ChatGPT produces outbound clicks — links that actually send a user away to another website — in only 5.2% of conversation sessions, according to a new arXiv preprint [S1]. That number, roughly one in twenty, is a fraction of Google's referral rate and it sits at the centre of a question every publisher, advertiser and small business with a website now needs answered: what happens to the open web when the search engine stops sending you traffic?

The answer, according to the researchers, is not simply "less traffic." It is a different kind of internet — one where the intermediary satisfies the user's question internally and the referral bargain that has funded content production for two decades starts to fray [S1].

How the study works

The paper, which has not been peer-reviewed [S1], uses URL-level Comscore U.S. desktop clickstream data to compare what happens when people seek information through ChatGPT versus Google [S1]. The researchers exploited the staggered rollout of ChatGPT Search — the feature that lets the chatbot browse the live web — to estimate how much traditional Google searching it displaced [S1].

Think of it as a natural experiment. When ChatGPT Search became available to more users, did those users keep Googling at the same rate, or did they substitute? The answer: wider ChatGPT Search access cut traditional search use by 9.4% [S1].

The click that doesn't come back

The 5.2% outbound-click figure is the study's most striking data point, but the texture around it matters as much as the headline number.

ChatGPT's remaining outbound clicks are not a scaled-down version of Google's referral stream [S1]. They skew toward specialised destinations and away from ad-supported sites [S1]. In other words, when ChatGPT does send you somewhere, it is not sending you to the same places Google would have — and it is particularly not sending you to the ad-funded publishers that have historically lived off search referrals.

Search-referral losses were largest for informational categories [S1] — the kind of queries where a user wants to learn something, not buy something. That is precisely the content type that has depended on Google traffic to fund ad-supported answers pages, tutorials, explainers and reference material.

The authors frame this as a structural shift: AI search satisfies information needs inside the intermediary — inside the chatbot itself — while weakening the referral bargain that has linked search, traffic and content production on the open web [S1].

What it means

For two decades, the web's economic logic has been straightforward. A user types a question into Google. Google shows ten blue links. The user clicks one. The destination site serves an ad. The publisher earns a fraction of a cent. Repeat billions of times daily and you have funded an entire ecosystem of free content.

ChatGPT collapses that chain. The user asks a question. The model synthesises an answer from its training data and live web access. The user never leaves. No click. No ad impression. No revenue for the site whose information was used to train or ground the answer.

This is not a hypothetical future. The 9.4% search displacement figure suggests it is already happening at meaningful scale on U.S. desktops [S1]. And the click pattern data shows the effect is uneven — informational publishers are losing the most, while specialised destinations are comparatively protected [S1].

For a regular person, the immediate experience is better: you get a direct answer instead of clicking through three mediocre blog posts to find one good paragraph. The cost is invisible to you but real to the system. The content that trains and grounds these models has to come from somewhere, and if the producers of that content can no longer monetise it through search referrals, the supply side of the open web weakens.

What it means for business

A two-person content agency that writes explainers, how-to guides or reference material for ad-supported sites is the first to feel this. If your traffic model is "rank on Google, earn from display ads," the 9.4% search-use decline is a direct hit to the top of your funnel [S1]. And the skew away from ad-supported destinations means the traffic that remains is less likely to land on your pages at all [S1].

A suburban real-estate agency that publishes suburb guides, market explainers or investment tips to capture search traffic faces a similar squeeze. The informational content that used to pull in prospective vendors and buyers is exactly the category the study identifies as most exposed [S1].

For businesses that sell through specialised platforms — a niche software tool, a professional database, a subscription research service — the news is less grim. ChatGPT's outbound clicks skew toward specialised destinations [S1], which means if your product is the destination rather than the ad-funded middleman, you may retain a share of referral traffic even as the overall pie shrinks.

The practical implication for operators: audit where your traffic actually comes from. If it is Google referrals to ad-supported informational pages, this study suggests structural headwinds. If it is direct visits, email, community or specialised platform traffic, your exposure is lower. No one should panic off a single preprint, but the direction of travel is clear enough to warrant a strategy conversation this quarter.

What we don't know yet

This is a single, non-peer-reviewed preprint [S1]. Every empirical finding is provisional until independent researchers replicate it.

The dataset covers U.S. desktop clickstream only [S1]. Mobile behaviour, international markets and app-based usage could look materially different — and most of the world's internet users are on mobile.

The two headline statistics — 5.2% outbound clicks and 9.4% search displacement — are tied to specific ChatGPT Search access expansion events [S1]. As ChatGPT Search rolls out more broadly and user habits evolve, those numbers could move in either direction.

The authors' conclusion about a "weakening referral bargain" is interpretive — it describes a potential structural shift, not a measured causal outcome [S1]. The data shows correlation and displacement; the broader economic argument is the researchers' synthesis of what those numbers imply.

What would settle the question: a peer-reviewed replication with mobile and international data, plus longitudinal tracking of outbound click rates as AI search features mature. Until then, the 5.2% figure is a signal flare — bright enough to take seriously, not yet bright enough to navigate by.

If this story sharpened your thinking, subscribe to keep reading — the next piece will land in your inbox before the consensus catches up.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.