The impact of AI on the web traffic received by websites worldwide continues to generate a flood of studies (and headaches) in this regard. The latest is a study conducted by Profound, a consulting firm specializing in AI optimization, which examines the way ChatGPT generates referral traffic.
According to an analysis published by Josh Blyskal, its head of AEO (Artificial Engine Optimization), the traffic originating from ChatGPT has decreased by 52% since July 21. Based on his conclusions, the reason does not appear to be any technical error, but rather a change in how OpenAI’s tool selects whom and how to cite in its responses.
This represents a loss of half the traffic: there is no need for me to explain the implications that such a change can have for brands, which now face a new challenge. AI favors those who provide answers.
The decline in referral traffic from ChatGPT coincides with a notable increase in citations to sites such as Reddit (+87%) and Wikipedia (+62%) in AI-generated responses. In the words of the author:
“The decline in referral traffic began precisely when citation patterns changed dramatically. Citations to Reddit increased by 87% starting July 23, reaching over 10% of all ChatGPT citations. Wikipedia simultaneously reached historic highs, with an increase of 62% from its July low, making up nearly 13% of citations yesterday.
The top three domains (Wikipedia, Reddit, and TechRadar, a portal focused on technology and electronics) have collectively grown by 53%, now controlling 22% of all citations. This means that one out of every five ChatGPT citations is directed to only three sites. At the same time, branded websites are seeing fewer opportunities for citation: millions of potential references are being absorbed by these dominant platforms. This is not due to a GPT-5 update, as the consolidation began weeks before the model’s release.”
For Blyskal, this demonstrates that OpenAI is experimenting with its RAG system (Retrieval-Augmented Generation) to prioritize sources that provide responses considered high-quality and useful for users. But what exactly is RAG?
A traditional language model is trained with a vast amount of data, but its knowledge is static; it is “frozen” at the moment the training ends. If you ask it about recent news, it will not be aware of it. Furthermore, sometimes it can “hallucinate” and invent information if it does not have enough data.
In the case of RAG, the approach combines generative models (such as GPT) with a search or document retrieval system. When a user asks a question, the system searches a database or the web for the most relevant documents. Thus, ChatGPT does not generate the response from scratch but instead drafts it using those retrieved documents as support, and it can link to or reference the sources (“according to Wikipedia” or “on Reddit, it is discussed that…”).
And now comes the part where the author’s arguments, apart from the data, may become more debatable:
“Reddit and Wikipedia do not prevail by being special. They win by default because they are the only ones that provide direct answers. When someone asks, ‘What is the best CRM for startups?’ and a brand’s website says ‘Request a demo,’ whereas Reddit has a thread comparing 10 options, Reddit is cited.”
This is accurate, but it is also true that there are thousands of websites worldwide producing quality content that goes beyond ‘request a demo.’ Well-structured content, useful for users, and developed according to the classic parameters of SEO. Thus, what sets Wikipedia or Reddit apart from the rest?
We certainly know that OpenAI has used both platforms in the past to train its models, which indicates a certain preference over other options. Moreover, in 2024, OpenAI and Reddit announced a collaboration that allows OpenAI to access Reddit’s data API. This agreement provides OpenAI with structured, real-time content from Reddit in an effort to improve the capabilities of its language models, such as ChatGPT, to understand and generate responses based on real conversations.
That is already quite indicative. However, what other factors may work in their favor?
It is clear that both have their strengths, each in their own way: Wikipedia offers comprehensive, neutral, and well-structured explanations, while Reddit focuses on discussions in which users compare options, share real experiences, and highlight advantages and disadvantages.
Another factor is coverage and updates. Wikipedia manages to remain relatively up-to-date thanks to its global community, while Reddit offers freshness by including conversations about very recent topics, such as the latest software updates or opinions on a new product. In addition, both platforms use natural language, the same that users employ when asking questions, while sometimes the media and brand websites may employ a more formal language.
Furthermore, structure also plays a key role. Wikipedia organizes information hierarchically and with references, which makes it easier for the model to understand and cite it. As for Reddit, it sorts responses by votes, which can help to identify which comments have greater relevance or consensus.
Finally, there is the perception of impartiality: Wikipedia is recognized for its neutral tone, and although Reddit clearly is not, its (theoretical) plurality of opinions makes it a useful source for presenting different viewpoints.
For the author of the study, the message is clear: if you wish to be cited by ChatGPT (and, by extension, to be visible in the new AI ecosystem), you must become a source of useful information. Rather than focusing on text optimized for conversion and keywords, brands ought to begin creating content that resolves real problems for their customers.
Some examples of this type of content include:
In reality, the analysis by Profound also highlights a rather evident reality of these new times in AI: instability. Just as a core update from Google can cause havoc (or success) for a website, a manual adjustment to the weighting of citations by OpenAI resulted in a 52% drop in traffic in less than a month. The traffic from AI tools is volatile and subject to the experiments conducted by technology companies.
The future of SEO and content marketing no longer depends solely on Google; rather, it depends on the direction taken by companies such as OpenAI. It may be that brands and marketing professionals who adapt to this new reality, prioritizing utility and clear answers, will be the ones who achieve visibility and trust in the era of artificial intelligence.
Or perhaps not; everything will depend on Sam Altman and his team.
Image: Flux Schnell
Your email address will not be published. Required fields are marked *
Δ