the earliest record carrying the wrong value is usually the parent, and syndicated press wires produce a characteristic burst of same-day copies that all descend from one release. Standard backlink and mention analysis tooling handles most of this, because tracing a claim across the web is mechanically the same job as tracing who links to what and when.
There is a hard case worth naming, because it is common and it is where the whole approach strains. Sometimes the origin is a filing, an archived article at a publisher with no corrections process, or a database whose amendment route runs through a form that nobody answers. You cannot fix those, and pretending otherwise wastes a quarter. The response is not to correct the origin but to out-corroborate it: date-stamped supersession on your own page, plus enough independently dated third-party records carrying the current value that the chronology is legible without the origin’s cooperation. It is slower and it is the only route available.
A worked example: Meridian Freight Systems, fourteen weeks
Meridian Freight Systems is a Manchester logistics-software firm, rebranded from Loadway in March 2024, having moved from Leeds the year before. It shares a name with an unrelated freight brokerage in Atlanta. Its buyers are operations directors who increasingly arrive having already asked an assistant what the company does.
Week 1. Baseline. Twelve prompts across four engines, run three times each, every cited URL recorded — single questions rather than follow-ups, since answers drift across a multi-turn exchange and a baseline needs to be repeatable. Four recurring errors: headquarters given as Leeds in 9 of 36 answers; founding year given as 2016 rather than 2013 in 14 of 36; the Atlanta brokerage’s services described as Meridian’s in 6 of 36; and the old Loadway name presented as current in 11 of 36.
Week 2. States assigned. Headquarters was stale-consensus. Founding year was contested. The Atlanta collision was an identity failure, not an attribute failure. The company’s claim to be the UK’s leading multimodal platform was unwitnessed, and was struck from the page entirely rather than defended.
Weeks 3 to 4. Origin trace. The 2016 founding year appeared in seven places with the same phrasing, all descending from a single 2016 funding announcement that had described the year of incorporation of a holding company rather than the trading entity. One origin, six copies. The Leeds address traced to two parents: an industry directory and a trade-title profile page.
Weeks 5 to 8. Corrections at the origins, plus a supersession block on the entity home covering the address move, the rebrand and the incorporation discrepancy, each with an effective date. The identity work ran alongside it: registration number and full registered address stated in visible text, and a description narrow enough that the Atlanta firm could not have used it.
Weeks 9 to 14. Re-testing on the same twelve prompts, plus a control run through extended research modes, which pull more sources and therefore surface stale ones the short answers had stopped citing. Headquarters errors fell from 9 of 36 to 1. The Loadway naming error fell from 11 to 2. The founding year moved more slowly — 14 of 36 to 6 — and the residual cases all cited the original funding announcement, which the publisher declined to amend. The Atlanta confusion fell from 6 to 0, and it fell fastest of the four, which fits: identity was the half the owned page could actually settle.
The honest read. Three of four errors were substantially reduced in a quarter, one was not, and the one that resisted was the one whose origin was outside the company’s reach. Nothing here produced a traffic number worth reporting. What it produced was a sales team no longer opening calls by correcting the prospect’s assistant — which is the return this work actually pays, and it is not a search-traffic return.
Key takeaway
Four error types, four different treatments, four different speeds. The identity error cleared fastest and the attribute error with an unreachable origin barely moved. Sort by state before budgeting, or you will spend the quarter on the one that cannot move.
The objection: first-party sources really are privileged
The strongest counter-argument deserves stating at full strength, because a great deal of competent practice sits behind it. Engines are explicitly built to prefer an entity’s own declaration where a declaration exists. Structured data on the entity home, a stable identifier for the organisation node, consistent linking out to official profiles, formal registration numbers — these are not decoration, they are the inputs a resolver is designed to privilege. And retrieval-grounded answers do correct quickly once the owned page is re-crawled, which is far faster than any third-party route. On this account the page is upstream after all, and the corpus follows it.
Most of that is true, and the piece would be wrong to wave it away. The resolution is that the privilege is real but scoped, in two specific ways.
It is scoped to fields rather than sentences. Where there is a defined property and a claiming mechanism, first-party declaration is decisive; the machinery exists precisely to let the entity speak for itself. Where the output is free text assembled from retrieved documents, there is no field to claim, and the declaration competes as one document. That is why the Searchable and Mergeflo numbers look the way they do despite near-universal adoption of the recommended markup: the markup wins the fields, and the answer was never built from the fields.
And it is scoped to identity rather than attributes, which is the split this article opened with. Privilege gets you correctly resolved. It does not adjudicate between your dated claim and forty older documents, because at that point the engine is not asking who you are; it is asking which value is current, and yours has no special standing on that question. The Meridian numbers show the shape: the identity error went to zero, the attribute error with a live origin fell by two thirds, and the attribute error with a dead origin barely moved.
So the objection wins its half and the argument survives on the other. First-party work is necessary and cheap, and skipping it makes the problem harder for no saving. It is not sufficient, and treating it as sufficient produces a year of confident publishing with nothing measurable at the end.
