ChatGPT vs Perplexity vs Gemini: Which AI Sends Referral Traffic

If you run an ad-funded site and want to know which AI is actually sending you visitors, the short answer is ChatGPT, by a wide margin. As of March 2026, ChatGPT accounts for 78.16% of AI chatbot referral traffic to websites, with Google Gemini at 8.65% and Perplexity at 7.07% (StatCounter data, reported by MediaPost). Gemini is the fast-growing number two, Perplexity is falling from its earlier peak, and Claude and Copilot trail in raw volume. But the ranking below ChatGPT shifts depending on whether you count consumer or business traffic, and your own GA4 will not show you the real picture without some setup. This guide walks through who sends what, why, and how to see it for your site.

Which AI sends the most referral traffic to publishers?

ChatGPT is the clear leader, and it is not close. StatCounter's March 2026 data puts the AI chatbot referral market at ChatGPT 78.16%, Gemini 8.65%, Perplexity 7.07%, Microsoft Copilot 3.19%, Claude 2.91% and DeepSeek 0.02%. Similarweb's panel data tells the same story from a different angle: in June 2025, ChatGPT accounted for more than 80% of AI referrals to the top 1,000 domains.

Here is the current picture in one place.

Two practical points follow from this. First, if you only have time to optimise for one AI referral source, it is ChatGPT. Second, do not confuse market share of the chatbots themselves with referral traffic. A model can have hundreds of millions of users and still send comparatively few clicks, because much of the value is delivered inside the chat window without a link out. ChatGPT wins on referrals partly through scale, with roughly 810 million monthly active users per Sensor Tower data reported by Digiday, and partly through how often its answers surface and link to source pages.

One caution on the smaller players. Claude's March 2026 figure of 2.91% more than doubled from 1.37% the month before, but weekly data showed that surge peaking at 3.6% in mid-March before falling back, which suggests it was driven by a news cycle rather than a durable shift. Treat single-month jumps at the bottom of the table with care.

Why is the number two spot contested?

This is the single most misunderstood point in AI referral analysis. The ranking below ChatGPT depends entirely on how you measure. Raw-visit panels such as StatCounter and Similarweb weight by total referral volume, and on that basis Gemini is second, powered by Google's integration across Search, Android, Chrome and Workspace. But studies focused on business and SaaS referrals tend to tell a different story, with Claude ranking higher because it over-indexes on technical and business audiences who use it for work.

So two things can be true at once: Gemini is the number two consumer AI referrer by raw volume, while Claude punches above its overall weight for B2B audiences. When someone quotes you an AI referral share figure, the first question to ask is which population it describes. A lifestyle or news publisher should lean on the consumer volume numbers; a developer-tools or B2B SaaS site should weight business-audience data more heavily. A single blended "AI referral share" number hides which of these two worlds you actually operate in.

Does Google Gemini send referral traffic?

Yes, and it is the biggest growth story in this space. Gemini rose from 2.31% of AI referrals in March 2025 to 8.65% in March 2026, overtaking Perplexity to become the number two AI referral source in early 2026. Over a comparable window, Similarweb data reported by Digiday showed Gemini's referral traffic growing about 388% year over year, versus ChatGPT's 52%. That is a very steep curve, and it is largely down to Gemini being baked into products people already use every day rather than a standalone app they choose to open.

There is a technical wrinkle that publishers need to understand about Gemini, because it changes how you control and measure it. Google's control for AI training, Google-Extended, is a robots.txt user-agent token, not a separate crawler with its own request string. As Google's documentation puts it, "Google-Extended doesn't have a separate HTTP request user agent string." It lets you opt out of Gemini Apps and Vertex AI training and grounding, and Google states that using it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal." Crucially, the underlying crawling is still done with the standard Googlebot user-agent. That means you cannot cleanly separate Gemini crawling from Search crawling by user-agent, and you certainly cannot block Gemini's fetches by disallowing Googlebot without also deindexing yourself from Google Search. For Gemini, the training and grounding controls sit on top of your core search crawling, so tread carefully.

