We have seen a clear decline in LLM referral traffic over recent months, especially from ChatGPT, and this trend lines up with the release of GPT-5.

Many senior marketers are asking why GPT-5 seems to send fewer clicks to websites than GPT-4.5. The answer is simple. GPT-5 gives people more complete answers inside the chat, so they do not feel the need to visit a site.

LLM referrals in Looker Studio, travel sector website

Below is a clear and simple explanation based on our research and understanding of OpenAI’s documents, testing, and reports by research bodies.

GPT-5 does not guess links as much as GPT-4.5

GPT-4.5 often tried to provide links even when it was not fully certain about the source. It sometimes created URLs or citations that looked real but were not always accurate. This behaviour made users click through because the model appeared to be pointing them toward helpful sources.

GPT-5 stopped doing this. It only shows a link when a real source has been retrieved. This aligns with OpenAI’s documentation, which explains that GPT-5 is stricter about source attribution and avoids fabricated citations. This change means people see fewer links overall.

Fewer links mean fewer natural opportunities for users to click out of the chat.

GPT-5 gives more “complete” answers

GPT-4.5 often provided shorter or more surface-level answers, which encouraged users to click websites to get a full explanation. These shorter replies acted like a teaser and created a need to seek depth elsewhere.

GPT-5 behaves differently. It gives longer, clearer, and more complete explanations inside the chat. Users often describe the answers as “everything I needed in one go.” This matches OpenAI’s positioning of GPT-5 as a model that handles complex reasoning and extended responses.

When users feel the model has already answered the full question, they do not feel compelled to click through to external content.

GPT-5 thinks more inside the model

GPT-5 has stronger internal reasoning. It can break down tasks, analyse problems, and form structured answers without relying on external retrieval. OpenAI documentation highlights that GPT-5 includes improved reasoning modes that allow it to complete more work “inside the model” instead of depending on outside sources.

GPT-4.5 leaned more on external references because it sometimes needed them to support or extend its answers. GPT-5 does not need to do that as often.

Less use of external retrieval directly results in fewer outbound links being shown to users.

People trust GPT-5 more

GPT-5 delivers answers that feel more confident, structured, and complete. Users often perceive its tone as more authoritative. When people trust a response, they are less likely to double check it on another website.

GPT-4.5 sometimes produced uncertain or partial answers, which naturally encouraged more verification. GPT-5 reduces that friction. OpenAI notes that GPT-5 aims for “higher accuracy and more reliable reasoning,” which leads to stronger user confidence.

High trust results in fewer clicks because users feel the conversation has given them everything they need.

GPT-5 is designed to keep the user in the chat

GPT-5 is built to answer a question fully in one reply. Its internal architecture can switch between lighter and deeper reasoning models, depending on what the user asks. This allows GPT-5 to complete a wider range of tasks within the chat itself.

OpenAI documentation describes this as part of GPT-5’s “unified model” design, which aims to reduce unnecessary steps and keep the interaction smooth.

Because GPT-5 can satisfy more user needs directly inside the conversation, fewer people feel the need to leave the platform.

This naturally lowers website traffic from ChatGPT.

In summary

  • GPT-5 shows fewer links.
  • GPT-5 gives deeper answers.
  • Users trust GPT-5 more.

GPT-5 reduces the number of times users click through to websites. This is a natural effect of better model design, not a negative intention.