ChatGPT is a referral channel now
Over the past 30 days, our self-hosted analytics recorded more visitors arriving at our equity research platform EvidInvest from AI assistants — ChatGPT, Microsoft Copilot, Perplexity — than from Google’s organic search results, where Search Console recorded zero clicks in the same window. The absolute numbers are small; the ordering is the story. Here is how that traffic actually shows up in analytics, and what it took for assistants to read the site at all.
AI assistants cite sources and users click them — but the visits are easy to mismeasure, and easy to lose entirely. ChatGPT tags its outbound links with utm_source=chatgpt.com (Copilot does the same with utm_source=copilot.com), Copilot answers lean on Bing’s index rather than Google’s, and an anti-bot firewall that shows assistants a challenge page makes your content invisible to the fastest-growing way people ask questions.
What the data shows
Three assistant domains appear as referrers on the platform in the last 30 days: chatgpt.com (10 visitors), copilot.microsoft.com (4) and perplexity.ai (3). For a site whose Google Search clicks currently round to zero, that makes assistants the largest measurable search-shaped channel — ahead of the search results pages themselves. And the referrer header undercounts: assistant apps often strip it, so some of these visits only reveal themselves through query parameters. ChatGPT appends ?utm_source=chatgpt.com to links it cites, and those tagged URLs showed up in our analytics more often last week than the chatgpt.com referrer itself. If you only segment by referrer, you are undercounting this channel.
Why Bing suddenly matters
The same 30 days show Bing referring roughly twice as many visitors as Google to the platform — unusual, until you remember that Microsoft Copilot answers from Bing’s index. A page that Google has not crawled can still be cited by Copilot the day after Bing picks it up, and Bing accepts push notification of new URLs via IndexNow instead of waiting for a crawl. For a small site, that is the cheapest indexation win available: we ping IndexNow on publish and Bing routinely lists new pages while Google is still deciding whether to visit.
The firewall problem nobody checks
The least obvious lesson: assistants have to fetch a page to cite it, and they fetch like bots — because they are. The platform runs a bot-protection challenge (a real necessity; scrapers dwarf humans in its raw traffic, as they do in its search data), and that challenge was serving an interstitial to every assistant fetch. The fix was not to disable protection but to give machines an honest, cheap surface: the RSS feed and plain-markdown mirrors of each article are exempt from the challenge, and llms.txt — generated from the sitemap so it can never drift — tells assistants what exists and where. Machine readers get the content; scrapers hammering the app get the challenge.
What to take from it
Treat AI assistants as a channel you instrument, not a rumor: segment by the utm_source tags assistants add, not just referrers; make sure Bing actually has your pages, because Copilot speaks from its index; and fetch your own site the way a bot would, because that is how the assistants see it. None of this replaces classic SEO — it compounds it. The entity work, structured data and honest sitemaps that feed Google’s crawler are the same signals assistants use to decide whom to cite.
Want your product visible where your customers ask questions?
EBD Sweden builds the measurement and AI-readiness plumbing — analytics you own, agent-readable content surfaces, and the AI agents that act on the data.