Your AI referral mix is not the market's

An AI referral share is the percentage of a site's visits from AI assistants that each assistant sends, and the published versions of that number disagree so widely that none of them can stand in for your own.

That matters because referral share is the figure most teams reach for when they decide which engine to optimize for first. If the benchmark says one assistant sends almost everything, the plan writes itself. The trouble is that three credible 2026 datasets describe three different markets, and the differences are too large to be noise.

#Three panels, three answers

Statcounter measures AI chatbot referrals across more than 1 million websites and over 3 billion page views a month. For March 2026 it reported ChatGPT at 78.16%, Gemini at 8.65%, Perplexity at 7.07%, Copilot at 3.19% and Claude at 2.91%.

BrightEdge reported on September 24, 2026 that ChatGPT's share of AI referral traffic to brands reached 95.1% in August, after a 2026 low of 89.5% in April. In August it put Gemini at 2.4%, Perplexity at 1.1% and Claude at 0.8%. The release does not describe which sites or industries sit behind the figures.

Goodie runs a GA4 panel of mostly B2B brands and averages shares per brand, so each brand counts equally. For March and April 2026 it reported ChatGPT at 62.6%, Claude at 18.5%, Gemini at 10.6%, Perplexity at 7.3% and Copilot at 4.0%.

Put Claude side by side. BrightEdge says it sends under 1% of AI referrals, Statcounter says about 3%, and Goodie says close to a fifth. ChatGPT ranges from 62.6% to 95.1%. The time windows are not identical, and BrightEdge's own series moved several points within the year, but no seasonal shift explains a gap of that size. The panels are measuring different sets of websites with different audiences.

#Why the panels disagree

Three choices drive most of the spread.

Who is in the sample. A panel of B2B software brands draws visitors who use assistants at work, and that audience is not the same as the general population reaching a million mixed websites. Goodie's authors say directly that their panel may overindex on Claude and Perplexity audiences, and they decline to break the data out by industry because the samples inside each category are too small.

How sites are weighted. Goodie gives each brand equal weight. A census of page views, like Statcounter's, lets the largest sites dominate. If a handful of high-traffic consumer sites get most of their AI visits from ChatGPT, a traffic-weighted share will lean that way even if many smaller sites look different.

What counts as an AI referral. Every one of these numbers depends on the assistant passing a referrer that analytics can recognize. Statcounter excludes Grok because it does not provide referral data in its header. Goodie classifies sessions by matching source and medium against named AI domains, and its authors note that about 31% of the panel's sessions arrived as direct traffic, some of which may be AI visits that lost their referrer.

#Usage share does not predict referral share either

A natural fallback is to rank engines by how many people use them. Goodie's report shows why that also misleads. Using Similarweb visits for January to April 2026, it found Gemini had 29.0% of visits to AI platforms but a normalized referral share of 10.3%, while Claude had 1.29% of platform visits and an 18.0% referral share in the same panel.

An assistant's audience size tells you how many answers it gives. It says little about how often those answers send someone to a website like yours. That depends on what people ask it, how it presents sources, and whether the person asking is in a buying or research mode that leads to a click.

#The biggest AI surface is missing from all three

None of these datasets includes Google's AI Overviews or AI Mode. Google's documentation says sites appearing in AI features are included in the overall search traffic in Search Console, reported under the Web search type, with no separate referrer. Goodie notes that in GA4 these visits arrive as google / organic and cannot be told apart from classic search clicks, so its study excludes them.

Any chart titled "share of AI referrals" is therefore a share of the assistants that identify themselves. For many sites the largest generative surface they appear on is not in the denominator at all.

#What to do with this

Use benchmarks for direction, not allocation. All three panels agree that ChatGPT leads and that more than one assistant now sends meaningful traffic. That is a reasonable conclusion. Deciding that a given engine deserves a tenth of your effort because a press release said 2.4% is not.

Measure your own mix first. Pull AI referral sessions by source from your analytics for at least the last full quarter. Define the source list explicitly and write it down, including how you treat Copilot, Gemini web versus app, and any assistant that strips referrers. A small site will have small counts per engine, so report the raw sessions next to the percentages.

Segment by page type and audience. A documentation section, a pricing page and a blog can have very different engine mixes. If your business serves developers or other technical buyers, check whether a panel built on a similar audience matches you better than a general census.

Keep Google's AI features in a separate line. Track AI Overviews and AI Mode through Search Console's AI feature reporting and your own citation monitoring, not through referral share. Do not add them into an assistant referral total that cannot see them.

Compare referral share with citation share. An engine that cites you often but sends few visits points to a presentation or intent issue. An engine that sends visits from few citations points to high-intent answers worth protecting. Neither pattern shows up if you only look at a market benchmark.

Re-check after model and interface changes. Referral behavior moves when an assistant changes how it shows sources. BrightEdge's own figures moved from 89.5% to 95.1% for ChatGPT between April and August 2026. Annotate product launches in your reporting and compare before and after on your own data.

The published shares are honest descriptions of the panels behind them. Your site is one more panel, and for planning purposes it is the only one that counts.