We have been talking about user-centred design for years. Today it is time to talk about design centred on all users. Artificial ones included.
When you design a website, you make decisions that affect much more than aesthetics: visual hierarchy, contrast, content structure, HTML semantics. Decisions we have so far measured in terms of UX and conversion.
Today those same decisions determine whether ChatGPT, Perplexity or Google AI cite you when someone asks about something related to your business.
Design has always spoken two languages. Now it has a third.
Good web design speaks to the people who use it. But it has always spoken to machines too: Google’s crawlers read your HTML, interpret your heading hierarchy and assess whether your structure makes sense.
Now there is a third party in the conversation: language models. And they read your website exactly like a screen reader does: without seeing the colours, without perceiving the spacing, without understanding that the big bold text “looks like” a title if it is not marked up as an H1.
The data confirms it. Traffic referred by AIs grew by 527% between January and May 2025. ChatGPT already has more than 800 million weekly users, and Google AI Overviews appears in at least 16% of all searches. And when an AI does cite you, the visitors who arrive convert 4.4 times more than traditional organic traffic.
Visibility in AIs is no longer a marketing problem. It is a design problem.
Which design decisions directly affect your visibility in AIs
Visual hierarchy is not the same as semantic hierarchy
That huge headline you designed in Figma at 64px and weight 800 is beautiful. But if in the HTML it is a styled <div> instead of an <h1>, to an AI it does not exist as a headline. It cannot understand what the page is about.
AIs build their understanding of content from its semantic structure, not its look. Properly nested H1, H2 and H3 headings tell the model what is important, what is secondary and how the content fits together.
Design decision: always check that the visual hierarchy matches the semantic one. What looks like a heading must be one in the code.
Contrast is not just accessibility. It is legibility for machines
The WCAG 2.1 AA standard requires a minimum contrast ratio of 4.5:1 for normal text. Many designs fail it with those elegant greys on white backgrounds.
But beyond legal compliance, low contrast is a sign of technically low-quality content. Language models that assess the quality of a page take technical accessibility signals into account as indicators of how trustworthy the site is.
Design decision: colour cannot be the only carrier of information. Ever.
Lack of consistency
- Each AI-generated module may lack integration and an overall vision.
- Design decisions are neither documented nor formally captured.
- It makes it harder for the project to evolve and for new members to join.
In short, Vibe Coding is useful for very simple projects or prototypes, but not for software that has to be maintained, scaled or held to professional quality standards.
Images without a description are invisible content
A hero with a striking photo, an infographic with key data, an explanatory diagram. Visually powerful. To an AI, non-existent if they have no descriptive alt text.
Wikipedia accounts for 47.9% of the sources most cited by ChatGPT when answering factual questions. What does Wikipedia have that your website does not? Structured text content, precise descriptions and semantic context in every element.
Design decision: alt text is not a technical field for the developer to fill in. It is content, designed by the content team together with the visuals.
Forms without labels are broken UX and broken GEO
An email field with only a placeholder and no visible label may look cleaner. But it is an accessibility failure (WCAG 1.3.1) and a negative signal of technical quality for AI crawlers.
Whatever breaks the experience for screen reader users also breaks language models’ trust in the quality of your site.
Design decision: labels are part of the design, not a technical extra. Design them visible or properly associated; do not remove them for looks.
Information architecture is your calling card for AIs
How you organise your content, how you title it, whether you use Schema.org structured data, whether you have a clear “about us” page with expert authorship… all of these are E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness) that language models use to decide whether your content deserves to be cited.
After implementing GEO best practices, a B2B company got 10% of its organic visits to come from generative engines within 90 days, with 27% of that traffic becoming qualified leads.
Design decision: information architecture is not just navigation and UX. It is the structure that tells AIs who you are and why they should cite you.
Accessibility and GEO: two problems, one solution
Web accessibility (WCAG 2.1 AA) and GEO share the same design principle: structural clarity benefits every reader, whether human, assistive technology or language model.
An accessible approach is not only more inclusive: it is also better prepared to be interpreted by AIs. These are not two separate lines of work, but one investment with a double return.
The web has always had the responsibility of making information understandable. Today that responsibility is growing: we no longer think only about people, but also about systems that filter, prioritise and decide which content deserves visibility.
And now, more than ever, good design is still the answer.