Citizens are increasingly asking their questions to ChatGPT or Google instead of a government website. At a roundtable organized by Smals, five specialists outline what that means: fewer visitors, many more crawlers, and a battle that is no longer about the click, but about the citation.
“The battle for clicks is over, the battle for the mention begins,” says Rafal Naczyk, web copywriter at Smals, opening the roundtable discussion. AI has drastically changed the search landscape, and that also applies to government agencies.
Generative Engine Optimization, or GEO for short, is the art of structuring your content so that generative AI finds, understands, and correctly cites it. For government institutions, this is not a marketing issue: they are expected to provide official information. What if the citizen never comes to them to get it anymore?
From 1.5 to 1 million visits
The figures from Bart Van Herreweghe of the FPS Chancellery speak for themselves. “On Belgium.be, we have more than 1.5 million visits per month, and we are seeing a spectacular drop to 1 million per month.” Almost a third less, although he nuances: these are raw figures from August 2024 to August 2025, after which it remained stable. Furthermore, not everything can be attributed to AI with certainty.
Mathieu De Keyzer, Functional Analyst at Smals, also observes a decline per site. At Student@work, the click-through rate from Google halved from 12 to 6 percent. He immediately makes the link with the introduction of AI summaries by Google in the spring of 2025.
Spiders
While visitors are dropping off, the opposite is happening on the other side. “At a certain point, I had the impression that more than 80 percent of the traffic consisted of spiders, or pieces of software that collect information to feed AI, which were slurping up the content,” says De Keyzer.
The reflex of many IT departments is to block. Van Herreweghe warns that platforms such as Akamai and Cloudflare automatically blocked AI crawlers on various federal sites until he intervened himself. “We’re not going to block our own websites for AI, are we?” he asks.
SEO versus GEO
A new way to search the internet also requires a new strategy to be found. In addition to Search Engine Optimization (SEO), there is now also Generative Engine Optimization (GEO). “SEO determines whether you are found, GEO whether you are accurately mentioned after you have been found,” says Bert Jehoul, Innovation Designer at BOSA.
Where we used to fight for the click, we now fight for the mention.
Rafal Naczyk, web copywriter at Smals
Writing style is decisive
The art lies in feeding the AI correctly. Naczyk points to FALC, a European writing method aimed at making content as accessible as possible for all citizens, including those with specific needs. Think of citizens with a cognitive or intellectual disability.
This fragmented writing style has a simple sentence structure: subject, verb, direct object, with a maximum of ten to twelve words. “We realized that with FALC we also apply elements that make the text accessible to LLMs, because we more or less use the structure that feeds them best.”
The FALC writing style for citizens with specific needs is also accessible to LLMs.
Rafal Naczyk, Web Copywriter at Smals
What AI certainly does not appreciate are PDF files. Large, sometimes bilingual PDF documents are difficult for both humans and machines to process, and Google stops indexing them after a certain size. “It doesn’t cost much to create a PDF for an annual report, but it will be virtually invisible,” says De Keyzer. AI also appreciates question-and-answer structures.
“In Gov we trust”
Making information digestible for AI is one thing, but how do you convince a language model that you are the authoritative source? Van Herreweghe relies on ‘In Gov we trust’. “We already have certain tools available to increase the level of trust, such as the .be logo, a product and service catalog, and domain names, of which there are now almost 4,000, even without subdomains.”
De Keyzer agrees with Van Herreweghe. “For LLMs, clear authority indicators are important, including the domains .fgov.be or .belgium.be. Institutions that do not have these must switch to canonical URLs. This lets Google or the AIs know that this is the authentic source of the website, even if other domains exist.”
“In marketing, nothing beats repetition,” Naczyk adds. He is convinced that the more we make it clear to AIs which websites they should focus on, the more they will also label those sites as authoritative. Furthermore, he emphasizes the use of other websites. “It may sound contradictory, but it is worth linking certain content with websites such as Reddit, YouTube, or Wikipedia. These are usually pages that the AI will scan first.”
Conflict of interest
There are no guarantees. Jehoul refers to research by the Open Data Institute, which tested eleven LLMs on 22,000 questions about GOV.UK information: the models often provide incorrect information and, above all, are not to the point. “An LLM will always say: I know the answer. And if I don’t know it, I’ll just make it up.”
In addition, De Keyzer warns of a conflict of interest now that Google is placing its advertisements on its own AI pages instead of on websites. “Are we going to trust Google to put authentic sources in the foreground?”
Own chatbot?
Van Herreweghe is outspoken: “My personal opinion is that there should be a bot on every government website that is managed by the government and that does not hallucinate. That is perfectly possible today.” He declares the portal site era over: focus on your own information and let MCP protocols make the connections.
Jehoul puts an example on the table: the Pulse.Chatbot from BOSA. “We very consciously chose not to have generated answers on the spot. A contact center employee is not going to invent answers on the spot either.” According to him, the added value lies in personalized advice after authentication, linked to e-Box so that the answer becomes verifiable. To illustrate: “When you log in to MyPension, for example, you could ask questions to the chatbot after logging in, which can then provide personalized information.”
I think Google has more information about you than the government.
Bart Van Herreweghe of the FPS Chancellery
Van Herreweghe fears that resistance may arise around such bots, but he turns the ‘Big Brother’ objection around: “I think Google has more information about you than the government,” he points out. Private chatbots are already profiling citizens today without any authentication. Without a federal strategy, such a government bot risks having gaps.
Maintain SEO habits
AI has reversed the findability of websites. Where the search engine used to refer visitors, the assistant keeps them and decides for itself which source to cite. Naczyk refers to McKinsey research that speaks of 20 to 50 percent of organic clicks disappearing. For government sites, which are supposed to provide official information, the task thus shifts from attracting visitors to being recognized as an authoritative source.
The remedy is well-known and not exotic: let crawlers in, establish your authority with canonical URLs and semantic markup, rewrite your most important pages in short, structured language, and prioritize text over PDF. All old SEO hygiene thus becomes more relevant than ever, according to De Keyzer.
Content that is structured enough to feed an LLM is also the content that the citizen understands fastest. “That is ultimately a win-win situation,” Naczyk concludes.
This editorial contribution was created in collaboration with our partner Smals. The video of the roundtable can be found here.
