Read Often and Visited Rarely

Published material has always been valued by the traffic it attracts, on the assumption that consumption and visits are roughly proportional. Network-level measurement now shows the two diverging by orders of magnitude. AI systems request pages at a rate that bears no relation to the number of people they send onward, and the referrals they do send are small in absolute terms, growing quickly from a small base, and inconsistent in quality depending on the month measured. This report sets out both sides of that gap and resists the single verdict the material does not support.

The scale of consumption

Cloudflare, which sits in front of a large share of the web and publishes its own network measurements, has quantified the gap directly as a crawl-to-refer ratio: HTML requests from a platform’s user agents divided by HTML requests carrying that platform’s hostname in the referer header. For the week of 19 to 26 June 2025 it reported that the ratios range from Anthropic’s 70,900:1 down to Mistral’s 0.1:1.

Cloudflare’s own caveat travels with the figure and materially affects the top of the range: traffic referred by Claude’s native application does not include a referer header, so the denominator is understated and the gap correspondingly overstated. The measurement is a reasonable proxy for the direction and a poor one for the magnitude at the extremes.

A second measurement in the first week of August 2025, across the same customer base, gave Anthropic at nearly 50,000:1, OpenAI at 887:1 and Perplexity at 118:1, with sharp variation by sector: in news and publications the ratios were 2,500:1, 152:1 and 32.7:1, and in computer and electronics 8,800:1, 401.7:1 and 88:1. The ordering is stable across both periods even where the absolute numbers move.

Most of that consumption is not connected to answering anyone’s question. Classifying its crawler traffic by declared purpose, Cloudflare found training traffic responsible for nearly 80% of the crawling from AI bots, with user-action and undeclared purposes together under 5%. The bulk of the reading is not a retrieval that could have produced a visit in the first place.

The scale of referral

Referral from these systems is growing at rates that make percentage change almost meaningless without a base. Adobe Analytics, examining more than a trillion visits to US retail sites, reported in March 2025 that traffic from generative AI sources increased by 1,200 percent compared to July 2024, with travel and hospitality up 1,700% and banking and financial services up 1,200% over the same window, holiday-period traffic up 1,300% year on year, and Cyber Monday 2024 up 1,950%.

The base is the point. Similarweb, whose figures it states are estimations and extrapolations from third-party source materials, reported that AI platforms generated over 1.1 billion referral visits in June 2025, up 357% year on year. TechCrunch, reporting the same Similarweb dataset in July 2025, set that 1.13 billion against 191 billion Google Search referrals to the top 1,000 websites in the same month — roughly 0.6% of the search figure. The comparison is carried here at second hand because the Similarweb page holding it no longer resolves; the 1.13 billion and 357% figures are confirmed at source.

Whether the traffic is any good

The quality evidence points in different directions depending on when it was taken, and the dates are not optional. Adobe, measuring US retail in February 2025, found AI-referred visitors showing 8% higher engagement, 12% more pages per visit and a 23 percent lower bounce rate, while being 9% less likely to convert — a gap that had narrowed from 43% in July 2024. On that measurement the visitor reads more and buys less, and the shortfall is closing quickly.

Semrush reached the opposite conclusion in a different category. Working from more than 500 high-value digital marketing and SEO topics translated into search terms and prompts, and publishing no sample size or confidence intervals, it reported in July 2025 that the average AI search visitor is 4.4 times as valuable as the average visit from traditional organic search, based on conversion rate. Semrush sells software to the audience the study measures, and the topic set is narrow enough that the finding should not be generalised beyond it.

These two results are not straightforwardly reconcilable, and the studies differ in vertical, period, method and disclosure. The defensible statement is that AI-referred traffic behaves differently from search traffic and that whether the difference is favourable depends on the category and the month.

What this means for an organisation

The gap between consumption and referral breaks the assumption that traffic measures the value of published material. On Cloudflare’s August 2025 figures, material can be read thousands of times for every visitor it returns, and roughly 80% of that reading is for training rather than for an answer that might have cited it. A page’s contribution to what these systems can say about an organisation is therefore invisible in any analytics view the organisation holds.

It does not follow that referral has stopped mattering. On the June 2025 comparison the AI figure is under one per cent of the search figure, so conventional routes still deliver almost all arrivals, and the growth rates quoted above describe a small quantity multiplying rather than a large one shifting. Both statements are true at once: the reading is enormous and the visiting is marginal.

The practical difficulty this creates is one of accounting rather than of strategy. Two different returns now come from the same published page, and only one of them appears in any system an organisation operates. Deciding what to publish, correct or retire on the evidence of the visible return alone means deciding on a partial measure, and the partiality is not small — it is the difference between 70,900 and one in the widest ratio Cloudflare measured.

Where a return cannot be counted, the substitute is to inspect the input. Whether the material exists, whether it is current, whether it agrees with the organisation’s other material, and whether anything outside the organisation’s own estate confirms it are all checkable without cooperation from the systems doing the reading, and they are the properties determining what those systems can accurately say. No traffic figure will register that the checks were done.