The short version
- People now ask AI assistants basic questions about government, and the models frequently answer with outdated office-holders, superseded policies and old figures.
- The root causes are structural: training-data lag, thin machine-readable official sources, and many conflicting copies of the same fact scattered across the web.
- The fix is not a press release. It is making the official record the most authoritative, structured and machine-readable version of each fact.
- Treat AI answers as a channel you monitor for drift, correct when wrong, and measure over time — the same discipline you already apply to the public record.
The new front desk is an AI assistant
A decade ago a citizen with a question about a visa rule, a benefits scheme or the current holder of a public office would search, land on an official page, and read the answer in your own words. Today a growing share of those same questions never reach your website at all. They are typed into a general-purpose AI assistant, and the person reads a synthesized answer that may draw on your pages, on years-old news coverage, on third-party explainers, or on the model's own training data — blended into a single confident paragraph with no byline and often no link.
For a commercial brand this shift is a marketing problem: the concern is whether the model recommends you. For a public body the problem is different and, in some ways, sharper. Nobody is asking the AI assistant to *recommend* a ministry. They are asking it to *state a fact* — who holds an office, what a current policy says, which scheme covers a particular need, what a figure or threshold is this year. When the model answers, most readers assume the government said so. If the answer is wrong, the error is attributed to you, and you were never in the room.
Why AI gets government facts wrong
The failures are rarely malicious and almost never random. They follow a small number of structural patterns that any communications team can learn to recognize.
Training lag freezes an old version of the truth
Large models are trained on a snapshot of the web taken at some point in the past, then released months later. Anything that changed after that cutoff — a new appointment, a revised threshold, a superseded regulation — is invisible to the base model unless it retrieves live sources at answer time. Government facts are unusually vulnerable here because they change on a schedule the model cannot see: cabinets are reshuffled, budgets are revised annually, programs open and close. The model will happily name last term's office-holder with total confidence, because within its frozen world that answer is still correct.
Thin machine-readable sources force the model to guess
Even when a model does retrieve live pages, it can only use what it can parse. Many official facts live inside PDFs, image-based notices, dense legal prose, or pages that state the current position only implicitly. If the single authoritative sentence — "the current threshold is X" or "the office is held by the Minister for Y" — is not written plainly in text a machine can lift, the model reconstructs it from weaker signals: a news article, a summary site, a cached older page. Absence of a clear machine-readable fact is itself a cause of error.
Conflicting copies leave the model to arbitrate
The same government fact usually exists in dozens of places: your own current page, your own archived pages, news coverage from several years, explainer sites, aggregators, and forums. When these disagree, the model has to pick a winner, and it does not automatically know that your page outranks a popular five-year-old article. Popularity, link volume and repetition can all outweigh recency and authority. A widely-cited old figure can beat a correct new one simply because more of the web still repeats it.
Confidence is not accuracy
AI assistants present stale government facts in exactly the same authoritative tone as correct ones. There is no visible uncertainty marker for a reader to catch. This is why drift goes unnoticed until a journalist, a business or a citizen acts on a wrong answer — the error carries the full apparent weight of an official statement.
Why this is a civic risk, not just a comms nuisance
When an AI assistant misstates a private company's return policy, a customer is inconvenienced. When it misstates a public rule, the consequences compound. A business may plan around a superseded regulation. A journalist may publish a figure that is a year out of date, attributing it to the government. A vulnerable person may believe they are ineligible for support they actually qualify for, or apply for something that no longer exists. Each wrong answer quietly transfers the cost of the model's error onto a citizen and onto public trust in the institution.
There is also a slower, corrosive effect. If AI assistants repeatedly return outdated or contradictory information about a public body, the institution starts to look disorganized or evasive even when its actual published record is accurate and current. The gap between what the government says and what the AI reports becomes, in the public mind, a gap in the government's competence. Correcting the record is therefore not reputation management in the commercial sense — it is part of the basic duty to communicate accurately.
6+
major AI assistants citizens now query — ChatGPT, Gemini, Claude, Perplexity, Grok and AI Overviews — each with its own view of your facts
Months
typical lag between a model's training cutoff and its public availability
0
visible uncertainty markers a citizen sees on a confidently wrong answer
A playbook: make the official record the source AI trusts
The goal is not to argue with the models. It is to make your version of each fact the cleanest, most authoritative, most machine-readable one on the open web, and then to verify that the AI assistants have picked it up. This is the same discipline as generative engine optimization, applied to accuracy rather than promotion. Five practices carry most of the weight.
