Ordinary people don't like websites. Why would they?

Most websites are crushingly bad at even basic aspects of delivery - especially findability and readability.

As soon as users discover an easier way to get what they want, we'll never see them again.

AI is just what they're looking for.

The simple truth is that Agentic AI is getting very, very good at interacting with websites.

Not perfect, but very impressive.

As it improves, delegating web engagement to an Agentic AI will be a no-brainer for most people. Imagine it. No more need to trawl through vast, confusing megamenus and impenetrable gobbledegook

The AI will do all the hard work and simply present you with results in a clear and easy to understand manner - typically much, much better than the source website itself.

This will have a profound impact on the discipline of Content Design. For instance:

• Will content that was designed for people continue to work when used by AI?
• Will we need to start optimising content for AI instead of people?
• Or will AI itself innovate new types of Content Design?
• If so, how can we influence that?

These are big questions.

But answers are emerging.

Testing Agentic AI on common Content Design patterns

In this article I share findings and screen-recordings from a series of content tests I recently conducted using Operator, the Agentic AI from Open AI.

My aim was to establish how well an Agentic AI can engage with traditional forms of content produced for direct human interaction - and what needs to change for a new era when engagement is intermediated by AI.

To do so, I assessed Operator against the 4 forms of content listed below.

1. Unitary content: This form of content is for when there is only 1 answer to a user need and no other possible response.
2. Series content: This form of content is for when the right answer to a user need comes from a predefined set of responses.
3. Algorithmic content: This form of content is for when the answer to a user need comes from a predefined set, but the output depends on the user.
4. Instructional content: This form of content is for when there is a known system or procedure that the user must follow.

Find out more about these in 'The 8 basic forms of content design'.

As you can see in the summary below, the results from my tests were ... let's say, mixed.

• Unitary: 5-star performance on Agentic AI    ⭐⭐⭐⭐⭐
• Series: 4-star performance on Agentic AI    ⭐⭐⭐⭐
• Algorithmic: 3-star performance on Agentic AI    ⭐⭐⭐
• Instructional: 1-star performance on Agentic AI    ⭐

Let's find out more.

Tests 1 & 2: Agentic AI engaging with examples of Unitary & Series content

(Learn more about the Unitary and Series forms of content.)

As you can see in the attached video, Operator performed quite well with these very simple forms of digital content.

(The video below begins at timestamp 16m:10s where Operator engages with the Unitary and Series content.)

And indeed why wouldn't it.

Information in the Unitary and Series forms is a very natural type for a machine to engage with. It's just plain old data and can easily be interpreted and presented by the AI.

In fact, this type of content has already been long presented in standard search as zero-click results in Google. For zero-click results, you don't need to visit a website to get what you need. The answers is given in search results.

Of course, success for zero-click content depends largely (or better to say entirely) on accurate, up-to-date information and a well structured page, e.g. built with good presentation, clear labelling and little ambiguity.

Although Large Language Models (LLMs) and Generative AI rely mostly on words, the under the hood stuff adds extra power, such as semantic HTML and structured data.

The presentation layer also plays an important role for Agentic AI, just as it does for us as people. Indeed, the quality of User Interface (UI) seems to be central at this point of AI development.

Just like us, an Agentic AI relies on features, cues, text and other UI elements to figure things out and arrive at the right answer.

Test 3: Agentic AI engaging with an example of Algorithmic content

(Learn more about the Algorithmic form of content.)

Next, I looked at content built using the Algorithmic form. To do so, I took the example of the Irish Citizenship Checker delivered by the Immigration Service.

(The video below begins at timestamp 17m:28s where Operator engages with the Algorithmic content.)

To be honest, I thought that the AI would really struggle with this.

I expected that it would struggle both with reasoning its way through the conceptual model and the UI, even though it is relatively simple.

And yet it excelled.

Of course, I had equipped Operator with some good background about a putative person looking for citizenship (a person born in Australia with an Irish granny). Nevertheless, it was able to run with this information and it did a great job.

So, it's a thumbs up then? Well, not quite.

In fact, despite showing that this form of content (that was specifically built for people) can also work for an AI, something about it just seems wrong .

Let me put it this way.

If you want a robot car, you don't build a robot to drive a car. You build a robot car.

Agentic AI kind of feels the same.

Although the AI can use this form of Content Design, it doesn't need it.

It's simply odd to watch a machine engage with a UI that was obviously built for people. It's clearly very cumbersome when a simpler approach would work better.

As such, I predict a future when the Algorithmic form of content is almost entirely removed from websites and replaced by the long descriptive form of content. Interstingly, this type of content is often of least help to real people, but could be of great help to AI which love straightforward narrative text.

That said, I also believe that the decision-type UI will remain. it's just that you or I as Content Designers won't create it.

Instead, we'll create the long form content for the source website, e.g. all the rules and requirements for citizenship.

The AI will then read and process that text and, when necessary, generate the algorithmic type interface on the fly.

Test 4: Agentic AI engaging an example of Instructional content

(Learn more about the Instructional form of content.)

And finally onto the Instructional form.

Here again my expectations were confounded. I had expected the AI to do quite well, given how orderly this type of content is.

(The video below begins at timestamp 19m:59s where Operator engages with the Instructional content.)

