Are we still designing for humans?

Today’s post is a collaboration between myself, Tom Greenwood, and Michael Andersen, a fellow digital sustainability advocate and author of the book Sustainable Web Design in 20 Lessons.

Michael got in touch with me recently with a rather interesting question — if people are increasingly getting their information directly from AI services like ChatGPT, Claude and Grok, should we still be designing websites for humans? Put another way, is the primary audience for a website in the future going to be non-human intelligence?

This is a rather provocative question, and as Michael and I both care deeply about the environmental and social impact of the web, it is one that we should not ignore.

Of course, the idea that we might have to consider robots when designing a website is not entirely new. There has long been a tension between the need (and desire) to provide the best experience to human users while also needing to optimise websites for higher search rankings. The robots have been exerting power over us for a long time already. However, until recently the robots were simply the gatekeepers controlling who actually finds their way into our websites. With the rise of AI services, the robots are stealing our content and locking the gates behind them. 

This poses some interesting challenges for any person or organisation that has a website, and in the process forces us to ask some deeper questions about what we actually want the Internet to be moving forward. It’s easy to jump to the conclusion that this is an inherently bad thing, and it’s certainly not all good, but Michael and I want to hold space to consider the challenges alongside the possibilities with open minds, to imagine how we might evolve the web positively for this new era.

In recent times we’ve seen a number of clients at Wholegrain Digital experience significant spikes in their web hosting traffic as a result of AI bots constantly crawling their pages for information. This can come at a real financial cost to the client, not to mention the environmental impact, and it raises the question over whether we should block AI bots from our websites.

At the same time though, Michael highlights that some organisations have seen significant drops in web traffic coinciding with the rising popularity of AI services such as ChatGPT. At Wholegrain we have seen ChatGPT appear as a new traffic source for some websites, but at this stage it’s mostly a small one, as fewer people tend to click back to websites from their AI answers compared to traditional search engine results. Not to mention the fact that Google now provides AI answers at the top of its search results these days, reducing the need for the website “visitor” to um… visit. 

So we have AI companies causing a significant increase in artificial traffic to websites, while also reducing the human traffic to those same websites. And yet the place where these AI tools get their information from are these very same websites. Websites that people have invested a lot of time and money into designing, building, and maintaining, as well as creating content for.

This article isn’t about arguing the injustice of this situation, but to sit with it and acknowledge that we are in something of a paradox. AI services need a thriving Internet full of rich information in order to do what they do, yet they simultaneously erode the incentive for people and organisations to invest in creating a thriving web. The snake is eating its own tail.

In our discussions, Michael highlighted a key feature of this changing dynamic being that websites are increasingly treated as sources of information, rather than destinations to visit. If AI companies have their way, much like social media, people will spend most of their time on these new platforms and not leave. In this sense they really don’t care about your website other than it being a source of data to mine. 

But it’s likely that if you have a website, then you do care that people visit. You likely have reasons, whether commercial or not, to want real human beings to visit your website. You could, as some are doing, block AI crawlers from your website to discourage them from mining your website and attempt to ensure that the information on your website is really only available by visiting your website. This may work for some, but as people increasingly use AI tools as search engines, blocking AI crawlers entirely is not dissimilar to blocking Google from your website. It’s a bold and admirable ethical stance, but it might not be very practical in many cases. So what can we do instead?

My personal view is that we may need to design hybrid websites, facing the tension head on by separating the needs of humans from the “needs” of robots, but neglecting neither. We can compartmentalise, at least partially, the destination for humans, and the source of data for machines. In doing so, we might actually create a win-win and unleash a new wave of human creativity in web design, reversing the long term trend of websites “all looking the same.” Let me explain.

One of the things that this new dynamic highlights is that some people go online in search of answers to questions, and LLMs are increasingly taking this role. However, the user experience of LLMs tends to come in the form of long text based responses, or a digital voice speaking these same answers. What they don’t provide is the rich, immersive experience that has for a long time been a significant part of what people enjoy about going online. It could be argued that in the past decade or so, as User Experience “best practices” have become more firmly established and websites have been optimised for search rankings, the web has become increasingly bland compared to its earlier days when website design was a truly creative practice.

