AI Harvesting of Information Lowers Traffic and Reduces Revenue for Websites

The administrators of many websites and blogs have complained about a decline in traffic (page views), especially over the last year. Office365ITPros.com is not immune to what’s happening across the internet. We’ve seen a drop of about 50% in traffic since the summer of 2023. There’s lots of theories being explored for why the drop in traffic happened. Thinking about the situation, I believe that technical, human, and disruptive change combine to deliver what we see today with the impact of generative AI on technology websites being felt and not in a good way.

Many point to the effect of Google Analytics 4 (GA4) and say that the results for website traffic reported by GA4 are markedly different to its predecessor (here’s an example). I don’t pretend to be an expert on Google Analytics, but when everyone’s talking about the impact of a change, it’s hard not to conclude that the introduction of GA4 has had some effect.

In addition, search engine result pages (SERPs) now include snippets of information that might be sufficient to answer user queries (Figure 1). If an answer is found in a snippet, there’s no need to follow the link to the source web site to find more information. Snippets are great for users while contributing to declining page views.

User behavior is also changing. More mobile devices are used, so if an article doesn’t seize the attention of the reader within the first few paragraphs, the user will move on and not follow links in the text. Another factor is that mobile devices can favor the consumption of video content rather than traditional articles.

But the biggest change affecting websites covering technology is the impact of generative AI tools like ChatGPT and Microsoft Copilot. AI tools build large language models (LLMs) using information from the internet and can regurgitate that knowledge in a more approachable fashion than regular search results (Figure 2).

I used Bing.com for the SERP example in Figure 1 because Microsoft Copilot uses Bing search to gather information for its responses. What you see in Figure 2 is the Copilot equivalent of the Bing result in Figure 1. Interestingly, Copilot dropped the reference to the article used in the Bing SERP. I wrote that article some years ago and it’s very outdated now. It seems like Copilot favored more recent articles in its output.

What we learn from this demonstration is that AI gathers the knowledge that people would have found through web searches and serves it up without the need to go anywhere near the source sites. The result is a dramatic fall in website traffic with the only page views counted by Google being those that occur when someone clicks a link in an AI response.

The Lesson of Stack Overflow

Stack Overflow is a stark example of what can happen to a thriving technology website. For years, Stack Overflow was the place for developers to go when they had a coding problem or needed an example to know how to do something. Then ChatGPT came along and the effect on the traffic handled by Stack Overflow was dramatic (Figure 3).

Simply because technologists adopt new technology faster than the general public, technology sites were always likely to experience an impact as generative AI began to have an effect. In the case of Stack Overflow, the people who used it to seek answers are prime candidates to adopt new technologies like ChatGPT. The numbers don’t lie.

An analysis of the effect of ChatGPT on Stack Overflow and Reddit published on nature.com noted, “We estimate that Stack Overflow’s daily web traffic has declined by approximately 1 million individuals per day, equivalent to approximately 12% of the site’s daily web traffic just prior to ChatGPT’s release.” The report also noted a decrease in posting activity on the site. In time, Stack Overflow bowed to the inevitable loss in revenue and laid off 28% of its staff in October 2023.

GitHub Copilot

The success of GitHub Copilot and other AI-based developer tools increased the pressure on sites that offer answers to developers. The advantages of having a tool that can literally write code (and comments) to meet the needs of a developer cannot be understated. I’ve used GitHub Copilot for about a year and although I am not a professional developer and only write PowerShell scripts, GitHub Copilot has removed much of the need to lookup code examples.

GitHub Copilot shows generative AI off at its best. The users know what they are looking for, recognize errors, the source material for the LLMs is based on working code, and the output must always meet the acid test that it either works or not. It’s much easier for a tool like GitHub Copilot to cope with code than it is for its Microsoft 365 Copilot counterpart, which must deal with the vagaries of writing styles and content found in Office documents.

Users can’t be blamed for switching focus. From their perspective, it’s much easier to use a tool like ChatGPT than clicking through multiple threaded posts seeking a definitive answer to a problem. Leaving aside the salient fact that generative AI is quite capable of producing horrendously inaccurate answers, the user experience is easier, especially when AI delivers what seems to be well-crafted and complete answers.

Killing the Goose that Lays the Golden Eggs

All of this is great for those who sell generative AI products. At least, it is for now. The danger exists that the source material ingested by the LLMs used by generative AI will dry up over time as websites close because of a lack of traffic and declining revenue. In the Microsoft 365 space, we’ve seen this happen earlier this year with the demise of the Tekki Gurus site.

If no new content is created and published in blogs and articles on websites, ChatGPT and Microsoft Copilot will increasingly rely on aging information. This might be fine for answering questions of historical interest; it won’t be for questions about technology.

In October, I commented that errors and hallucinations generated by Microsoft 365 Copilot run the risk of corrupting the Microsoft Graph by being included in documents and files that are subsequently included in Microsoft Search and the semantic index. Once in the Graph, the bad information becomes available for Copilot to reuse and spread. No doubt users will pay attention to what Copilot generates in its responses and will attempt not to reuse erroneous content. But humans are humans and sometimes the pressure of work leads to mistakes.

Some websites won’t be affected. I think sites that offer very specific product content are less likely to see dramatic falls in website traffic patterns than those which specialize in covering general-purpose technology, like Microsoft 365. Sites offering news coverage and other time-dependent content will be less affected because of the time required to populate the LLMs with new material. Sites selling products won’t be affected because generative AI just doesn’t do this kind of thing (yet), and so on.

In-Person Technology Conferences Score

In a weird sort of way, in-person technology conferences become more important in the new world. Human interaction with conference attendees, asking questions at sessions, and the ability to have offline conversations with experts to explore their knowledge are real advantages that artificial intelligence cannot deliver. Virtual conferences offer the chance to learn and share knowledge too, but that in-person connection is where magic happens.

With that in mind, I look forward to meeting people at the ESPC event in Stockholm next month. Perhaps someone there can convince me that AI won’t continue to kill websites that publish valuable information about how technology works, but given the evidence available today, I can see only one outcome.