Quick answer: Between 4 May and 1 September 2026, AI assistants sent 1,225 sessions to 1clickreport.com: ChatGPT 515, Claude 406, Gemini 243, Perplexity 33, NotebookLM 22, Copilot 6. Over the last 90 days that is 7% of all sessions, our third-largest source after direct and Google. The pages they cite are how-to guides with specific settings, not comparison or product pages. ChatGPT referrals fell 80% from May to August while Perplexity grew, and visitors from claude.ai signed up at 25% against 2.7% from Google. All of it is our own data, and the methodology is below so you can run it on yours.
Most of what is written about AI search traffic is either a vendor pitch or a survey. This is neither. It is one site, four months, every session an AI assistant sent us, pulled from GA4 and Mixpanel on 2 September 2026, with the ugly parts left in. We build a marketing analyst that runs inside Claude and ChatGPT, so we care about this more than most, but the numbers are the numbers.
What 120 days of AI referrals look like
| Engine | Sessions, 4 May to 1 Sep | Share of AI sessions |
|---|---|---|
| ChatGPT | 515 | 42% |
| Claude | 406 | 33% |
| Gemini | 243 | 20% |
| Perplexity | 33 | 3% |
| NotebookLM | 22 | 2% |
| Copilot | 6 | 0.5% |
| Total | 1,225 |
For scale: over the most recent 90 days the site had 12,716 sessions. Direct was 7,409, Google organic 3,011, and the six AI assistants together 894, which is 7%. Nothing else came close: LinkedIn sent 111, Bing 60, Product Hunt 36. Over the full four months the AI share was 8.5%. Most published benchmarks we have seen put AI referrals well under 1% of a B2B site's traffic, so this is an unusual site, and the rest of this post is about why.
The trend by engine: ChatGPT fell 80%, Perplexity grew
| Sessions from | May | June | July | August |
|---|---|---|---|---|
| ChatGPT | 193 | 166 | 130 | 39 |
| Claude | 101 | 115 | 118 | 79 |
| Gemini | 22 | 92 | 84 | 44 |
| Perplexity | 4 | 9 | 4 | 16 |
Three things are in that table. ChatGPT went from 193 sessions in May to 39 in August. Claude held its level until August. Perplexity, tiny in absolute terms, is the only engine that grew every month it was measured.
The ChatGPT drop has a mechanical explanation, and it is not that ChatGPT stopped liking us. ChatGPT's web search retrieves through Bing. In late August we discovered that our IndexNow feed, the mechanism that tells Bing a page changed, had sent ten URLs on 27 August and nothing before or since, because our deploy hook only announces pages whose sitemap date is within two days and nobody had touched a sitemap date in weeks. The five pages ChatGPT cited most were published in June and had not been visibly updated. Bing had no reason to re-crawl them, and an engine that weights freshness stopped citing them. Both were fixed on 2 September: every content change now bumps its sitemap date, which triggers the ping on deploy. We will report what happens to the September number.
Claude's line needs one caveat. Some of its referrals are not citations at all. 1ClickReport is an MCP server that people connect to Claude, and the OAuth hop back from claude.ai to our app registers as a claude.ai referral. GA4 cannot tell a citation click from a connector redirect. If you sell anything that plugs into an assistant, split those before you celebrate.
Which pages the engines actually send people to
| Landing page | AI sessions, 120 days | What it is |
|---|---|---|
| (not set) | 281 | Sessions with no landing page recorded, mostly cross-domain hops into the app |
| Meta Advantage+ campaign setup | 209 | Step-by-step guide with settings |
| Meta ads creative safe zones | 87 | Pixel specs and templates |
| Homepage | 79 | Product |
| Meta Andromeda update guide | 79 | What changed and what to do |
| Meta ad policy changes | 57 | Compliance guide |
| Google May 2026 core update | 33 | Analysis with our own data |
| Facebook ad algorithm changes | 25 | Guide |
| Competitor analysis tools | 23 | Listicle |
| Advantage+ detailed targeting | 17 | Guide |
The pattern is not subtle. Nine of the top ten are how-to guides about Meta Ads, each with concrete settings, thresholds and dates. Our comparison pages, which rank on Google and carry the buying intent, are almost absent: the AgencyAnalytics comparison got 8 sessions from all sources in the last month, none of them from an assistant. The engines cite pages that answer a question completely. They do not cite pages that argue for a purchase.
The 281 sessions with no landing page are an honest gap. GA4 records "(not set)" when the session starts on a page without the tag or crosses from our marketing domain to the app domain mid-hop. Nearly a quarter of our AI traffic is unattributable at the page level. If you run a marketing site and an app on different subdomains, expect the same hole and set up cross-domain measurement before you trust any per-page number.