We now have an opportunity to reverse that trend by prioritising the human visitor not just in terms of ease of use, but in terms of the richness of experience that we offer people. Websites that are truly beautiful, immersive, joyful and thought provoking in their design, and which people want to visit because they are places that people actually want to be rather than just sources of dry information.

This also includes the tone of voice of the content, which is generally very bland in LLM tools but can be much more expressive on your own website, especially if there is now less need to awkwardly jam our content full of SEO keywords. This will allow us to enrich people’s experiences and convey the spirit of our brands in ways that an LLM likely won’t, making our own websites a more authentic expression of what we really have to say.

We could compare this to the shift we are seeing in high streets, whereby generic, uninspiring shops have struggled to compete with online retailers, but it has left an unmet need for unique, enjoyable experiences that people cannot get from shopping on Amazon. When high street retailers shift their perspective to see themselves as destinations to be experienced rather than just places to buy things, people do make the effort to leave their homes and go into town. Similarly, when we view websites this way and design experiences that people genuinely enjoy visiting and even want to share with their friends, we create something that the LLMs cannot replicate, even if they steal some of our content.

Of course, my comparison with high streets runs out here as some of the organisations that have the greatest opportunity to do this are likely those with some element of e-commerce. There is an inherent opportunity to inspire potential customers with rich visual information about the products, and interactive tools to help them understand products better and choose what is right for them. Of course we want to do this in a way that is accessible and environmentally efficient, but there are huge opportunities here for creativity. To quote my colleague Chânelle Sharp, a designer at Wholegrain, “the design of websites will shift to work even harder for the user now”. I think that is a good thing, with web design in the near future becoming much more about digital storytelling.

People are time poor and so instead of wasting their time with large reams of text intended mostly for Google, user experiences could actually cut back significantly on written text in many cases and just give people exactly the information they need, in a beautiful and engaging format. And despite the downsides of these AI tools, it may just be that AI coding tools in the long term will make it easier for organisations to create rich functionality and design features such as visualisations, personalisations, dashboard, calculators, and interactive journeys, that were previously beyond their budgets. It may well be therefore that AI coding will gradually reduce one of the key limiting factors in doing really exciting things online.

If we design websites primarily for humans once more and relegate the AI and search crawlers to second class citizens of the web, then you may be wondering, how do you get your website featured in search results and AI responses? I think the answer might be to have a secondary layer of a website that is purely informational, with large repositories of high quality written information and raw data relating to your organisation and area of expertise. These repositories can include high quality original data, white papers, academic research and the like, making them a rich source of quotable information for the LLMs. These pages can be crawled by crawlers without getting in the way of human visitors, and to quote Chânelle again, “if the information is purely informational, then maybe it doesn’t even need to be designed”. This second layer of a website could be super efficient by virtue of not being mixed in with the presentation elements such as imagery and custom fonts that are there only to improve the experience of humans.

But what if your visitors do want to access some of this information from your data repository while they are on your website? This is where a more efficient LLM tool tailored just to your website may provide the answer. Tools such as HOLP, which we used for the UK Green Building Council, can provide answers to the questions of the people who have arrived at your website, without them having to leave your website and unlike mass market AI tools, they can provide trusted answers as they are trained to provide information from your own websites information repository (as well as its user facing web content). This is a win for the human visitor who is seeking specific information that may not be easy to find in the standard user facing content, and for the organisation who is able to meet the unique needs of such visitors without complicating the user experience for other people.

I’ll hand over to Michael now to discuss how some of this might be achieved on a practical level.

I want to begin by saying how excited I have been to write this alongside Tom. Ever since I first read his book Sustainable Web Design, he has been a constant source of inspiration to me. The hybrid idea he lands on, a light and human surface sitting on top of a denser layer meant for machines, is the thread I want to follow, though I would not want to give either of us too much credit for it, since the web has been built in roughly this shape for a long time and most of us have simply taken it for granted. What feels worth examining now is not the layers themselves, but the question of who that hidden layer is increasingly being built for, and that is where my own corner of the work comes in.

That corner is the code itself, where I tend to sit rather than up above it looking at the strategy. Tom has been describing the shape of all this from above, the way you might look down at a city and see how the roads connect to one another, and I want to stay down at street level for the rest of this, because it is down there that the same question we kept circling stops being philosophical and turns into something you can put your hands on. It becomes a question about which file you create, and what you put in it, and who you imagine reading it. So where Tom has been holding the why of all this, I want to spend my half on the how, on what the hybrid he describes actually looks like when you are the one who has to build it.