1. Give every changeable fact one canonical page
For each fact that people ask about and that changes over time — current office-holders, current thresholds, current scheme eligibility, current policy status — designate a single canonical URL that always states the present position in plain text. Not a PDF, not a news post, not an announcement dated to one moment, but a living page whose job is to say what is true *now*. When a fact changes, you update that page rather than publishing a new one and leaving the old one to compete with it.
2. Write the fact in text a machine can lift verbatim
State the answer in one unambiguous sentence near the top of the page, in the exact form a person would ask it. "The current [scheme] is open to residents who meet the following criteria …" is retrievable; a criteria table buried in an appendix PDF is not. Assume the model will quote one sentence and design that sentence to be correct out of context.
3. Mark it up as structured data
Machine-readable annotation lets an AI assistant read your fact without inferring it from prose. Government-relevant schema — organizations and their office-holders, services and their eligibility, effective dates on regulations — turns an implicit fact into an explicit, typed statement. See structured data for GEO for the mechanics. Crucially, always publish an explicit effective or last-reviewed date; recency is one of the few signals that reliably helps a model prefer your current page over an older copy.
4. Publish a machine-readable fact sheet the crawlers can find
Beyond individual pages, maintain a consolidated, plainly-structured index of your key current facts and point AI crawlers to it. An `llms.txt` file and clean, unblocked access for the AI assistants' retrieval bots make it far more likely they read your authoritative source rather than a third-party summary. Confirm the bots can actually reach these pages — AI crawler analytics will show you which AI assistants are fetching your content and which are not.
5. Retire or clearly date the old versions
Superseded pages are a major source of drift. When a fact changes, either redirect the old page to the canonical one or stamp it unmistakably as archived, with the superseding date and a link to the current position. An undated old page that still reads as current is a landmine — the model has no way to know it has been overtaken.
Correct once, at the source
When you find an AI assistant stating a wrong fact, resist the urge to treat it as a one-off. Fix the canonical page, add the structured data and the effective date, retire the conflicting copies, then re-check the AI assistants after they next crawl. A correction made at the source propagates across every model; a correction made in a single conversation vanishes when the chat ends.
Monitoring for drift
Publishing accurately is necessary but not sufficient, because you cannot see what the AI assistants are saying unless you look. Drift is silent: a model updates, a competing source gains traction, a page falls out of the retrieval set, and suddenly the answer changes without anyone touching your website. Communications teams already monitor coverage; AI answers deserve the same standing watch.
In practice this means maintaining a list of the factual questions the public actually asks about your remit, phrased the way they phrase them, and periodically checking how each major AI assistant answers. The questions that matter most are the ones where a wrong answer causes real harm: eligibility, current office-holders, current figures, whether a program still exists. Track how the answer changes over time and which sources the model cites, so you can see drift the moment it starts rather than after a wrong answer reaches the public. Measuring AI-search visibility covers the monitoring approach; for the public sector, the metric is accuracy, not prominence.
Measuring accuracy over time
Because the target is correctness, the scoreboard should be built around it. A useful accuracy program tracks a handful of concrete measures rather than a single vanity number.
- Factual accuracy rate: across your list of public questions and the major AI assistants, what share of answers are currently correct?
- Currency lag: when a fact changes, how many days until each AI assistant reflects the new position? This is the clearest measure of whether your source is authoritative.
- Source attribution: when the model cites a source, is it your canonical page or a third-party copy? Rising citation of your own pages is the leading indicator that corrections are landing.
- Drift incidents: how often does a previously-correct answer regress, and how quickly do you catch and fix it?
- Coverage: for how many of your high-stakes public questions does an authoritative, dated, machine-readable canonical page exist at all?
Reported over time, these turn an anxious, reactive posture into a managed one. You will not achieve a permanent hundred percent — the models change beneath you — but you can drive the accuracy rate up, the currency lag down, and the share of answers grounded in your own official record steadily higher. That trend line is the real deliverable: evidence that the public is getting the government's actual position when they ask a machine, and that the gap between the record and the answer is closing.
None of this replaces authoritative publishing, public-information law or the human spokesperson. It extends them to a channel that has quietly become one of the first places people go for facts about their government. Treating AI answers as part of the official communications surface — published to, monitored, corrected and measured — is how a public body keeps its own record from being overwritten by a confident approximation of it.