Do AI visitors convert? Yes, on a sample too small to brag about
| Source, last 30 days | Visitors | Signups | Rate |
|---|---|---|---|
| Direct, no referrer | 591 | 7 | 1.2% |
| 255 | 7 | 2.7% | |
| claude.ai | 12 | 3 | 25% |
| 5 | 1 | 20% |
Twelve visitors and three signups is not a study. But it is the same direction we have seen every month since we started measuring, and it makes sense: someone who asked an assistant a specific question and clicked the source it cited has already read the answer and decided it was credible. The Google visitor is still deciding. The practical consequence is that a small AI referral number can be worth more than a large organic one, which changes what is worth publishing.
One more number for anyone selling an AI-native product. Google Search Console shows 5,670 impressions over 90 days for queries containing "Claude", "MCP" or "AI analyst", and 7 clicks. Google has almost no demand for the category yet. The assistants are where the category lives, which is why an 80% swing in ChatGPT referrals matters to us more than a ranking change would.
How to run this on your own site
- GA4, sessions by source. Report on the session source dimension and keep chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, copilot.com and notebooklm.google.com. These land in the referral channel, so the channel grouping hides them. Our GA4 setup guide has the exact filter and a custom channel definition.
- GA4, landing page by source. Same filter, break down by landing page. Expect a large "(not set)" row if you have more than one domain.
- Product analytics, signups by referrer. We use Mixpanel's initial referring domain on the signup event. Whatever you use, capture the referrer and the UTM at first touch, on the signup event itself. We did not capture UTMs on signup for months and lost the page-level attribution as a result.
- Search Console, category queries. Filter queries by the words that name your category and note the impressions against clicks. It tells you whether your buyers search or ask.
- Bing Webmaster Tools, IndexNow. If ChatGPT matters to you, this page is the one to watch. Zero submissions in the last day means ChatGPT is reading a stale copy of your site.
What we changed after reading our own data
- Every content change now updates the page's sitemap date, which triggers an IndexNow ping to Bing on deploy. The ChatGPT drop was, at least partly, self-inflicted.
- llms.txt, the file written for the engines, still described pricing we had retired. It now matches the site. Consistency is what the engines weight as trust, and we were failing it in the one place built for them.
- The homepage gained Organization and SoftwareApplication schema with our other profiles linked, so an assistant can tie the name on a review site to the site itself.
- On 24 posts the first thing an engine could extract was a sales box. It now sits below the first section on every post, so the extractable passage is the answer.
- The five most-cited Meta guides get a visible update date and a short, sourced key-facts block at the top this month. That is the freshness the engines are asking for, and it is the experiment we will report next.
Frequently asked questions
How do you track AI referral traffic in GA4?
Filter sessions by source for the assistant domains: chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, copilot.com and notebooklm.google.com. GA4 groups these under referral, so the source field is what to read, not the channel. Our GA4 setup guide walks through the exact filter.
Which AI engine sends the most traffic?
Over 120 days ChatGPT sent 515 sessions, Claude 406, Gemini 243, Perplexity 33, NotebookLM 22 and Copilot 6. By August, Claude had overtaken ChatGPT because ChatGPT referrals fell by 80% between May and August while Claude held.
Why did our ChatGPT referrals drop?
ChatGPT retrieves through Bing. Our IndexNow feed had gone quiet, the pages it cited had aged without a visible update, and the sitemap dates never changed, so Bing had no reason to re-crawl. Freshness and a working Bing feed are the two levers, and both were fixed on 2 September.
Do AI referrals convert better than Google?
In our last 30 days of Mixpanel data, 3 of 12 visitors from claude.ai signed up, against 7 of 255 from Google. That is 25% against 2.7%. The sample is tiny, but every month has pointed the same way, and it matches what other SaaS teams report.
What kind of page do AI engines cite?
How-to guides with concrete settings, thresholds and dates. Our five most-cited pages are all Meta Ads guides. Comparison pages and product pages barely appear, even though they rank on Google. The engines cite answers, not pitches.
Sources and method
- GA4 property for 1clickreport.com, sessions by session source and by landing page, 4 May to 1 September 2026, pulled 2 September 2026.
- Mixpanel, unique users by initial referring domain and the sign_up_completed event, 30 days to 2 September 2026.
- Google Search Console, query filter for the category terms, 90 days to 1 September 2026.
- Bing Webmaster Tools, IndexNow and URL submission reports, 2 September 2026.
- Related reading: why Search Console impressions can no longer be trusted on their own and what llms.txt did for our citations.