Designing Websites for AI Search: What's Changing in 2026

There is a lot of confident advice circulating about redesigning your website for AI search, and a striking amount of it describes things Google has publicly stated it ignores. Meanwhile the change that is genuinely happening, and that genuinely affects how you build pages, is barely being discussed. So let's separate the two, using what Google itself has actually published rather than what the industry has decided must be true.

The headline is simpler than the panic suggests. AI search hasn't changed what makes a good website. It has changed who is reading it. Your site now has three distinct audiences, and only one of them is a person.

Your website now has three readers

For twenty years, web design optimised for two: the human visitor, and the crawler that indexed your page so humans could find it. Those two have been in rough alignment, which is why "build it for people" has been reliable advice.

2026 adds a third. The agent. Google's guidance describes autonomous systems that visit your site to perform tasks on someone's behalf, and it's specific about how they perceive a page: by analysing visual renderings such as screenshots, inspecting the DOM structure, and interpreting the accessibility tree. That last item deserves a pause, because it quietly rewrites a piece of design orthodoxy.

The insight If agents read your site through the accessibility tree, then accessibility is no longer only a compliance obligation or an ethical one. It has become a distribution channel. The screen-reader users you were building for and the AI agents deciding whether to recommend you are reading the same underlying structure.

This is the most useful reframe available right now, because it converts a vague anxiety ("how do I optimise for AI?") into work your team already knows how to do. Semantic markup. Real buttons instead of clickable divs. Labelled form fields. Sensible heading order. Alt text that says something. All of it was already good practice, and all of it is now legible to a machine that may be deciding whether your business is the answer.

What actually changed in 2026

Three things, and they matter to designers more than to marketers.

1. AI answers became the default surface, not a feature

Google reported at its 2026 I/O that AI Mode passed a billion monthly users, with usage more than doubling each quarter since launch. Whatever your view of that, the practical consequence is that a large share of your audience now meets a synthesised answer before they meet your page. Your homepage is no longer the first impression. A paragraph of yours, quoted inside someone else's interface, is.

2. Search began generating its own interface

The genuinely radical announcement was generative UI: Search now assembles custom layouts on the fly, building tables, visual tools, comparisons and even small interactive apps in response to a question. Sit with the implication. The interface the user experiences may be built by Google, from your content, without your design. Your careful layout, your considered typography, your CTA placement, none of it necessarily survives the trip.

That doesn't make design worthless, but it does change its job. Your page must now work both as a destination (for the humans who do click) and as a source (a thing that can be cleanly extracted, understood, and attributed). Designing purely for the first and ignoring the second is how sites become invisible while still technically ranking.

3. Agents started arriving as visitors

Search agents that run in the background, comparison agents, booking agents that complete tasks on a user's behalf. If an agent can't parse your booking form, it will complete a competitor's instead, and no human will ever know you were considered and rejected. This is a conversion problem disguised as a technical one.

What Google says you can safely ignore

Now the useful part, and the reason to read primary sources rather than the fourteenth blog post recycling the thirteenth. Google's official guide to optimising for generative AI features includes an explicit mythbusting section. Several of the most popular pieces of AI-search advice appear in it, as things that don't work.

A necessary caveat on the fourth row, because it's the one most likely to be misread: structured data is still worth doing. It makes you eligible for rich results and remains sound practice, as we cover in our on-page SEO checklist. The claim being debunked is narrower: it is not the secret entry ticket to AI answers that many guides imply. Don't rip it out. Just stop expecting it to be the thing that saves you.

Also worth noting, Google's own position on the acronyms: from its perspective, optimising for generative AI search is optimising for search, and thus still SEO. That's a slightly self-serving framing, and the distinct discipline described in our generative engine optimisation playbook is still worth understanding, not least because Google is not the only AI surface your buyers use. But it does tell you something important: there is no separate secret ruleset. There's the same ruleset, applied with more discipline.

The design checklist that actually matters

Strip away the hacks and a short, unglamorous list remains. This is the work.

Notice what isn't on that list: novelty. No new file formats, no AI-specific markup, no reformatting your prose into machine-shaped fragments. The 2026 design brief is mostly a demand that you finally do properly the things you've been half-doing since 2015.

The one genuinely new discipline: designing for agents

If your site involves any transactional flow, this is where to spend new effort. An agent trying to book, compare, or purchase on a user's behalf needs a page it can operate. In practice:

• Use native controls. A <button> is operable; a div with a click handler and no accessible name is a dead end.
• Label everything. If a field's purpose is only conveyed by placement or colour, an agent can't infer it. Neither can a screen reader, which was always the point.
• Keep flows predictable. Exotic custom widgets, gesture-only interactions, and content that appears only on hover are agent-hostile.
• Don't hide key facts in images. Price, availability, and specifications rendered as pictures are invisible to a text-based parse.

None of this is exotic. It is, almost line for line, an accessibility audit, which is the quiet joke at the centre of AI-era web design: the most future-proof thing you can do is the thing you were supposed to be doing anyway.

Content design: the commodity problem

Design can only carry you so far if the content is replaceable. Google's guidance is unusually blunt here, drawing a line between commodity content, the generic listicle that could have come from anyone, and non-commodity content built on first-hand experience and a genuine point of view. Its own illustration contrasts a generic tips post with a specific, experience-led piece about a decision the author actually made and what it cost them.

The logic is brutal but fair: if a language model could have produced your page from general knowledge, it doesn't need your page. It'll produce it. Your only durable protection is saying something a model couldn't have said, which means original data, real experience, actual opinions, and specifics that only you possess.

One structural note that changes how you plan pages. Google describes query fan-out: a single question gets expanded into a set of related queries behind the scenes. The instinctive response is to build a page for every variation, which is precisely wrong, and Google names it as scaled content abuse. The correct response is the opposite: build fewer, deeper pages that comprehensively answer a topic and its neighbours, which is exactly the compounding content strategy that was already the better bet.

The trade nobody enjoys

Be clear-eyed about what you're optimising for. Much of this work makes you more likely to be quoted in an answer, and being quoted does not guarantee a click. You may become more influential and less trafficked at the same time, which is a genuinely uncomfortable position to defend in a reporting meeting, and precisely why zero-click content demands new success metrics.

It also means falling organic traffic no longer has a single diagnosis. You might have lost rankings. Or you might be ranking perfectly, being cited constantly, and simply not being clicked. Those require opposite responses, and telling them apart is a job for a proper end-to-end funnel audit rather than a panicked redesign. If the fix does turn out to be structural, that's what website design and development work is for.

The bottom line

Ignore the hacks. Google has said in writing that llms.txt does nothing for it, that chunking isn't required, that you don't need to rewrite your copy for machines, and that structured data isn't the price of entry to AI answers. What remains is the unglamorous truth: be indexable and snippet-eligible, put your content in the HTML rather than behind JavaScript, use semantic and accessible markup, make your main content obvious, be fast, and publish things a model couldn't have written itself. The one genuinely new discipline is designing for agents that operate your site rather than merely read it, and that work turns out to be indistinguishable from accessibility. Which is the real lesson of 2026: the web didn't need a new playbook. It needed us to take the old one seriously, now that the cost of ignoring it is being left out of the answer entirely.