AI Referral Traffic Study

Edition 1 reported AI referral counts for 12 selected sites. Edition 2 adds total-session denominators across a 108-property portfolio frame, of which 104 recorded sessions in the window. The findings describe this owned portfolio, not a representative sample of websites.

Across 104 portfolio sites with measurable sessions, sources matching the five-assistant matching rule described in the Method section accounted for 1.84% of all sessions, or 7.99 AI-attributed sessions for every 100 organic-search sessions, in the 90 days to 21 September 2026. ChatGPT-matched sources accounted for 86.8% of that traffic.

AI-matched sources accounted for 1.84% of recorded sessions

6,165 sessions matched an AI-assistant source out of 334,744 total sessions across the 104-site panel. The median site's own AI-attributed share was 0.65% of its sessions, the middle half of sites ran between 0% and 3.67%, and 36 of the 104 sites (34.6%) recorded zero matched AI sessions in this window (a further 4 properties had zero sessions of any kind and are excluded from these figures entirely). Edition 1 reported counts for a selected 12-site subset only; it did not publish this shape.

One property supplied 97,123 of the panel's 334,744 sessions, 29.0%, in the Careers, education and travel sector. Excluding it raises the pooled AI share from 1.84% to 2.28%. Across the full panel, 246,655 sessions, 73.68%, were classified as Direct; this extraction does not establish why one property's Direct total is so large. Both the 1.84% and 2.28% figures are published.

Finding 1: ChatGPT still sends most of the visible AI traffic, and it's concentrated in very few sites

The single highest site accounts for 15.8% of all matched AI sessions in the panel on its own. The top 3 sites account for 40.7%, and the top 10 for 77.1%. A small number of sites carry most of the matched traffic, and this analysis does not isolate which content features caused that concentration.

Two external benchmarks, each opened and read directly, give different figures for ChatGPT's share. BrightEdge reports ChatGPT generated 95.1% of AI referral traffic in August 2026, up from 89.5% in April 2026. That figure comes from its own AI Market Pulse tracking; sample and method are not disclosed. Similarweb reports a separate measure: ChatGPT received approximately 53% of worldwide visits to generative-AI platforms themselves in May 2026, down from around 76% a year earlier. That is a panel-estimate of traffic to the AI tools, not referral clicks from them, and it identifies Claude as the fastest-growing platform proportionally over the same period. Neither source's population or denominator matches this panel of 104 owned sites. Both measurements agree only on direction: ChatGPT dominant, and Gemini or Claude the platform gaining fastest depending on the source. BrightEdge, 24 September 2026; Similarweb, 29 July 2026.

Finding 2: a third of sites recorded no matched AI sessions

Zero matched AI sessions occurred on 36 of the 104 active sites (34.6%) in 90 days. This matches the concentration pattern above: matched AI traffic here collects around a handful of sites, the same pattern the first edition found, now visible against a full denominator instead of a hand-picked shortlist.

Finding 3: descriptive sector results

Eight sectors contain at least the predeclared minimum of 3 active properties. A ninth (Local services and events) fell to 2 properties after excluding client and unconfirmed-ownership sites and is omitted from this table rather than published below the minimum. These are small, unequal groups: they provide descriptive results, not reliable industry rankings. Share is AI-matched sessions as a percentage of that sector's own total sessions.

Affiliate product reviews recorded the most AI-matched sessions of any sector, 1,634, followed by Food, health and wellness at 1,062; Affiliate product reviews is only the fourth-largest sector by total sessions, so volume and share do not track each other here. Online tools and templates had the highest share by proportion among the eight published sectors, 8.26%.

Finding 4: the original 12-site panel, re-measured over a mostly-overlapping window

The first edition's panel was never a random sample. It was the 12 sites already known to receive the most AI traffic, chosen to show what strong performance looks like. Those 12 properties are included in the full 104-site panel above and also reported separately here for comparison, not a disjoint sample.

The refreshed edition-1 window (28 May to 26 August 2026) spans 91 inclusive days. Edition 2's window (24 June to 21 September 2026) spans 90. The two share 64 days, leaving 27 days exclusive to edition 1 and 26 exclusive to edition 2. These overlapping totals alone cannot establish a sustained trend. The cause of any difference between them was not measured. Historic labels were mapped to current GA4 property IDs using internal references, site subject and an old-window re-pull of each property. Re-pulling the original window independently gives 4,745 AI-matched sessions across the 12 properties, against 4,717 originally published, a close but not exact match. The original edition did not publish property IDs, so the mappings retain some uncertainty.