Why does ChatGPT sometimes send no traffic?

Even though ChatGPT is the top referrer, plenty of publishers see very little ChatGPT traffic in their reports. There are three separate reasons, and they are worth teasing apart because the fixes differ.

1. You blocked the wrong bot. OpenAI runs three distinct crawlers, each with a different job:

• GPTBot (user-agent GPTBot/1.3) fetches pages to train models. IP ranges are published at openai.com/gptbot.json. OpenAI notes that disallowing it signals your content should not be used for training.
• OAI-SearchBot (OAI-SearchBot/1.3) surfaces pages so they can be shown and cited in ChatGPT search. IPs at openai.com/searchbot.json. OpenAI is explicit that "sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers."
• ChatGPT-User (ChatGPT-User/1.0) fetches a page live when a user explicitly asks about it. IPs at openai.com/chatgpt-user.json.

The referral traffic flows from OAI-SearchBot and ChatGPT-User, not GPTBot. A lot of publishers, worried about training use, blocked GPTBot and assumed they had opted out of the training grab while keeping the shop window open. If you accidentally blocked OAI-SearchBot as well, you removed yourself from ChatGPT's search citations, which is the exact mechanism that produces the click. Check your robots.txt: blocking GPTBot alone preserves the referral path; blocking OAI-SearchBot closes it. Full user-agent and verification detail lives in OpenAI's crawler documentation.

2. The referrer header got stripped. Many AI apps and in-app or embedded browsers drop the HTTP referer header. When that happens, a genuine ChatGPT click arrives at your server with no source and lands in GA4's Direct bucket. You did get the traffic; you just cannot see where it came from. This is why sudden, unexplained spikes in Direct can be a proxy signal for AI referrals, and why default dashboards systematically understate AI's real share.

3. The answer never linked out. Sometimes ChatGPT answers a question fully in the chat window and never surfaces your page as a citation at all. That is not a measurement problem, it is the core structural risk of AI search: your content can inform an answer without ever earning a click. Which leads to the next question.

Is being cited by an AI the same as getting a click?

No, and this is the distinction that matters most for revenue. Being cited means an AI answer was built on your page. Getting a click means a human left that answer and landed on your site, where your ads can actually load and your subscription CTA can convert. The gap between the two is where ad-funded publishers lose money, and Google's AI Overviews are the clearest example of that gap in action.

Pew Research Center studied 900 US adults across 68,879 unique searches in March 2025 and found that when a Google AI summary was present, users clicked a traditional search result in only 8% of visits, versus 15% without a summary. They clicked a link inside the summary itself in just 1% of visits. And 26% of AI-summary pages ended the browsing session, compared with 16% for pages with only traditional results. Ahrefs, studying 300,000 keywords, found that AI Overviews reduce the organic click-through rate for position-1 content by 58%, with position-1 CTR on AI Overview keywords falling from 0.073 to 0.016 between December 2023 and December 2025.

So citation without a click is not a hypothetical. It is the dominant outcome of AI Overviews, and it decouples your visibility from your traffic. You can be more visible than ever inside AI answers while your sessions and ad impressions fall. That is precisely why the AI traffic you do capture is worth more per visit.

The traffic is tiny, but it is your best-converting traffic

Keep the scale in perspective. AI referral traffic is still only about 1% of total web traffic across major industries. In June 2025, all AI platforms combined drove 1.13 billion referral visits, up 357% year over year, while Google search sent about 191 billion. AI referrals grew fast but from a small base, and they are nowhere near replacing search volume.

What makes them worth chasing is quality. Analysis of 1,200-plus publisher and news sites using Microsoft Clarity data, reported by Digiday, found LLM referral traffic converts far better than search: 1.66% sign-up conversion versus 0.15% for search, and 1.34% subscription conversion versus 0.55% for search. A visitor arriving from an AI answer has often had their question partly resolved and is clicking through with clear intent. For a publisher with a paid tier or a newsletter, one AI-referred visitor can be worth several from generic search. The strategic move is not to obsess over AI's small share of your pageviews, but to protect and grow the highest-intent slice of your audience.

