Search traffic is collapsing, direct traffic is booming: here's my data

Love it or hate it, AI is having a big impact on the world. If you are involved in managing a website, you probably have noticed the indirect effect of AI in your website analytics data. One obvious way that this manifests itself is in declining levels of overall traffic.

Tools like Google’s AI search summaries are obliterating traffic for many websites as people find all the information they need without needing to click through (and that’s assuming that there is even a link to click on). A Pew Research Center study from last July found that users were much less likely to click on links when their Google search results had an AI summary compared to search results without one. 

Direct traffic

In a blog post from January, I wrote about the different ‘buckets’ of traffic that website analytics platforms record. This includes obvious things like ‘social media’ and ‘search engine’ traffic, but also the broader (and somewhat nebulous) category of ‘direct’ traffic. Historically, this was really referring to people who went directly to your website by using browser bookmarks, clicking on links in emails, or even just by typing a web address.

Direct traffic is effectively any traffic that cannot be put into any of the other buckets. My earlier post focused on the subset of direct traffic which has been referred to as ‘dark social’. This will include people who click on links in messaging apps like WhatsApp…these apps don’t leave any digital fingerprints so the data is recorded as ‘direct’.

AI traffic

However, traffic from AI tools is increasingly showing up in the ‘direct’ bucket too, and this is something that has been capturing a lot of attention recently, e.g. see:
‍ ‍

• The AI Traffic Attribution Crisis: Why Your Analytics Are Wrong (Updated With Real Data) - November 2025

• The 70% Problem: Your AI Traffic Is Hiding in GA4 Direct - June 2026

• Your AI Traffic Is Hiding: A Practical Guide to Tracking ChatGPT, Claude, and More - June 2026

Even though analytics platforms like Matomo are trying to now identify AI traffic, it’s not easy and people like me who spend a lot of time looking at website data, are finding it harder to get a clear picture of how websites are being used.

Here is some actual data from three websites that I manage. The chart below shows the percentage of traffic that is recorded as search engine or direct at three different time periods (the first six months of 2024, 2025, and 2026). The websites are anonymised. 

View fullsize

Click to enlarge

In all three cases, these websites have seen a big decline in the share of traffic coming from search engines and a corresponding big rise in the share of direct traffic. I’m only focusing on the share of traffic from different sources rather than the total level of traffic.

It’s possible that this rise in direct traffic has nothing to do with AI…but that’s the problem with direct traffic, it’s all unknowable.

If this change is largely driven by AI (which seems reasonable) then it means that the future of managing a website is going to be a very different experience. We might expect to see declining levels of overall traffic. We might also expect to see that the vast majority of traffic is direct traffic.

The quality of your traffic is going to matter much more than quantity but identifying high-quality vs low-quality traffic is another challenge altogether.

2,000 links and one new colour

It was recently World Refugee Day (20 June) and for much of the last six months I have been racing to help launch a new website project for the Careif charity where I volunteer my time.

We chose 20 June as our target launch date for CARE - the Careif Archive of Refugee Empowerment. CARE is a collection of resources relating to the mental health of refugees and asylum seekers aimed at both professionals working in these areas as well as those directly affected and in need of help. You can read more about it in the launch news story which I wrote.

I’ve been helping build these pages at Careif since last year. There is a lot of content (it brings together links to over 2,000 external resources) and it means that the number of pages at Careif has effectively tripled.

Building the site

Squarespace page folders

Although there were some talks about whether this might be a completely separate website to Careif, we decided to build it as part of the site.

That has led to a few challenges in working with the Squarespace CMS. Although Squarespace lets you arrange pages in folders, you can only have one level of hierarchy. The new CARE pages predominantly concern adult resources and child resources, but there are also other sections (e.g. resources for humanitarian workers).

I can create URLs for all of these pages that reflect this hierarchy, e.g.

• careif.org/care/adult-resources

• careif.org/care/child-resources

• careif.org/care/humanitarian-resources

However, Squarespace doesn’t let you recreate that hierarchy visually in the CMS. This is because, although you can create folders to help arrange pages, they can only be one level deep (I think the primary purpose of folders is to arrange pages that will appear in navigation menus). This may seem a minor point but I think it points to the challenges of using Squarespace if you are going to develop a big website.

So I ended up having to create three separate folders to make it easier to manage these pages, with one folder each for adult and child resources and another folder for everything else.

A new colour

When building the ‘new’ Careif website a few years ago, I had settled on a very simple colour palette which was based on the (pre-existing) Careif logo. As the CARE pages have greatly expanded the Careif site, I felt like we needed one new colour to be used on CARE pages which was distinct from the existing colours but which still complemented them.

Using one of the many online colour palette tools that you can find (I used the excellent coolors), I settled on ‘raspberry’ as the CARE accent colour. This is applied as an underline on page titles and also as a new secondary button colour.

Rethinking the content model

There is still work to be done in creating one more page of resources for CARE. In putting these pages together, I feel that ultimately all of the page content should be completely rebuilt as some form of database with all resources tagged/categorised appropriately. 

