More and more people are skipping your website and going straight to ChatGPT, Google’s AI Overviews, Claude, or Copilot. They type “how do I book a tip slot” or “where’s the nearest walk-in centre open now” and take whatever comes back at face value.

The AI answers whether you’ve done anything about it or not. And right now, it might be quoting last year’s prices, citing a page you forgot to unpublish, or pulling from a Reddit thread instead of your own site.

That probably sounds like a problem that needs a new skillset, but the good news is, it doesn’t.

The best thing you can do for AI search is the thing you’d do anyway: write clearly, keep your content current, and sound like a real organisation talking to real people. AI models decide what to trust based on what humans trust. They’re trained on the content people actually read, share, link to, and come back to. The signal they reward is quality, not trickery.

Public sector communicators have been doing this for years. Plain English, accessible structure, direct answers, GOV.UK style conventions. That work now has a second life: it’s what makes AI choose to cite you rather than someone else.

This guide is about improving what you’re already doing to include what AI needs from you. Most of the steps will feel familiar, and that’s a big help to getting this right.

A BBC and European Broadcasting Union study reviewed 3,000 AI-generated responses to questions about current affairs. 45% contained at least one significant issue. 20% had outright factual errors, including outdated information presented as current. 

That’s news content, heavily covered by professional journalists. Public service information, with its constantly changing dates, prices, and eligibility criteria, is even more vulnerable.

When AI gets your information wrong, the consequences affect people’s lives. A parent misses an admissions deadline. Someone turns up at a charity with bags of clothing they’ve been told aren’t wanted. A resident drives to a tip that’s closed. You end up fielding the complaints, correcting the record, and doing the work twice. And you are left with people trusting you a bit less because they didn’t get the right information and have been inconvenienced. 

We tested it ourselves. We asked ChatGPT and Google straightforward questions about real public services, the way a resident would ask them, and checked the answers against the source.

Some answers were spot on, some were confidently wrong. But the pattern was consistent: where organisations had clear, current, well-structured content on their own site, the AI quoted them accurately. Where there were gaps, stale pages, or buried answers, the AI went elsewhere, to Reddit threads, commercial competitors, old news articles, or sometimes nowhere real at all.

The lesson isn’t that AI is unreliable, it’s that AI is a mirror. It reflects whatever content it can find. If the clearest thing it can find about your service is a three-year-old PDF, that’s what it’ll quote. If the clearest thing is your well-maintained service page, it’ll quote that instead.

Which means instead of asking “how do we optimise for AI?” ask “is our content good enough for AI to trust?” And for most public sector communicators, the answer is: it’s closer than you think.

Let’s break it down into five sections.

In this example, we asked Google how to find an NHS dentist in Manchester. The AI overview cited Reddit and a private dental practice blog alongside NHS sources. Where the official answer was clear, the AI used it. Where there was a gap, it went elsewhere.

Where to find the right questions

You’ve got the data already. Look at tour contact centre’s top ten queries, your social media inbox themes, and your website search logs. If you use a social listening tool, even better! It can show you the questions people ask about you that never reach your inbox.

Tip: Don’t just test your main service pages. Test the smaller, changing stuff: seasonal opening hours, current prices, application deadlines, temporary service changes. That’s where staleness lives, and it’s where AI is most likely to get things wrong.

You’re not auditing to learn a new system. You’re auditing to find out where your existing content is already working and where the gaps are letting someone else speak for you.

What do you do if you find a wrong answer?

Report it. Google AI Overviews have a three-dot menu with a feedback option. ChatGPT has a thumbs-down button on every answer. Use them. But don’t wait for the platforms to fix it for you, because there’s no guarantee they will, or any timeline for when. The faster route is fixing the source: update your own page so the correct information is the clearest thing available, and the AI will follow. Reporting buys you a chance, but fixing your content puts you in control of the truth.

AI tools are looking for content that directly answers the question someone asked. If your page buries the answer three paragraphs down, behind a welcome message and a policy statement, the AI will often skip you and find someone who gets to the point faster.

If you’ve ever been trained in plain English or digital content design, you already know how to do this. The difference is that it now has a direct consequence beyond your own website: it determines whether AI quotes you or quotes someone else.

Practical tips:

This might be the most impactful thing you can do, but it’s potentially a bit of a slog.

Old news releases, superseded service pages, last year’s PDFs, archived consultations. They sit on your site doing nothing, except when an AI tool finds them and quotes them as current.

Content designers at the Department for Business and Trade found exactly this happening on GOV.UK. A Google AI Overview was telling people it costs £40 by post or £13 online to set up a charity in the UK. The actual cost, since February 2026, is £124 by post or £100 online. The primary guidance page had been updated, but the AI found an unmaintained legacy page with the word “charity” on it and quoted that instead. As one user research participant put it: “What else does government have out there that is ‘designed to trip you up’?”

DBT’s response was to audit their content estate and identify 150 pages that hadn’t been updated in five years, had fewer than 11 views, and had no current owner. They didn’t just add “this page is out of date” banners, because user research showed people skip them and AI bots ignore them entirely. Instead, they redirected every page to either its National Archives version, a current GOV.UK page, or the underlying legislation. You can read their full approach on the GOV.UK digital trade blog. 