What the upgraded page actually carries
Everything above dictates the contents, and the dictation is unusually specific. The page opens with the disambiguating material, because identity is the half it can settle alone: canonical legal and trading names, registration number, full registered address, and a description narrow enough to fail if a namesake tried to reuse it. Vague category language is the single biggest own goal here — a description that could describe forty companies has told a resolver nothing about which one you are.
It then carries the contested and superseded attributes, and only those. Founding year if founding year is disputed; ownership if ownership changed; product naming if the old names are still circulating. Uncontested facts get one plain line and no emphasis. Unwitnessed claims are cut, without exception and without negotiation, because an unwitnessed superlative on a source-of-truth page is the fastest way to make the whole page read as marketing to a system deciding whether to trust it.
Three structural properties matter more than the wording. The URL never changes, because every correction you file elsewhere will point at it and a moved page turns a year of outreach into 404s — the same fragility that makes broken and redirected URLs a live technical liability. The facts appear in rendered text rather than only in markup or in client-side components, since not every reader executes scripts and a crawl that never sees the content cannot corroborate anything — worth confirming directly as part of any audit of how your pages surface without a search results page. And every claim carries a date, because a fact without a date cannot be ordered against a document that has one.
Where should the entity home live?
Put it on your primary domain at a permanent path, distinct from a marketing about page, and link to it from your footer so it is reachable from every page. It should be a plain, fast, text-first document that renders fully without scripts, because its readers include tooling that will not run them.
What it should not become is an interactive asset. Anything that invites exploration invites partial reading. The page is one screen of stateable facts plus a dated supersession record, and its virtue is that a stranger can read the whole thing and leave with the right values — the discipline that separates a working reference page from a collection of assets built to be browsed.
Why the correction campaign is a link-building campaign
The origin trace produces an artefact that will look familiar: a list of publishers, databases, directories and trade titles carrying a claim about you, ranked by how much traffic the claim is doing. That is a prospect list, assembled from evidence rather than from a scrape — and the outreach that works it is ordinary link building with a different opening line, which is why it usually belongs with the person who already owns outreach rather than with a separate entity workstream.
The ask attached to it is unusually strong. A correction request is not a favour — it offers the publisher an accuracy fix on a live page, which is a thing editors are professionally obliged to care about and a materially easier conversation than asking for a mention. Correction requests routinely come back with the citation attached, because the natural way to substantiate an amendment is to point at the record that supports it. This is the same mechanic as reactive commentary and journalist queries, where the currency is being useful to a deadline rather than being persuasive, and it shares a failure mode with newsjacking: send it late and it is worthless.
Two cautions, both learned expensively by people who ran this as a volume play. Never name the competing entity in your own copy when the problem is a namesake; asserting a relationship is how you create one in a graph that reads co-occurrence as evidence. And do not manufacture corroboration. Seeding agreeing profiles across low-quality directories is detectable as a copying pattern by the very machinery described above, and it converts an accuracy problem into the sort of manipulated-signal problem that negative-signal defence work exists to clean up.
Measure it as accuracy, not as visibility. The metric is the share of a fixed prompt set returning each fact correctly, re-run monthly on the same prompts, with the cited URLs logged so you can see which document is still doing the damage. Reporting it as citations or referring domains borrows credit from other work and hides the only signal that tells you whether the campaign is working.
The Monday checklist
Seven steps. The first three are a week’s work; the rest is a quarter.
- Fix a prompt set and baseline it. Ten to fifteen questions covering the facts a buyer acts on, run across the engines your market uses, three times each, with every cited URL recorded.
- Assign a state to every wrong or missing fact. Uncontested, contested, stale-consensus or unwitnessed. Do this before deciding anything, because the state sets both the budget and the expected speed.
- Separate identity failures from attribute failures. Same-name confusion is fixed on your own property with names, identifiers and a narrow description. Attribute conflicts are not.
- Trace each contested fact to its origin. Match on identical phrasing, shared error patterns and earliest publication date. Expect the count of real sources to be a fraction of the count of documents.
- Publish supersession, not the answer. One dated line per changed attribute: old value, new value, effective date. Never delete an entry once it is live.
- Work the origins, aimed by fact class. Databases and registries for firmographics, listings for operational detail, editorial for capability. A correction filed in the wrong corpus changes nothing.
- Re-run the same prompt set monthly. Report per-fact accuracy and the URLs still carrying the error. Twelve weeks is a fair first read; six months is a fair verdict.
The failure threshold: if after two full quarters a fact is still wrong and every remaining citation traces to an origin that has refused or ignored amendment, stop working the origin. Switch to out-corroboration — dated third-party records carrying the current value — and accept a longer horizon. Some facts are simply held elsewhere, and recognising which ones early is worth more than persistence.
The agent web has changed what an about page is for. It was a brochure, read by people who had already decided to be interested. It is now the reference a machine reconciles other people’s records against, on questions the reader never sees being asked. The reframe is a relief: you are no longer trying to be persuasive on your own page. You are trying to be checkable, dated and hard to confuse with anybody else — and then doing the unglamorous work of making the rest of the web agree.