One possible solution for this in Squarespace would be to create every individual resource as its own blog post (with relevant tags/categories applied) and then use a third-party plugin to have a front-end that lets you search and filter all blog posts.

This would be a lot of work but could make managing the content easier as there would only be one content item (a blog post) for each resource, whereas currently some individual resources are duplicated across different pages.

Tackling link rot

Now that the site has grown so much, an additional challenge is to be better at checking for ‘link rot’. With over 2,000 links to external sites, we can expect those links to break regularly as sites shut down, get merged with other organisations, or just rearrange all of their content. This is a common challenge with managing any site that links to a lot of external material.

At RCPsych we use Dr Link Check to find broken links (there are various paid plans available) and I think we will try using this at Careif too now.

Time for a break

This has been a lot of work to get the CARE pages together and it is pleasing to see that they are already generating traffic. For me personally, this is now the third big piece of web development I’ve done outside of work this year. I’ve built an artist’s website for a friend (meganyelets.com), I’ve rebuilt this website and now I’ve helped launch CARE. I think I have earned a summer break now!

I’ve had the decorators in: new look to website plus accessibility improvements

Regular visitors to this blog — which is probably just me — might notice that things have changed a little bit. I’ve taken some time over the last week to update the look and feel of my site.

Outwardly, it’s not a very bold change. This website still consists of three pages, with the blog page as the main page that includes summaries of each post. I’ve ditched the ‘Read More’ links as each blog post title is also a link.

The most obvious visual change is that I have refreshed the look of the site and doubled down on red as an accent colour.

However, there are many more changes that have happened _behind_ the scenes, including many accessibility improvements.

Squarespace upgrade

The main reason I made any changes to the site is that this website has been running on Squarespace v7.0 for many years. The latest version of Squarespace is v7.1; while this might not sound very different, it’s quite a fundamental change.

Upgrading to 7.1 is a non-reversible action and after I made the change, it immediately changed how my site looked (for the worse). Squarespace 7.1 just does a lot of things differently and some of the things you could do in 7.0 just don’t exist any more.

I have spent many hours just making very small refinements to get the site broadly looking like it used to. Although this was a bit of a time sink, it made sense given that the charity site I manage (careif.org) and the site I built for my artist friend (meganyelets.com) are both running Squarespace, so it makes sense for all of these sites to be running on the same version.

Accessibility updates

One of the things I really wanted to do with this refresh is improve the accessibility of this site. This follows a lot of work I’ve been doing in my main job to help improve the accessibility of the RCPsych website.

This means that this website now includes:

• Improvements to hover states: you’ll see a yellow background when hovering over a link, the link text will darken and the underline will thicken.

• Improvements to focus states: keyboard users can tab through elements and see a black box border on links.

• Use of ‘active’ state: clicking on a link slightly darkens the yellow background colour to reinforce that a link has been clicked.

• Making use of visited link status: if you click on a link, then your browser will remember this and change the link colour to grey. This is not so much of an accessibility improvement, but could potentially help people with memory problems, providing a clearer indication of what they have previously clicked on.

Incidentally, removing all of the ‘Read More’ links on my blog landing page is another accessibility improvement. Link text should be unique on a page, as someone using a screen-reading device would otherwise struggle to distinguish between different link destinations if they all say the same thing.

These are the sorts of improvements that can take a bit of time to implement, but it’s a one-off investment that provides lasting benefits. You will never know which users of your website have disability requirements, so making it as accessible as possible removes barriers that prevent some users from benefiting from content.

Rekindling the joy of building a website

Art by Megan Yelets

Last Thursday I was waiting to see my 8-year old’s school play and was chatting to a parent friend who was standing in the line next to us.

She is a great artist who is currently showing her work at an exhibition in London, but she revealed that she didn’t have a website or own her own .com domain.

A couple of hours after the play ended I bought her domain name for her. I also offered - if she wanted - to build her a website for free.

I have quite a bit of experience using Squarespace to build websites (including this site) and I felt like this might be the sort of thing where a built-in template would probably meet most of her requirements without much need for tweaking.

We met the next day to discuss what options were available and less than a week later her website — meganyelets.com — is now live (built to her specifications)!

Building a site for an artist

This is a very different type of project compared to anything else I have done in my years of managing and building websites. After we had reviewed some other artist’s websites, Megan had requested that she wanted a big cover image on the home page that would lead into a gallery of some of her images.

It quickly became apparent to me that we needed the whole focus to be on the art itself. Consequently this led me to keep everything else relatively plain and unobtrusive.

I probably had the first draft of the site all complete with a couple of hours. I then spent many more hours doing lots of fine tweaking.

Frustratingly, I discovered that although Squarespace lets you add custom code and CSS while you are building the site (i.e. while you are effectively on a free plan) but as soon as the site goes live, your custom code will be removed if you are on their base level plan.

