Is ChatGPT losing traffic share in 2026? That headline is about the wrong number

July 29, 2026

The viral "ChatGPT dropped to ~54%" headlines measure how many people visit chatgpt.com versus gemini.google.com — not how much traffic those assistants send to your website. Those are two different numbers, measured different ways, and only one of them describes your audience: the referral share in your own first-party log.

So is ChatGPT losing its grip in 2026? By one common yardstick, yes — Similarweb's worldwide web-visit data put ChatGPT at 53.9% of assistant visits in May 2026, down from roughly 76% a year earlier. But that figure counts visits to the assistants' own domains. It says nothing about the split of AI referral traffic arriving on your pages, which depends on what your visitors read, what they ask, and which assistant surfaced you. If you want to know who sends you traffic, the global headline can't tell you — your referrer log can. This is a follow-up, so if you haven't set that tracking up yet, start with how to see ChatGPT, Perplexity and Gemini traffic in your analytics; this post is about reading the numbers once you have them.

Usage share is not your referral share

Here's the trap. The most-shared 2026 stat is usage share — the proportion of visits or app sessions across the major assistants. It's a real, useful measure of where the AI market is going. But it is not a measure of referral traffic, and treating it as one is like reading a TV network's ratings to guess who links to your blog.

The May 2026 mix, as estimated by Similarweb's worldwide web-visit data (published July 2026, and echoed in Momentic's "Top AI Chatbots" roundup):

Two caveats travel with that table, and they matter. First, it's a web-visit measure across seven assistants — it undercounts app and embedded usage, by an estimated 15–60%, so the real picture skews further from the browser than these numbers show. Second, change the methodology and the numbers move: Sensor Tower's True Audience data — unique users across mobile apps and the web — through the end of the same month put ChatGPT at 46.4%, Gemini at 27.7%, and Claude at 10.3%. Same month, different measuring stick, an eight-point gap on ChatGPT. Neither is "wrong" — they're counting different things (web visits alone versus unique users across apps and web). That disagreement is the whole point. No global share figure is your data, because none of them measured your site.

The number that is your data has one source: your own first-party referrer log. It's the only one that reflects your topics, your audience, and your geography.

The May 7 change that quietly moved AI referrals to your homepage

The other thing the headlines miss is where AI referrals now land. On 7 May 2026, OpenAI started surfacing prominent, clickable brand-name links in ChatGPT answers instead of burying sources in footnote-style citations. The effect on referral patterns was fast and large, and it's been corroborated across several independent trackers (Similarweb, SE Ranking, Search Engine Roundtable, Profound, 33 Sticks).

According to those reports, ChatGPT's homepage share of its own referral clicks jumped from roughly 26–32% before the change to about 60–63% within days, and held there. In the same window, homepage referrals rose about 354.7% and total ChatGPT referrals rose about 157% — a near-tripling.

The practical consequence for your analytics: AI-driven visits increasingly hit your homepage rather than the specific page the assistant read. If you assume a referral lands on the URL that got cited, your per-page attribution is now distorted — a brand-link click to your root domain looks like homepage interest when it's really "the assistant recommended us and the visitor arrived at the front door." When you see AI referrals concentrating on / after early May, that's the branded-links change showing up in your data, not a shift in what people wanted.

What you can actually check in your own log

This is where honesty matters more than a sales pitch. Simplytics classifies every referrer into a channel — Direct, Search, Social, AI, Email, or Referral — and the AI channel recognizes each assistant as a distinct referrer host: chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, deepseek.com, grok.com, meta.ai, and more. So the assistants that are gaining in the estimates above — Gemini, Claude, Grok, DeepSeek — are already broken out for you, and you can watch your mix shift instead of inferring it from a worldwide chart.

One deliberate detail: AI is matched before Search, so gemini.google.com is counted as AI rather than folded into Google Search. That ordering is where a lot of tools quietly lose the signal. As covered in what GA4's native AI Assistant channel still hides, GA4 still files Perplexity under Referral, counts AI Overviews and AI Mode as Organic Search, and keeps its recognized-referrer list undocumented and expanding — so the AI total you read there is missing pieces you can't see.

Two honest limits, because overclaiming would defeat the point. Simplytics shows each assistant as its own referrer and gives you an AI channel total, in your own first-party log — it does not ship a built-in market-share stat, and it doesn't have a dedicated "which AI referrals hit your homepage" cross-tab. What it does have is a Sources (referrers) report and a Pages report; to spot the homepage effect for yourself, watch two signals together after early May — your AI channel total climbing, and / rising in your Pages report at the same time. The tool won't cross-tab those for you, so it's a correlation you read, not a number a dashboard hands you.

And Simplytics isn't the only tool that breaks out AI referrers — Plausible, Fathom and Simple Analytics do too. The honest differences are the ones we'll stand behind: it's cookie-less with no consent banner, the data lives in the EU (Warsaw), each assistant is its own referrer in a clean AI channel with no GA4-style misfiling, and it runs $1/month rather than the $9–15/month typical of hosted privacy tools. A lot of functionality for a lot less money — see the side-by-side with Google Analytics for the rest.

The bottom line: the "ChatGPT is slipping" story is about usage share of the assistants' own domains — a third-party estimate (53.9% by Similarweb, 46.4% by Sensor Tower, same month) that says nothing about who sends traffic to your site. Since 7 May 2026, more of that AI traffic lands on your homepage than on the page that was cited. Both facts point the same way: stop reading the global headline as if it were your audience, and read your own referrer log, where the only share that describes your visitors actually lives.