Which AI engines send your traffic? Why ChatGPT's share ranges from about half to nearly nine in ten

Why the published numbers disagree

ChatGPT's share of AI traffic depends on what a study counts. Count visits to AI tools, and ChatGPT holds about half. Count referrals from AI tools to other websites, and it sends roughly three in four. Narrow the sample to one sector or one period, and the share moves again.

That is not a sign that one study is wrong. Each answers a different question. Similarweb, which sells web analytics, measures where people go to use generative AI. SE Ranking, which sells SEO software, measures where the AI visits that reach ordinary websites come from. Both are useful, and neither is your number.

Our view: most of the confusion in board decks this year comes from quoting one study's share as if it described a company's own traffic. This post lines up five 2026 datasets, explains why they diverge, and sets out a method for reading the split by engine in your own analytics.

Visits to AI tools are not referrals from them

The first and largest source of disagreement is the unit. A visit to chatgpt.com is a person using ChatGPT. A referral is that person clicking a link in an answer and landing on someone else's site. The two numbers move differently, because engines differ in how often they link out and how often people click.

Similarweb's numbers are about the first unit. Its AI search statistics, published in July 2026, show ChatGPT's share of worldwide generative AI web traffic falling from about 76% to about 53% between June 2025 and May 2026.¹ Over the same period Gemini's share rose from under 9% to about 27–28%, and Claude's from about 2% to about 9%.¹ Similarweb notes that ChatGPT's absolute visits stayed roughly flat. Its share fell because rivals grew.

SE Ranking's numbers are about the second unit. Across 101,574 websites in 250 countries and territories, measured through Google Analytics from January to April 2026, ChatGPT sent 74.78% of AI referrals.³ Gemini sent 11.56%, Perplexity 7.23%, Copilot 3.51% and Claude 2.62%.³

Two different questions, two different shares

Share of AI tool traffic

• Counts visits to chatgpt.com, gemini.google.com and others
• Measures where people use AI
• Panel or clickstream data
• Tells you where your buyers ask

Share of AI referrals

• Counts clicks from AI answers to websites
• Measures which engine sends visitors
• Site analytics data
• Tells you where your visits come from

Read together, the two sets say something neither says alone. Gemini is used far more than its referral share suggests. A plausible reason is that much Gemini use happens alongside Google Search, where people who want a website click a search result instead of a link in the answer. That is our reading, not a finding of either study.

How small is AI traffic, really?

Small, and growing fast. In SE Ranking's sample, AI search accounted for 0.32% of all visits in January to April 2026, against 0.02% in 2024.² That is sixteen times larger in two years, and still about one visit in three hundred.

The small base is why shares swing. When AI traffic to a site is a few hundred visits a month, one engine's update, one viral answer or one new citation can move an engine's share by many points. A share computed on a small base needs a note on the base.

It is also why referrals undercount influence. Many people read an answer, then search for the brand or type its address. Those visits arrive as branded search or direct traffic, not as AI referrals. Our note on how ChatGPT mentions turn into search visits covers that path. Our view: treat AI referrals as the visible tip of AI influence, and never as its full size.

Panels, sectors and periods move the share again

Once the unit is fixed, three more choices separate the studies: whose sites are in the sample, which sectors they belong to, and which months are measured.

Whose sites

SE Ranking's sample is broad: a large number of sites across most countries. Conductor, which sells an AEO and SEO platform, measured AI referrals across 1,215 of its US enterprise customers' domains. Octane11 is a B2B marketing analytics platform. Its quarterly AI search series measures business-to-business websites only.⁶ Each panel reflects its customers, and each vendor sells in this category.

Which sectors

Sector changes both how much AI traffic arrives and which engine sends it. Similarweb found that the share of US ChatGPT prompts whose answer carried a citation rose from 1.6% in June 2025 to 6.8% in May 2026, with travel at about 23% and professional services under 4%.¹ An engine that rarely cites in your sector cannot send you many referrals there, however often it is used. Our note on citation rates by sector covers this in detail.

Which months

Period matters because the market moved fast between mid-2025 and mid-2026. A dataset that ends in September 2025 describes a market before Gemini's rise. One that ends in May 2026 describes the market after it. The business-to-business series cited above tracks each engine's share of AI referrals to business-to-business sites month by month across most of the past year, which shows how much a single snapshot depends on the month it was taken.⁶ Two snapshots taken a few months apart can disagree without either being wrong.

Five datasets side by side

Put the studies in one table and the spread makes sense. Each row counts a different thing, in a different panel, over a different period.

Sources as footnoted. Each vendor sells products in this category.

Conductor's figure is the high end. Its 2026 benchmarks report, first published in November 2025, found ChatGPT sent 87.4% of AI referrals across those US domains in data from May to September 2025, nearly a year old by this post.⁴ AI referrals made up 1.08% of visits in that sample, with IT the highest sector at 2.80%.⁴ The older period and a different panel both help explain the gap with SE Ranking's 2026 data. We set out why the report's dates matter in our note on Conductor's benchmark dates.