If you have ever sat and built a website properly, not just made it look nice but actually built it so that it holds together, then you already know that a page is never only the thing the visitor sees. Underneath the colours and the images and the words arranged on the screen, there is a second version of the same page that almost no human ever really looks at. There is structured data, little labelled notes in the markup that say this is the price, this is the author, this is the date, this is the review, and here is how many stars it got. There is JSON sitting in the background, describing the page in a way that is clean and predictable and easy for software to read. And none of this is particularly new to anyone who builds for the web, because we have been doing it for years, mostly so that search engines could understand our pages well enough to rank them, and most visitors have never had any idea it was there.

To make that a little less abstract, it helps to picture what one of those hidden notes actually looks like in practice. If I have written an article, I might tuck a small block of structured data into the page, written in a format called JSON-LD, which is really just a tidy way of listing plain facts so that a machine can read them without having to guess. In ordinary words, the block says something like, this thing is an article, its headline is this, the person who wrote it is this, it was published on this date, and it belongs to this larger publication. None of that changes a single pixel of what you see on the screen, and you could read the article perfectly well without ever knowing the block was there. But a machine that lands on the page no longer has to work out who the author is by hunting around the layout and hoping it has guessed correctly, because you have simply told it, in a vocabulary that thousands of other sites use in exactly the same way. That shared vocabulary is what people mean when they talk about schema, and the whole point of it is that meaning stops being something a machine has to reconstruct and becomes something you have stated plainly.

What is changing now is the kind of reader that layer is starting to serve. There is a proposed convention that appeared back in 2024, suggested by Jeremy Howard, who is one of the people behind Answer.AI, and it is called llms.txt. The idea is almost touchingly simple. You place a plain markdown file at the root of your website, the same way we have placed a robots.txt file there for decades, and in it you lay out, in clean and readable text, the parts of your site you actually want a language model to read. No clutter, no menus, no scripts, no advertising, just the meaning, served plainly.

And because it is only markdown, building one is genuinely something you could finish in an afternoon, which is part of why it interests me so much. You open a plain text file, you name it llms.txt, and you place it at the very root of your site so that it sits at a predictable address the way robots.txt always has. At the top you write the name of the site and a sentence or two describing what it is and who it is for. Then you write a short list of links to the pages that actually matter, your best writing, your core explanations, your reference material, and after each link you add a handful of words saying what a reader will find there. That is more or less the whole thing. If you want to go a step further there is a fuller version, often kept as a second file, where instead of only pointing at your important pages you include the full cleaned up text of them, so that a model has the real content in front of it rather than just a map showing where the content lives. Either way, what you are really doing is curating. You are deciding, on purpose and in advance, what you would want a machine to understand about your work if it only ever saw the tidy version and never visited the site at all.

I want to be honest about what it is and what it is not, because it would be easy to oversell it. It is not an official standard, and Howard himself has said as much. The big AI systems do not reliably read it yet, and by now more than one fairly large study has looked at the question and found that adding the file does not measurably change how often a model mentions you, with one analysis of around three hundred thousand sites finding essentially no difference at all. Google has even said, fairly plainly, that it has no plans to support it. And yet hundreds of thousands of sites have already added one, including a few names you would recognise, and people who have watched the web for a long time keep comparing it to the early days of schema markup, back when that too looked like a niche curiosity that later turned out to matter a great deal.

And this is where I want to put the practical heart of it, because it would be easy to read everything above and imagine that building Tom’s hybrid means building two entire websites, one warm and human and one stripped bare for the machines, two of everything and twice the work. You can do that, and for certain things you genuinely should, but for most of a site the truth turned out to be stranger and a good deal kinder than that. If you have already built the human layer properly, with markup that is clean and honest about what each thing is, with real headings and real buttons and images that are actually described, then you have already built most of the machine layer without ever thinking of it that way. What you add on top of that is thin. A little structured data to state plainly what the important things on the page are, and an llms.txt file at the root to hand a model the tidy version of what you most want it to read. That is the bulk of the work, and almost none of it is a second website. It is mostly just the first website.