That’s 150 pages from one department. GOV.UK alone has over 700,000 published pages. And every public sector organisation has its own version of this problem. Councils with old cabinet reports still indexed. NHS organisations with pre-reorganisation service pages. Schools with last year’s admissions guide as a PDF. Charities with COVID-era donation pages still live, listing restrictions that ended two years ago. Transport providers with timetables from the last franchise. 

Local Government Reorganisation is about to make this worse. As councils merge, there will be thousands of pages referring to authorities that no longer exist, services that have moved, and contact details that have changed. AI will keep quoting them until someone unpublishes or redirects them. If you, or your comms lead for the change, can flag this as a risk during the process, you’ll be a step ahead of the issue.

Practical steps:

• Audit your site for pages that haven’t been reviewed in 12 months or more. Most CMS platforms can generate this list.
• Either update, redirect, or unpublish. If a page is genuinely historical (a news release, a past consultation), add a clear banner: “This page is from [date] and may not reflect current information.”
• PDFs are the worst offenders because they’re harder to update and often sit outside your CMS review cycle. If a PDF contains information that changes annually (prices, dates, eligibility), consider replacing it with a web page you can keep current.
• Your old content isn’t just clutter. It’s a competing source of information, and AI might be treating it as your official position.

How do you know it’s working?

The simplest way is to re-run the audit. Pick the same ten questions quarterly and compare what AI says now against what it said last time. If your fixes have landed, you’ll see your own pages being cited where Reddit or old news articles were before.

If you want to go further, check your analytics. GA4 can show referral traffic from AI sources, including ChatGPT, Perplexity, and Google’s AI Overviews. The data isn’t perfect yet, but it gives you a rough picture of how much of your traffic is now AI-mediated and whether people are landing on the right pages. The DBT content team is doing exactly this, tracking how their content appears in AI answers before and after changes, and it’s how they’re building the case for ongoing investment in content maintenance.

This doesn’t need to be complicated. A quarterly audit, a glance at your referral sources, and a note of what’s changed. That’s enough to show progress and make the case for keeping it going.

This is where wrong AI answers do the most damage. During a flood, a major transport disruption, an outbreak, or any fast-moving situation, people will ask AI what’s happening and what to do. If you haven’t published clear, current information on your own site, the AI will find something else to cite, and that something else might be hours old, speculative, or wrong.

You already know this from your crisis comms and misinformation work. The principle is the same: be first, be accurate, be the source of record. The difference is that “the source of record” now includes being the source that AI tools find and quote.

This is where the human-first argument matters most. During an incident, people don’t need machine-readable schema or keyword-optimised headers. They need someone they trust telling them what’s happening in words they understand. That’s always been your job, and it still is. AI just made it more visible and accessible.

Practical steps:

• Publish on your own site before or at the same time as social media. Social posts are harder for AI to find and cite than web pages. Your site is the anchor.
• Update the page rather than publishing a new one. A single page with a clear timestamp and rolling updates gives AI one authoritative source. Five separate news releases across a week give it five chances to quote the wrong one.
• Use clear, factual headings: “School closures in [area], 15 January 2027” rather than “Important update for parents and carers”.
• If you’ve published prebunking or myth-busting content, make sure it’s linked from the main incident page. AI sometimes surfaces FAQ-style content well, so a clear “common questions” section can help correct the record before rumours spread.

There’s a bit of an irony here. If you respond to AI search by churning out AI-generated content onto your site, you risk undermining the exact thing that makes AI trust you.

AI models are trained on content that humans found worth reading. Content with a clear voice, actual expertise, a point of view is how the LLMs learned to judge what’s authoritative and what’s filler. Generic, AI-written pages, the kind that could have come from any organisation in your sector, don’t build those trust signals, they dilute them.

And as we know, trust matters with the messages you’re giving people. Your residents aren’t comparing you to a competitor’s landing page. They’re deciding whether to trust official information from their council, their health service, their fire and rescue service. That trust comes from sounding like a real organisation staffed by real people who understand the community, not from content that reads like it was churned out by prompts.

Use AI to help with research, structure, and editing. Use it to draft technical or informational sections where accuracy matters more than voice. There’s a lot it does well. But the voice, the judgement calls about what to include and what matters most, the “here’s what this actually means for you”, keep those human. That’s what makes your content recognisably yours, and what makes it worth trusting, by the people reading it, and by the AI deciding whether to cite it.

You can watch our short video on how to use AI in the public sector on YouTube or read our blog on our website. 

The organisations that will be quoted most by AI in five years’ time won’t be the ones that optimised hardest for algorithms. They’ll be the ones that consistently publish content good enough for people to rely on, update it frequently, and have built up trust with their communities.

• Run the AI audit: ten real questions, three platforms, log what’s wrong and what’s cited.
• Check your top service pages to answer the question in the first line.
• Date-stamp every page with changing information.
• Identify and deal with stale content: update, redirect, or unpublish.
• Review old PDFs with annual information and replace them with updatable web pages where possible.
• Make sure your incident publishing process puts content on your website, not just social media.
• Check your Google Business profile is current (AI pulls from this too).
• Brief your team: AI search is a content quality issue, not a tech issue.
• Build the audit into your regular content review cycle.
• Don’t replace human voice with AI-generated content on your public-facing pages.