Edition-1 panel total this window: 3,872 AI-matched sessions on 171,630 total sessions (2.26% AI share, 8.41 per 100 organic sessions), against 4,717 published in the original edition. Given the 64-day window overlap, that difference alone does not establish a falling trend. The historic subset's median property-level share is 3.97%, compared with 0.65% for the full panel.

What this means if you own a site

1. Matched AI sources accounted for 1.84% of this portfolio's recorded sessions. Measure your own baseline before setting a target; a hand-picked panel of your best-performing pages will show a higher number than your whole site.
2. A third of sites with any traffic recorded no matched AI sessions. A zero on your own site is a common outcome in this panel (34.6% of sites), not a rare one.
3. The top 10 sites here carried 77.1% of all matched AI sessions. This analysis does not isolate which content features caused that concentration; publishing checkable factual data is the pattern the first edition also found among the highest-traffic sites.
4. Compare the same properties using a consistent rule, and disclose window length and overlap. A historic panel measured again over a mostly-overlapping window will read as change.
5. This method excludes unclicked citations and cannot isolate Google AI Overviews or AI Mode traffic from Google organic search. The size of that omitted activity is unknown.

The AI Visibility Audit provides a site-specific baseline and action plan for your own site.

Method

• Source: Google Analytics 4, GA4 Data API and Admin API, one property per site, pulled 24 September 2026 (UTC; 25 September in the UK).
• Window: 24 June to 21 September 2026 (90 days). Dates use each property's own configured GA4 reporting time zone (Analytics Admin default), not one common zone.
• Panel: every GA4 property in the accounts I hold as sole or majority owner. That covers the main portfolio account, my own sunnypatel.co.uk property, seo.associates and She Cooks She Eats. Two accounts were excluded, without naming the businesses involved. One is a sales-prospect's property, not yet a client and not mine to publish. The other is a client-adjacent property whose ownership I cannot confirm as mine to publish. One duplicate GA4 property for the same site (a second, URL-named copy of generator.express) was dropped. Sessions for every property were pulled once, on 24 September. Before publication, a separate ownership review on 25 September removed a further 7 individual properties: 2 are client sites, 5 have ownership I cannot confirm. None are named here, for the same reason as the two excluded accounts. That correction was ownership-driven only. No property was added or dropped based on its session or AI numbers. Every figure below was recomputed against the corrected 108-property frame. 4 of those 108 recorded zero sessions in the window. They remain in the downloads but are excluded from the share and percentile statistics below, leaving 104 sites in the panel.
• AI classification rule, applied by matching sessionSource (case-insensitive substring) with one script against one immutable raw extraction per property, independent of what GA4 labels the channel group as:
  • ChatGPT: source contains "chatgpt" or "openai"
  • Microsoft Copilot: source contains "copilot" (the observed source string in this data is copilot.com, not copilot.microsoft.com)
  • Perplexity: source contains "perplexity"
  • Claude: source contains "claude"
  • Gemini: source contains "gemini"
• What this cannot see: generic bing sources are excluded from the Copilot count and counted as Bing organic search instead. GA4 cannot distinguish a Copilot click from a plain Bing search click once the source collapses to bing. This extraction cannot separate Google AI Overviews or AI Mode from other Google organic sessions at all. That traffic is invisible here, not zero. Of the 6,165 sessions matched to the five assistants, 5,770 (93.6%) already carried GA4's own "AI Assistant" channel label. The remaining 395 sessions, across 46 property rows, matched the source rule but GA4 had filed them under another channel, mostly "Unassigned".
• Other AI sources footnote: six additional source domains contributed 24 sessions outside the five-platform scope; these are excluded.
• Organic sessions are rows where GA4's sessionDefaultChannelGroup is exactly "Organic Search". Non-search channels are excluded even when a source name looks search-like.
• Sites are labelled by sector, not domain: downloads use sector labels to protect commercially sensitive site identities.
• Selection bias. This convenience sample describes one owner's portfolio. Shared infrastructure and content practices limit how far it generalises to other websites; treat every number below as a description of this portfolio, not a benchmark for "websites in general."