ChatGPT's share depends on what is counted

Referrals, 1,215 US domains, May–Sep 2025⁴87.4%

Referrals, 101,574 sites, Jan–Apr 2026³74.78%

Visits to AI tools, worldwide, May 2026¹About 53%

The engine no referral study counts properly

Google's AI features are the largest blind spot in every referral study. A click from an AI Overview or from AI Mode usually reaches your site as Google organic traffic, with no separate referrer. Referral studies therefore count Gemini's standalone app and website, not Google's AI answers inside Search.

The scale is large. Alphabet said on 22 July 2026 that AI Mode had passed 1 billion monthly active users and the Gemini app 950 million.⁷ It also said Google is sending billions of clicks to websites every week through AI features in Search.⁷ Those are company statements without a breakdown, but they describe traffic that most analytics setups file under organic search.

AI Overviews appear often enough to matter. Conductor found them on 25.11% of 21.9 million US searches in data from 15 September to 12 October 2025.⁵ If a quarter of your organic queries show an AI Overview, part of your "organic" traffic is AI traffic that no referral share captures. Our guide to tracking AI Overviews and AI Mode separately sets out what Search Console can and cannot show.

Our view: any engine split that leaves out Google's AI features overstates ChatGPT's role for sites that depend on Google. That is most sites. When you present the split, say which Google traffic it excludes.

How to read your own split by engine

Your own analytics answer the question the studies cannot: which engines send your visitors. Set it up once, keep the rules fixed, and read it monthly.

Reading your AI traffic by engine

1. Fix a referrer list

   One row per engine domain: chatgpt.com, gemini.google.com, perplexity.ai, copilot.microsoft.com, claude.ai and the rest. Date it.

2. Make a channel

   Group those referrers into an AI channel, separate from referral and organic, and keep the engine as a dimension.

3. Report the base

   Show sessions per engine next to each share, so a swing on a small base is visible.

4. Add landing pages and conversions

   Which pages AI visitors reach, and what they do next.

5. Set it beside share of answer

   Compare each engine's referrals with how often that engine names you on a fixed prompt set.

Two details save trouble. Some engines add a source tag to the links they send, which helps attribution, but not every link carries one, so keep the referrer list as the main rule. And some app traffic arrives with no referrer at all, so a share of "AI referrals" is always a share of the AI visits you can see.

The comparison with share of answer is where the split becomes useful. At Sigzen AI, our AI answer monitor checks six engines (ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI), through official APIs where they are offered and a licensed AI-answer data provider otherwise. If an engine names you often but sends few visits, the answer may lack a link, or the link may point at the wrong page. If it sends visits but rarely names you, a few strong citations are carrying you.

Four misreadings to avoid

Most errors with these studies come from four habits. Each one is easy to spot once you know the unit, the panel and the period behind a number.

Treating a share as a size

A falling share is not a falling engine. Similarweb's data shows ChatGPT's share of AI tool traffic dropping while its visits stayed roughly flat.¹ If you cut ChatGPT work because "its share is collapsing", you cut work on an engine that sends as many people as before. Ask for the absolute count before acting on any share.

Mixing units in one chart

A slide that puts Similarweb's usage share beside a referral share from your own analytics compares two different things. The gap between them is real information, as the Gemini example above shows, but only when the chart labels each bar with what it counts.

Dropping the period

A share quoted without months cannot be compared with anything. The Conductor and SE Ranking figures end seven months apart, within a year in which Gemini's share of AI tool traffic roughly tripled.¹ Quote the months every time, including in a footnote on a slide.

Assuming your site matches the panel

A panel of thousands of sites averages out exactly the differences that matter to you. A travel site, where ChatGPT cites often, and a law firm, where it rarely does, can see very different engine splits in the same month. Our view: an industry benchmark tells you whether your split is unusual, and only your own data tells you what it is.

Which number belongs in the board report?

Put your own numbers in the board report and use the studies only for context. A board needs three things: AI referrals as a share of your sessions, the split by engine with the base shown, and your share of answer on a fixed prompt set. All three trended monthly, with the method stated once.

Quote an external study only to answer "is this normal?", and then quote it with its unit, panel and period. "ChatGPT sent about three in four AI referrals across 101,574 sites in January to April 2026" is honest.³ "ChatGPT has 75% of the market" is not, because it mixes units and drops the date.

Our view: when AI traffic is under one percent of sessions, pipeline is the stronger argument. Tie AI-referred sessions to leads and opportunities where you can, and say plainly where you cannot. Our guide to proving AI visibility in pipeline covers the B2B version, and our board-report metrics piece covers the zero-click side.

What would change the picture

Three developments would move the shares again. First, a change in how often an engine links out: the share of ChatGPT answers carrying a citation already rose several-fold in a year,¹ and another change would shift referrals without any change in use. Second, Google reporting its AI features separately in Search Console, which would bring the largest hidden source into view. Third, more AI use through apps and agents that send no referrer, which would make every referral share less complete.

None of these would change the method. Fix the referrer list, show the base, separate Google's AI traffic where you can, and set referrals beside share of answer. The studies will keep disagreeing, and your own split is the one that should guide your decisions.