Similarweb: ChatGPT outbound clicks are highly concentrated

Similarweb has put a number on what many B2B marketers suspected but could not prove: ChatGPT's outbound traffic is not spread across the web. It flows to a small cluster of destinations, and if your domain is not in that cluster, you are receiving effectively nothing.

Search Engine Journal's write-up of the Similarweb data makes a second point that deserves equal attention. AI search is not cannibalising Google's share of referral traffic so much as sitting on top of it. Users who consult ChatGPT are, in aggregate, still going to Google. The two channels coexist, at least for now. That framing matters because it reframes the competitive question: the problem for most brands is not that AI search is stealing their Google traffic. It is that AI search is creating a parallel referral channel to which they have no access at all.

The concentration problem, stated plainly

Referral traffic from large platforms has always been unequal. Google's own results follow a power-law distribution: the top three results take the majority of clicks. ChatGPT appears to follow a steeper version of the same curve. When a model generates a response and includes a citation, the click typically goes to one of a limited set of high-authority, frequently cited domains. Academic publishers, major news outlets, and a handful of specialist reference sites collect the bulk of outbound traffic. Everyone else waits.

For B2B brands in financial services, multilateral institutions, or industrial groups, this is a structural problem, not a content quality problem. A procurement officer at a large infrastructure firm asking ChatGPT about cement standards or carbon pricing mechanisms will receive an answer. That answer will likely cite a regulatory body, a major newspaper, or an international standards organisation. The firm whose white paper actually contains the most authoritative treatment of the subject may not appear at all, because the model's citation behaviour is shaped by prior training, domain authority signals, and the density of pre-existing references to a source, not by the accuracy of any individual document.

The Similarweb figures do not offer a precise breakdown of which domains dominate, but the directional finding is consistent with independent research showing that a small number of sites account for a disproportionate share of AI-generated citations. That pattern is self-reinforcing. Sites that get cited get visited; sites that get visited accumulate the behavioural signals that further entrench their authority.

What the "layering" finding changes

The more consequential insight from the Similarweb analysis is that AI search and Google are additive, not substitutive, for now. Users are not abandoning Google queries; they are running ChatGPT sessions in addition to them. This has two implications that pull in opposite directions.

First, it means Google SEO remains worth protecting. Brands that conclude AI search has made traditional search investment redundant are making a premature decision on thin evidence. Second, the additive structure means a new layer of referral competition has appeared above the Google layer, one that most brands have no existing strategy to address. Winning in Google does not transfer automatically to winning citations in ChatGPT. The signals that matter to a language model during inference are not identical to the signals Google's ranking algorithm weights. Domain authority overlaps, but structured data practices, the presence of a source in training corpora, and the frequency with which other cited sources refer back to your content all play roles that Google optimisation does not fully address.

For a multilateral institution or a UN agency, the implication is sharp. These organisations are often the authoritative original source on topics like disaster risk, microfinance penetration, or labour market data. They publish the primary research. But if their content is not among the domains ChatGPT habitually cites, the answer a policymaker or journalist receives will attribute the finding to a secondary source that happened to cover the original study. The institution loses the citation and, with it, the visibility signal that shapes future model behaviour.

The brands best positioned to capture AI referral traffic are those already cited frequently in third-party editorial content, those with high domain authority, and those whose content is structured to match the informational intent of the queries models are being asked to answer. That is a description of a media relations strategy as much as an SEO strategy. Earned coverage in outlets that the models treat as authoritative proxies is, at this point, more valuable for LLM visibility than almost any on-site optimisation.

The Similarweb data does not solve the measurement problem. Referral traffic is a lagging and incomplete signal; most AI-generated answers produce no click at all, which means the citation has already done its work, shaping perception without generating a visit. Brands waiting for traffic analytics to confirm their AI visibility strategy is working will be waiting for a signal that may never clearly arrive.