Kingy AI Referral Index

The one chart

Every Tier A finding, on three measures, grouped by how it was measured. A study that published a range is drawn as a range — none of these values has been restated, rounded or converted to a midpoint.

Reconciled ranges

Where the Tier A studies actually agree, once the Seven Deltas are accounted for. Tier B records and anything pending primary verification are excluded from these figures.

The Seven Deltas

Published estimates of AI referral traffic disagree by more than an order of magnitude. They are not all wrong. Most of the spread resolves into seven recurring differences in what was measured.

1. 1 Dark traffic

   An assistant that sends a visitor without a referrer header is invisible to almost every method here. Cloudflare says so plainly about its own figures: traffic referred by Claude's native app carries no referrer, and it believes the same holds for other native apps, so its ratios "may overstate the respective ratios, but it is unclear by how much." Chartbeat makes the same point from the publisher side, noting that dark social is harder to attribute. The practical consequence is that nearly every AI-referral share you will read, including ours, is a floor rather than an estimate — and nobody can currently say how far off the floor sits.

   #dark-traffic-delta

2. 2 The denominator

   Share of what? Conductor's 1.08% is a share of all website traffic. Chartbeat's "less than 1%" is a share of total pageviews. A share of referred traffic — visits that arrived from any external link — is a much smaller denominator and produces a much larger number from identical data. Our own books show the size of the effect: the same week is 1.3% of pageviews and 2.8% of referred traffic. Neither is wrong; they answer different questions. Most published figures do not say which question they answered, which is why this index records a denominator note on every record.

   #denominator-delta

3. 3 The measurement window

   The studies here cover windows of one week, five months and thirteen months. Cloudflare's crawl-to-refer ratio is a single week in June 2025; Conductor aggregates May to September 2025; Chartbeat runs December 2024 to December 2025. In a channel growing at the rates reported, window length alone can separate two honest figures by a wide margin, and a short window can catch a spike. Where a source does not state its window at all, we cannot place the finding in time and it is quarantined to Tier B — that is why the Daily Mail figure below sits there.

   #window-delta

4. 4 Referrer attribution

   What counts as an AI referral is not a settled question, and the boundary is often drawn by tooling rather than by choice. Conductor excludes Google's AI Mode traffic entirely, because Google Analytics classifies it with organic and AI Overview data without differentiating it. The Daily Mail reports that it cannot separate clicks originating in an AI Overview from ordinary search clicks — "it's all bundled in together." Two studies can therefore measure the same visitor and disagree about whether AI sent them.

   #attribution-delta

5. 5 Panel versus site-side

   Each method sees a different slice of the web and cannot see the others. Pew observes 900 opt-in panellists through an app that records their browsing — excellent for behaviour, tiny and US-only. Ahrefs aggregates Google Search Console across 300,000 keywords — excellent for search, blind to everything that is not a Google result. Chartbeat sees its own client network of news sites; Cloudflare sees requests crossing its network. When these disagree it is frequently because they are not measuring the same population, not because one is wrong.

   #measurement-source-delta

6. 6 Category mix

   The single most under-reported source of disagreement, and the easiest to demonstrate. Inside one study, one method and one dataset, Conductor reports Information Technology at 2.80% and Communication Services at 0.25% — an eleven-fold spread with no methodological difference at all. A cross-industry average therefore tells you very little about your own industry, and two studies sampling different industry mixes will disagree even when both are measured perfectly. This is why the reconciled ranges on this page are published per category rather than as one headline number.

   #category-mix-delta

7. 7 Self-report, forecast, and the small-base problem

   Two distinct failures share this heading. The first is method we cannot check from the source in front of us. Similarweb's 357% year-on-year growth appears in a release that gives its window and its sample, but not its measurement method — so we catalogue it and set it aside. Not because it is wrong, and not as a comment on methodology the company documents elsewhere, but because that release does not let us test the figure. The second is arithmetic on a small base. Chartbeat reports ChatGPT referrals growing more than 200% year over year and AI accounting for less than 1% of pageviews. Both are true. Quoted alone, the first implies a transformation the second rules out: tripling a sub-1% channel leaves it a sub-3% channel. Growth figures without their base are the most common way these numbers mislead, and they mislead in the direction of whoever is quoting them.

   #self-report-delta

Study catalog

Every primary study this index reconciles, grouped by how it measured. This table is generated from the same records as the chart and the downloadable dataset.

Changelog & next wave

This is Wave One. Waves are quarterly. Each wave is archived at publication and never altered afterwards; method changes are appended to the methodology changelog, never applied silently.

Next wave: October 2026.