I only realised this after the site went live and so I then had to find a workaround as I didn’t think it was worth a £60 per year upgrade to the plan just to get this one extra feature.

Next steps

Megan is very happy with the results and I hope that she now has a platform to help showcase her lovely art to a wider audience.

As for me, there is still some work to be done and I hope to set her up with a newsletter service so she can more easily share updates with people interested in her work.

Why ‘dark social’ and AI traffic makes it harder than ever to make sense of website analytics

An increasing challenge for those of us working in the world of website analytics is the impact on traffic from Artificial Intelligence (AI) tools such as ChatGPT and Gemini. The traffic – or as I will come on to discuss, the lack of traffic – from such platforms is a relatively new phenomenon. However, it adds to other relatively recent challenges such as a rise in what is known as ‘direct’ traffic as well as the mysterious sounding source of traffic known as ‘dark social’.

Putting things in buckets

Let me back up a bit. Website analytics platforms routinely try to categorise traffic to web pages into different ‘buckets’ based on what they know about how someone arrived on a website.

Analytics platforms can tell you that a page view on your website originated from someone who arrived using a search engine, or from a public post on LinkedIn, or that your site was linked to from another website such as Wikipedia.

These different buckets all have special names, but the one that is occupying more of my time lately is the bucket called ‘direct’ traffic.

Historically, this captured traffic from people who went ‘directly’ to your website by typing an address in your browser or by clicking on a browser bookmark. But these days, the ‘direct’ bucket of traffic captures a lot more than that.

A better choice of name would be ‘untraceable’. If Google Analytics (and other website analytics platforms) can’t detect how the visitor came to your website, that traffic gets put in the ‘direct’ bucket.

This means that direct traffic can also includes things such as clicking on links in emails, or links in most messaging apps. Other sources of direct traffic include links on platforms like Apple News or links inside private Facebook groups.

The dark side clouds everything

‘Dark social’ is the broad (and nebulous) term that has emerged to describe a subset of this untraceable direct traffic. It was coined by Alexis Madrigal in 2012 in his blog post Dark social: we have the whole history of the web wrong.

It is typically used to describe traffic that might come from some social media platforms and messaging apps. WhatsApp is very likely a big source of ‘dark social’ traffic for many websites.

Because direct traffic is – by its very nature – untraceable, you can’t really ever know where it came from. Some tell tale signs of dark social traffic is that the traffic is to a specific page that might have been shared by someone, e.g. a news story or blog post.

It is also probably is more likely to have come from a mobile device as most social media is still consumed on phones. If you didn’t know, analytics platforms can track a lot of data about the technology used to visit a web page (desktop PC vs mobile phone, Mac vs Windows, Firefox vs Chrome etc.)

There are other signals that can be also used to potentially identify subsets of direct traffic but ultimately this is all just educated guessing and you can never know for sure.

Rise of the machines

AI is rapidly changing the landscape of how website traffic is captured. First and foremost, if you run an information-heavy website with lots of resources, then AI tools might be leading to a big decline in your traffic.

This is simply because many more people are getting their answers from an AI chatbot and no longer need to click through to a website to find out more. This is assuming that chatbots provide a link to your site which may not always be the case…and even if they do, that link might be very easy to miss.

A lot of AI tools leave a digital fingerprint that enables web analytics platforms to put them in the ‘referral’ bucket of traffic. I.e. where traffic to your site was a referral from another website or tool.

A recent update to the website analytics platform that I routinely use (Matomo Analytics) has started capturing this known AI traffic in a new ‘AI Assistants’ bucket which lets me see traffic from popular tools such as ChatGPT, Copilot, Perplexity etc.

AI Assistants traffic as it appears in Matomo Analytics

This traffic is growing rapidly but doesn’t neatly explain all AI-related traffic. If you use Google’s new AI mode in their search interface and then click through to a website, this is still captured in website analytics platforms as search engine traffic. But if you use a tool like Siri or Alexa to open a webpage, this might show up as direct traffic.

A tangled mess

All of this means that it is getting harder than ever to disentangle the various strands of traffic that bring people to your website. Dark social contributes an unknowable, and possibly increasing, subset of direct traffic and I imagine that the growth of WhatsApp is a big part of this.

AI tools - which are causing a reduction in traffic to many websites - might record what traffic they do generate in several different ways.

I’m finding that I’m increasingly using terms like ‘presumed WhatsApp’ traffic in some of my regular analytics reporting as the real answer about where our website traffic is coming from is increasingly ‘I don’t know’.

Regular recording of data is helpful as you can then start spotting trends that might reveal which areas of your website are seeing differences in traffic.

I’d also recommend using Google’s search console tool if you are concerned about AI tools displacing and replacing traditional search engine traffic. I’ve already noticed that the biggest drops in search engine traffic are to those sections of the website which are ‘informationally rich’. I.e. the content that is more likely to have been regurgitated as part of an AI tool’s answer to someone’s question.

Good luck with your website analytics detective work…it’s a jungle out there.