Where the two layers do genuinely come apart is in the place Tom pointed to, the large stores of pure information, the white papers and the raw data and the long reference material that exist mainly to be read rather than to be experienced. For that kind of thing it really does make sense to build a separate and lighter layer, plain pages or plain markdown with almost no presentation wrapped around them, so that a crawler can read them cheaply and a person who actually wants the underlying detail can still reach it without having to wade through it on the main site. So when I try to picture the hybrid in concrete terms, it is not two parallel websites standing side by side at all. It is one human site built well enough that it serves the machines almost for free, sitting next to a small and efficient archive for the heavy reference material that nobody really wanted dressed up in the first place.

And this is the moment I really wanted to reach, because when you sit with all of this at once, you start to notice something a little strange, and a little hard to name. The future Tom was describing, the website that is one thing for people and another thing for machines, is in some sense already half built. We just never called it that. We built the human layer because we cared about people, and we built the machine layer because we wanted to be found, and now that machine layer is slowly turning into the thing that reads our work on the reader’s behalf. It is a bit like discovering that a path you have been clearing and maintaining for years leads somewhere you never expected it to go.

There is one corner of all this that I have not been able to stop thinking about since my call with Tom, and it has to do with accessibility. For most of my career, the work of making a website accessible has been understood as work we do for people. The Web Content Accessibility Guidelines, which most of us just call WCAG and which are looked after by the people at the W3C, exist so that someone who cannot see the screen, or cannot use a mouse, or who processes information differently, can still move through a website and understand it. In practice, when we talk about meeting those guidelines, we usually mean building to a level they call AA, which most teams treat as the realistic target, and the work of getting there is made up of a hundred small and very ordinary decisions. You write alternative text that genuinely describes what an image shows rather than leaving it empty. You use headings in a real order instead of choosing them for how big they happen to look on the screen.

You label the fields in a form so that it is clear what each one is asking for. You make sure every part of the page can be reached and operated with a keyboard alone, and that the text has enough contrast against its background to be read by someone whose eyes are tired or beginning to fail. None of it is glamorous, and almost none of it ever shows up in a screenshot, but it is the difference between a site a person can actually use and one that closes its door to them. Here in Europe that work has recently stopped being optional in a lot of cases, because the European Accessibility Act came into force in the summer of 2025 and now expects a great many digital products and services to meet exactly these guidelines. So accessibility has been on a lot of minds lately, and rightly so.

But here is the part that feels strange to me. The whole craft of accessibility has always been, underneath everything else, the craft of making meaning explicit. You write your markup so that a screen reader, which is itself a machine, can correctly understand what is a heading and what is a button and what an image is actually showing. You describe things in plain words so that software can relay them faithfully to a person who needs them in some other form. In other words, we have spent years and years learning to make our websites legible to a machine, because a machine was the thing standing between our content and the human who needed it. We just thought of that machine as an assistant working on behalf of a person, which it was.

And once you see it laid out like that, the overlap with everything I was describing earlier stops feeling like a coincidence and starts to look almost like the same task wearing two different faces. The alternative text you wrote for a blind visitor is meaning made explicit, and a language model wants that exact thing. The heading you marked up honestly so that a screen reader could announce it is structure made explicit, and a model trying to make sense of your page wants that too. The form field you labelled, the button you named properly, the landmark you added so that someone could skip straight to the main content, every one of those is intent stated plainly rather than left to be guessed, and plainly stated intent is precisely what this new reader is hungry for. Which means that in very practical terms, if you build a site to that AA level for the people who actually need it, you have quietly built a large part of your machine layer at the very same time, and the two reasons for doing the work, the human one and the technical one, turn out to be pointing in the same direction.

Now another kind of machine is arriving, the language model, and it turns out that much of the work we did to make the web kinder to people has also, almost by accident, made it more readable to this new reader. The structure, the explicit labels, the meaning made plain, all of the things we built for the person using assistive technology turn out to be more or less exactly the things a language model also wants. I find that both hopeful and slightly unsettling, and I do not think I am supposed to resolve which one it is. It is hopeful because it might mean that good, accessible, carefully made websites are simply ready for this new era without anyone having to compromise anything. And it is unsettling because there is something odd about realising that work you did for the most human of reasons also serves something that is not human at all.