How can I see which AI engines refer traffic to my site?

This is where most publishers get an incomplete picture, because GA4's defaults do not capture AI properly. GA4 now auto-groups ChatGPT, Gemini and Claude under an "AI Assistant" channel using an "ai-assistant" medium. That sounds tidy, but it leaves Perplexity and Copilot sitting in the generic Referral channel, and it does nothing about the referrer-stripping problem that dumps real AI clicks into Direct. Rely on the default report and you will undercount AI and misattribute the rest.

The fix is a custom channel group in GA4 with a Source regex that covers every major AI source, ordered above the standard Referral channel so it captures those hits first. A production-grade pattern covers, at minimum:

• chatgpt.com, chat.openai.com, openai.com
• perplexity.ai
• claude.ai
• gemini.google.com
• copilot.microsoft.com
• deepseek.com, grok.com, meta.ai, you.com

This matters because the four leaders, ChatGPT, Claude, Gemini and Perplexity, account for the overwhelming majority of measurable AI referrals, but GA4's defaults miss two of the wider set. A regex covering those four plus Copilot captures nearly all of the signal. Without it, you are looking at a materially wrong number.

Two limits to be honest about. First, GA4 can only report clicks it can see, so referrer-stripped AI traffic in Direct will still be undercounted; treat unexplained Direct spikes as a soft signal rather than proof. Second, Google Search Console does not expose AI Overview or AI Mode clicks as a separate report. Those impressions and clicks are folded into standard web Search performance, so you cannot isolate AI-Overview-driven CTR loss in GSC directly. You have to infer it from CTR decline on queries you know trigger an Overview.

Building and maintaining that regex, then reading it against your revenue, is exactly the job Ramprt does for free from your existing GA4. It reads read-only, adds no tracking script, and shows AI referrals by engine alongside what that traffic earns versus your site average. If you would rather set it up yourself, our guide on how to track AI traffic in GA4 walks through the channel group step by step, and the AI crawlers explained piece covers the bot side.

Do not forget the crawler side

Referral traffic is only half the story. Before any AI can cite you, its crawler has to be allowed to fetch your pages. There are four crawler families to recognise and verify: OpenAI (GPTBot, OAI-SearchBot, ChatGPT-User), Perplexity (PerplexityBot and Perplexity-User, documented with IP ranges in Perplexity's crawler documentation), Google (the Google-Extended token, with Gemini grounding piggybacking on Googlebot) and Anthropic (ClaudeBot). Add all four to any allow or deny list and to your log analysis, including ClaudeBot even though Claude's share is low in raw-volume panels, because it ranks higher for some business audiences.

One rule for verification: match the user-agent AND check the source IP against the vendor's published JSON, using reverse-DNS or IP-range checks. User-agent strings are trivially spoofed, which is exactly why OpenAI, Perplexity and others publish verifiable IP ranges. If you are deciding what to block, remember the OAI-SearchBot lesson: blocking a training bot is a training decision, but blocking a search or grounding bot is a traffic decision, and the two often live in the same robots.txt.

Putting it together for your site

ChatGPT is where the AI referral traffic is, so make sure OAI-SearchBot and ChatGPT-User can reach your pages. Gemini is the growth curve to watch, but you cannot separate it from Google Search at the crawler level, so manage it through Google-Extended and accept that its referrals ride on your broader Google visibility. Perplexity is real but slipping, and Claude punches above its raw-volume weight for business audiences. Across all of them, the traffic is small but converts far better than search, and AI Overviews are quietly cutting the search clicks you used to rely on. The one thing you should not do is trust GA4's default AI reporting. Build the custom channel, read AI referrals against revenue, and you will finally be measuring the shift that is happening to your traffic rather than guessing at it.