And in a way it did get the basics right. Operator seemed to have no problem recognising the sequential nature of the information and relayed it quite well to the user.

However, it failed at something the instructional form is specifically designed to do. It only presented a summary of the content, not the full set of instructions.

As explained elsewhere, the Instructional form is for cases where there is a system or procedure that the user must follow. The procedures are not up for debate. If you miss a step, you will fail.

But just as the procedures of the Instructional form are non-negotiable, its delivery must also be non-negotiable.

That is, the user needs all the information. The AI can't cut it short or paraphrase it. Everything is needed.

Yet, Generative AI is biased towards summaries. Plus, the longer the text, the more chance of creating a hallucination.

So definitely more work is needed on this form.

AI needs to learn when concealing complexity works for users and when it doesn't.

A core part of what we as Content Designers do is reduce dense information into easy to follow steps. However, we also know that sometimes the complexity is the product.

Sure, we make it easy for users to traverse by presenting in consumable chunks with clear language, etc - but we don't cut corners.

AI has to learn to do that too.

And that opens the interesting question. Can AI learn to do Content Design?

Can AI learn to do Content Design?

Before we answer that, let's pause to check out findings.

We've seen that some forms of Content Design that have been created for direct human interaction also work quite well for Agentic AI - especially the Unitary and Series forms.

No real change is needed for those at the moment, apart from greater attention to UI, labelling, sematic markup and use of structured content.

In this way, AI optimisation will become the new SEO or GEO as it is being called.

However, I believe other content patterns will be superseded as Agentic AI comes to intermediate web engagement for many people.

The Algorithmic form is a good example.

In future, we as Content Designers just won't need to design content into that form. Detailed textual descriptions will work just fine.

The AI will do the work of presenting a decision-type UI to users if needed.

Lastly, there are cases where a lot more work is needed. The Instructional form is the clear instance.

Agentic AI needs to be trained not just how but why content is presented in varying forms. That is, to help human beings with inherently limited working memory to engage with complex information.

This will require a sophisticated consideration of information intent as well as human capacity.

But, I do expect it to happen.

Your website is not an art project - it's a machine for doing things

It's easy for AI to be a toy, but I want it to be a machine that'll actually do some work.

That's what we as Content Designers have done. We've worked hard to help ordinary people find, read, understand and take action on the information they want.

I have a saying that I've used with every web team I've worked with.

"Your website is not an art project. It's a machine for doing things!"

That motto is plastered on my personal homepage.

We Content Designers are brilliant at examining, rethinking and enhancing user journeys to create content that allows people to actually do things.

AI can do that too, but it will partly depend on us as Content Designers.

In the first instance, AIs are going to rely on us for "sustenance" in the form of accurate up-to-date information on source websites.

In a way, we're the content chefs. We make the meal that AIs consume and so we get to set the menu.

AI indigestion

If we really wanted to we could AI severe indigestion.

We could craft our content so AI will really struggle intermediating it to users.

It would be possible to make the AI experience so limited and so unsatisfactory that we force people to come home to our old school websites.

But I don't think that's what we're about.

I remember having a conversation along these lines a number of years ago when implementing the structured data format for FAQs for the first time.

I was somewhat ashamed that my site had hundreds of FAQs. The least I could do was add some markup so results could be directly presented in search.

"But Shane. If there's zero clicks, that means people won't visit the site and traffic will go down!"

That's correct. But users will have the information they want and that's what we're about, right?

Anyway, a curious thing that FAQs may be one of the best forms of content to deliver to AI. LLMs just love this straightforward Q&A form and gobble it up.

That's a hell of a slap in the face for us!

The newest and most sophisticated form of technology leads to the redemption of the oldest and most despised form of content.

There's something cosmically comical about that.

Half-website / Half-database

The potential redemption of FAQs shows just how much we need to rethink Content Design as AI gets going.

As we have already seen, not everything that's needed for real people is needed for an AI.

For example, UI elements that are useful for people often just get in the way of Agentic AI.

This especially includes things that obscure content, such as opening/closing accordion panels. Added to this are overlays, popups and interstitials.

They may all have to go.

If you want my vision of the future (wrong as it almost certainly is) the OpenData-style of website presentation seems most likely, e.g. www.data.gov.uk.

That is, websites will be plain and boring but packed full of really useful information, with a simple user interface that is highly optimised to be machine readable. Half website, half-database.

A good example is the website of the Irish Heath Service.

Interestingly, optimising for AI may also ensure that we finally deliver fully accessible websites. That's because using ARIA and WAI standards with clear structure and labelling, can really help guide an Agentic AI around intricate web interfaces.

The machine phase of web begins

In short, the fundamentals for creating successful Content Designs aimed at AI are already in place. It just depends on what we want next.

Sure, not everything on a website should be served via AI.

There will always be some need for which the right answer is a regular website - or an email, a phone call, a text, a letter or a face-to-face meeting, etc.

However, that doesn't change my belief that ordinary people just don't like using websites and will avoid them if they can.

One other thing is also possible.

If websites are rarely looked at in future and adopt a more OpenData style of plain presentation, it also means that visual design will become less important.

And so finally, after decades of bitter and impassioned argument, no one will care any more what the bloody homepage looks like.

At that point, the art project phase of web will end. And the machine phase will begin.

It'll take some adjusting. But that's what Content Designers are good at.