And then it opens a question I genuinely do not know the answer to. If a language model can read the raw, structured meaning of a page and then reshape it into whatever form a particular person needs, spoken aloud, simplified, translated, slowed down, made calmer, then it is at least possible that AI becomes a better companion for people with accessibility needs than many of the tools we have today. The same shift could even help companies that have never been able to afford a genuinely good, accessible experience, because the model sitting in front of their content could, in principle, give their visitors a kinder journey than their own budget ever allowed. I want to hold that possibility gently rather than promise it, because we have all heard a great many promises about what AI is supposedly going to fix. But I would be lying if I said the thought did not move me a little, that the very thing reshaping the web might also, if we are careful and a bit lucky, make it gentler toward the people it has so often left behind.

I want to slow down here, because there is a worry underneath all of this that I have been turning over for a long while, and I am wary of how I say it, because it would be very easy to make it sound like I am complaining about the way people read now, and that is not what I mean at all.

At my job we have spent time talking with young people about how they find things out, and one of those conversations has stayed with me. A teenager said, more or less, that AI made the person lazy, that when something can simply be handed to the person, the person loses the urge to go and chase the answer. I want to be careful and call this what it honestly is, which is one young person being candid in one conversation, not a study and not a verdict. But it landed on something I had been feeling for a while without quite having the words for it. Because if the answer can be had instantly, without arriving anywhere at all, then a certain kind of journey simply stops happening. Not the journey through a checkout or a menu, but the one where you go looking for something, and you end up somewhere you did not plan to be, and you leave a little changed by having been there.

For a long time I think we treated that journey as part of the value of the web, even though we never wrote it down anywhere or put it in a proposal to a client. You arrived at a site, you read more than you came for, you understood the people behind it, you caught their tone of voice, you felt that there was a person or a team or a belief on the other side of the screen. That was never really about the information. The information was almost the excuse for the visit. And my worry is not that people have become lazy, but that we are building a version of the web where that kind of arrival is no longer necessary, at least for everything that can be reduced to a quick answer. I do not think that is the whole web, and I do not think it has to be a tragedy. But I think it is worth noticing what we might be letting go of while we are busy admiring everything we are gaining.

And I notice, sitting here at the practical end of all this, that my worry is really the shadow of Tom’s hope. He is hopeful that we will build places worth coming to, websites so rich and so genuinely worth the visit that people choose to arrive even when a machine could have spared them the trip. I share that hope, more than I probably let on. But the thing I keep turning over, down here where the pages actually get made, is whether the arriving itself is quietly being designed out of the web, and whether all the craft in the world can pull someone back to a place they no longer have any reason to go.

So I find myself ending more or less where Tom began, holding a handful of questions rather than answers, which I have come to think is the honest place to stand.

Are we still designing for people, or are we slowly becoming translators, standing in between humans and machines, trying to serve both at once without fully belonging to either? When a language model has already read everything on the page before the person ever arrives, what is the visit itself actually for, and what is left inside it that still matters? And if the model becomes the thing that shapes how our work is understood, I find myself wondering whether it might end up building better journeys through that work than we do, and whether, a little further down the same road, we will still need the style guides and the UX designers and the graphic content, or whether those become a luxury rather than a default.

I wonder too about the places that exist almost entirely to be read; the Wikipedias and the reference sites and the great open stores of knowledge, and whether they will one day be built in a wholly different way, written first as plain markdown for the machines rather than dressed up in markup for us.

I am not sure either, when it really comes to it, whether companies will simply hand the whole customer journey over to the AI sitting in front of their content, having decided that a more accessible and more helpful experience, delivered by a machine, is worth more to them than the elaborate one they once paid so much to design. And if plain text is becoming the shared layer that the machines hand out for free to anyone who asks, then what becomes the thing worth keeping precious, the thing we still make slowly and by hand precisely because it cannot simply be read out by something else and passed along? I do not have clean answers to any of these, and I am not sure I trust anyone who says they do.

What I do know is that the question Tom opened with, whether we are still designing for humans, is not really a question with a yes or a no waiting at the end of it. It is more like a question we are going to be living inside of for the next several years, whether we choose to or not.