The short version
- A wrong AI answer about an institution is a symptom. The five root causes are: the fact is not machine-readable, the page is not indexed, the fact is stale in the model, there is a language gap, or your page is outranked by a third-party source.
- Each cause has a different fix. Republishing a fact that is merely stale in one engine is wasted effort; earning authority against an outranking source does nothing if the real problem is a scanned PDF.
- You can tell them apart with evidence, not guesswork: fetch the official page the way a crawler does (no JavaScript), check whether the current fact is in the server HTML, and compare that against what each engine returned.
- Only one of the five — stale in the model — resolves itself with no work. The other four are things you fix, and three of them are entirely within your control.
- This is the diagnosis our accuracy engine computes for every wrong answer, so a correction is aimed at the actual cause rather than the most visible symptom.
A wrong answer is a symptom, not a diagnosis
When an AI assistant misstates a fact about your institution, the natural response is to reach for the most familiar fix — usually "update the website". But the website is often already correct. The reason the model is wrong could be that a crawler cannot read your page, that it has never fetched the current version, that one particular engine is simply behind, that the fact only exists in the wrong language, or that a third-party source beat your page in that answer. These are five genuinely different problems, and the fix for one is useless against another.
So the first move is not to correct — it is to diagnose. The good news is that diagnosis does not have to be guesswork. Three pieces of evidence, gathered mechanically, separate the five causes: whether the current fact is present in your page's server-rendered HTML, whether any engine at all returned it correctly, and which sources the failing engine actually cited. From those, the cause falls out. Here is each one, what it means, and how to fix it.
How the evidence is gathered
The official page is fetched the way a text-only crawler sees it — a single request, no JavaScript executed. If the current fact is in that raw HTML, the page is machine-readable. Cross-referencing that against what each engine returned, and which domains it cited, is enough to compute the specific cause rather than infer it.
Cause 1 — Not machine-readable
The most common cause, and the most fixable. The current fact is not in your page's server-rendered HTML at all. It is there for a human — but it lives in a script-only widget that a crawler does not execute, a scanned PDF that is really a photograph of a page with no text layer, or an image a machine cannot read. The fact is published, free, and authoritative, and still completely invisible to the crawler.
When your authoritative page is unreadable, the model does not give up — it fills the vacuum from whatever it *can* read: an old news summary, a third-party explainer, a superseded cached copy, or its own training data. The confident wrong answer is the model reconstructing a fact your own format hid from it.
The fix
Restructure the fact as plain text or structured data. State it as ordinary server-rendered HTML near the top of a canonical page, transcribe key provisions out of scanned PDFs, and add JSON-LD so the fact is explicit. This cause is entirely within your control — see the full treatment in why AI gets government facts wrong.
Cause 2 — Not indexed
Here the page is fine. The current fact is in your server-rendered HTML, plainly readable — but no AI crawler has fetched the current version yet. The page may be new, recently updated, blocked by an overzealous robots rule or firewall, or simply not yet re-crawled since the fact changed. The distinction from cause 1 is sharp: there, the crawler read your page and the fact was not in it; here, the crawler has not read the current page at all.
You can tell this cause from the outside: your page carries the fact, yet no engine returns it. Nobody has the current value, which means nobody has read the current page. This is a reach problem, not a format problem.
The fix
Get the page crawled and confirm the fetch, then re-measure. Make sure nothing in your robots rules, firewall or rendering setup is turning the bots away, point them at the page through your llms.txt index, and use AI crawler analytics to verify the current version was actually fetched before you expect the answer to change.
Cause 3 — Stale in the model
The one cause that needs no work. Your page is correct, it is machine-readable, and it has been fetched — you can tell because other engines already answer the question correctly. This particular engine is simply behind: it is running on an older training snapshot and has not yet refreshed against your published record. Because the correct answer is reachable and other engines return it, this one will catch up on its own.
This is the difference between an answer that is *outdated* and one that is *incorrect*. An outdated answer sits behind a correct, published, dated record and will self-correct as the model re-reads you. Recognising this cause matters because the wrong response — republishing, restructuring, chasing citations — wastes effort on a problem that is already resolving.
The fix
Monitor; no work needed. Keep the canonical page current and dated so the lagging engine has a clean record to refresh against, and watch the answer rather than touching the page. The most valuable thing a diagnosis does here is tell you to do nothing.