A Python script applies the matching rule to the stored raw extracts and writes the aggregates below. Downloads contains property-level results and summary statistics.

Limitations

• Referrer loss. Browser privacy settings, in-app browsers and some assistant apps can strip or rewrite referrer data before it reaches GA4. Missing attribution can cause an assistant-origin session to be recorded under another source label, such as "(direct)" or "(not set)"; this study cannot identify every session affected.
• Dark traffic. Any AI answer that resolves a query without a click, or any citation inside an AI response nobody follows, is invisible to a referral method by definition. This study counts GA4 sessions attributed to matched source labels, not citations or mentions, and does not authenticate that every matched session is a genuine human click-through.
• GA4 channel rules. Google publishes default channel-grouping rules, which can change; 395 sessions counted here, across 46 property rows, were not labelled "AI Assistant" by GA4 itself, only matched by this study's own source rule, and a further 24 sessions from six source domains outside the five predeclared platforms were excluded entirely.
• Selection bias. This is 108 sites I own and operate, sharing a builder and hosting pattern. It describes this portfolio, not the whole web or an industry-representative rate.
• One property's session total looks anomalous (the Careers, education and travel property's 97,123 sessions, noted above); it is reported as pulled, not corrected, with the sensitivity check (1.84% including it, 2.28% excluding it) both published.
• Edition-1 identity mapping. The twelve historic sector labels were matched to current properties using internal references, site subject and an old-window re-pull, not a stored property ID, as detailed in Finding 4.

Edition history

• Edition 2, 25 September 2026 (this page): full 108-property frame (104 with sessions), 90 days to 21 September 2026, plus the original 12-site panel re-measured over a mostly-overlapping window.
• Edition 1, 2 August 2026, refreshed 26 August 2026: 12-site hand-picked high-AI-traffic panel, 91 days to 26 August 2026. Kept above, unchanged, as a historic comparison.

How to cite

Sunny Patel, "AI Referral Traffic Study," sunnypatel.co.uk, edition 2, 25 September 2026, GA4 data, 108-property portfolio frame (104 sites with sessions), 90 days to 21 September 2026. https://sunnypatel.co.uk/blog/ai-referral-traffic-study/

Reuse these findings with their denominators

These figures all cover 24 June to 21 September 2026, using each GA4 property's reporting time zone. The edition-2 JSON download contains the anonymous property aggregates used below. Matched source labels describe recorded sessions; they are not counts of citations, people or verified human visits.

Reusable summary: In my owned portfolio, GA4 source labels matching five declared assistants accounted for 6,165 of 334,744 recorded sessions (1.84%) across 104 properties with traffic from 24 June to 21 September 2026. This convenience sample does not represent the web, and the extraction cannot identify every lost referrer or separate Google AI Overviews and AI Mode from ordinary Google organic sessions.

Downloads

• Edition 2 data: CSV: one row per property (sector label and an anonymous site number, no site names), all 108 properties in the frame
• Edition 2 data: JSON: the same per-property rows plus sector splits, panel totals and the sensitivity check
• Edition 1 data: CSV: the original 12-site panel, unchanged
• Edition 2 verification script: Python: recomputes the table above from the public anonymous JSON, with no network requests or additional packages

To check the arithmetic, save the edition-2 JSON and Python script in the same folder, then run:

python verify-ai-referral-study-v2.py ai-referral-traffic-study-2026-09-21-v2.json

The script prints a JSON summary, including the input file's SHA-256 hash, the 108-property frame, the 104-property analysis panel and the sensitivity denominators. It uses unrounded property counts before rounding reported percentages. It does not pull fresh GA4 data, rerun source classification from raw session rows, authenticate human visits or verify the original extraction. The new reuse table and script were added on 2 October 2026; the data and edition date remain unchanged.

Related Articles

• AI Visibility Results - the current snapshot of this same data, refreshed on its own page
• AI Search Statistics: Fact-Checked - the popular numbers traced to their primary sources
• What Is LLM Optimisation? - the practice behind these referral numbers
• Optimise Content for AI Search - content patterns that earn citations
• Why Your Brand Is Not Appearing in ChatGPT - diagnosis for invisible brands