Cause 4 — Language gap
A fact can be correct in one language and wrong in another, and this is its own distinct cause. The current fact simply is not published in the language the question was asked in — a crawler finds it only in another language, usually English, and the AI assistant answers the native-language question from a weaker source or an outdated translation. For institutions that serve the public in more than one language, this cause hides underneath every other one until you measure per language.
The critical discipline is to never fix a language gap with more English. Adding more English pages does nothing for a citizen asking in another language; the model needs the fact published natively to retrieve and quote it in that language. A single English source, however authoritative, cannot close a gap that exists precisely because the native-language publication is missing.
The fix
Publish the native-language version of the fact — as machine-readable text with its own effective date — and re-measure in that language. Treat each language you serve as a separate accuracy surface, because a fact that passes in one can silently fail in another.
Cause 5 — Outranked
The most galling cause, because you did everything right. Your official page carries the current fact, it is machine-readable, and it has been fetched — but in this answer the engine cited a different, non-official source instead. A journalist's paraphrase, an aggregator, a community-maintained entry or an old article won the answer over your authoritative page. Unlike "not indexed", where the engine cited nothing, here you can see exactly whose source beat yours.
This happens because models weigh popularity, link volume and repetition, not just authority and recency. A widely-repeated third-party version can outweigh a correct official one simply because more of the web echoes it. The fix is therefore not on your own page — it is about authority in the wider ecosystem.
The fix
Earn authority so the official record wins the answer — get your canonical page cited, and get the wrong third-party source corrected or de-prioritised. This is where disciplined digital PR supports accuracy: keeping the third-party record — encyclopaedic entries, directories, credible coverage — aligned with your official one, so the source the model trusts is also the source that is right.
Why telling them apart is the whole game
Five causes, five fixes, and almost no overlap between them. Restructuring a page as machine-readable does nothing if the real problem is that a popular old article outranks you. Chasing citations is wasted if your fact is trapped in a scanned PDF no crawler can read. Republishing is pointless if the engine is merely stale and about to refresh on its own. And more English pages actively fail to close a native-language gap. The reason so many correction efforts stall is that they treat every wrong answer the same way — usually by editing the website — when only one or two of the five causes are even about the website.
Notice, too, how the causes divide by control. Three of them — not machine-readable, not indexed, and language gap — are entirely within your power to fix directly. One — outranked — you influence through authority rather than control outright. And one — stale in the model — you do not fix at all; you wait, and it resolves. Knowing which bucket a given wrong answer falls into tells you not just *what* to do but *whether to do anything*.
5
distinct root causes behind a wrong AI answer — each with a different, specific fix
3
of the five are entirely within your control: machine-readability, indexing and language
1
resolves with no work at all — stale in the model — which is why diagnosis before action saves effort
How the diagnosis is computed
This is not a framework we invented for the article — it is the diagnosis our accuracy engine computes for every wrong answer it finds. When GenAI Ranker measures how the six major AI assistants answer questions about your institution and grades an answer as outdated or incorrect, it does not stop at the symptom. It fetches your official page the way a crawler does, with no JavaScript, and checks whether the current fact is actually in the server-rendered HTML. It cross-references that against whether any engine returned the fact correctly, and against which domains the failing engine cited.
From that evidence it computes the specific cause — not machine-readable, not indexed, stale in the model, language gap, or outranked — and states it in one honest sentence with the matching fix. When the fix is a content change, it drafts the corrective assets a verified record needs — plain-HTML fact statements, JSON-LD, and llms.txt entries — for your team to review and release. And because the whole loop is measured per engine and per language with an audit trail, you can prove a correction landed rather than hope it did. The method for building the durable version of this is in how to build a verified record.
For public institutions
If AI assistants are giving outdated answers about your mandate, leadership or eligibility, the useful first step is a diagnosis, not a rewrite — knowing which of the five causes you actually have. Book a meeting with our public-sector team and we will run the questions across all six engines and show you the graded results and the computed cause behind each wrong answer.
The institutions that stay accurately represented in the AI era are not the ones that publish the most. They are the ones that, when an answer is wrong, ask *why* before they act — and then fix the actual cause. A wrong answer is a symptom. The five causes are the diagnosis. Matching the fix to the cause is how a public body keeps its own record from being overwritten by a confident guess.