ChatGPT now sends 92.4% of all trackable AI referral traffic to the open web. That's the number in Previsible's "2026 State of AI Discovery Report," published July 6, 2026, built from 6.77 million sessions across 166 GA4 properties tracked between November 2024 and May 2026. Over that stretch, ChatGPT's referral traffic grew 12.8x. Claude grew 64x and passed Perplexity in March 2026. Perplexity is down 61% from its own peak. Microsoft Copilot has fallen 96% from where it stood in August 2025. Read that data on its own and the story looks settled: ChatGPT won, and everyone else is splitting what's left. Read it next to a different report, tracking a different number, and the story gets more useful. Usage and referral traffic are not the same metric, and the gap between them is the actual finding here.
The AI referral traffic numbers Previsible just published
Previsible's report is the largest referral-traffic dataset published on this question to date: 6.77 million sessions, pulled from 166 GA4 properties, covering November 2024 through May 2026. The topline number is stark. ChatGPT commands 92.4% of all trackable LLM referral traffic across that dataset, and its share has climbed the entire time the study covers — a 12.8x increase in raw referral volume in 19 months. Search Engine Land's coverage and Previsible's own writeup both frame this the same way: whatever debate exists about which AI engine people prefer to use, there's very little debate left about which one actually sends visitors to your site when it cites you. That's a different question than usage share, and it's the one that shows up directly in your web analytics every time someone lands on a page carrying an LLM referral signature.
"Trackable" is doing real work in that 92.4% figure, and it's worth being precise about what it means. GA4 attributes an LLM referral when a visit arrives with a recognizable referrer or UTM signature from a known AI engine's domain. Sessions that arrive with no referrer at all — a growing share of AI-driven traffic across every engine — don't show up in this dataset, on any engine. That doesn't undercut the finding; it means the true scale of AI-driven traffic to the open web is larger than 6.77 million sessions, and 92.4% describes the share of the visible, attributable portion. Within that portion, ChatGPT isn't just ahead. It's most of the market.
The other three engines in that dataset tell three different stories, not one. Claude's 64x growth is the standout number, and it happened fast enough to overtake Perplexity's referral traffic in March 2026 — a crossover that would have looked unlikely a year earlier, when Perplexity was the engine most GEO teams were told to prioritize right after ChatGPT. Perplexity's 61% decline from its own peak isn't a rounding error; it's a reversal, on a platform that spent 2025 positioned as the search-native challenger. And Copilot's 96% collapse from its August 2025 levels means Microsoft's assistant, for referral-traffic purposes, has become close to irrelevant. None of those three moved together. They moved in three different directions, on three different timelines.
AI referral traffic vs. AI search usage share: two different numbers
Set the Previsible numbers next to Kevin Indig's "AI Halftime Report: H1 2026," published on Growth Memo the day before, and the picture stops being simple. We covered that report in detail in AI search market share 2026: ChatGPT's share of AI-search activity — Indig's term for who actually opens the app or the box and asks a question — fell from 78% to 56% between January and July 2026, a 22-point drop in six months. Gemini rose to 30% over the same window. Claude rose to 10%. Both reports cover almost the identical period. Both are credible, independently produced, and neither contradicts the other on the facts. What they disagree on is which question they're answering. Indig is measuring where the questions go. Previsible is measuring where the clicks come from. ChatGPT is losing ground on the first number and gaining on the second, at the same time, inside the same six months.
| METRIC | CHATGPT FIGURE | WHAT IT MEASURES |
|---|---|---|
| AI search usage share (Growth Memo, Jul 27, 2026) | 56% (down from 78%) | Share of AI-search activity — who gets asked |
| AI referral traffic share (Previsible, Jul 6, 2026) | 92.4% | Share of trackable LLM referral sessions — who sends the click |
“Usage share tells you who's getting asked. Referral share tells you who's actually sending the click when a source gets named. ChatGPT's lead is shrinking on the first number and hardening on the second — and the second is the one that shows up in your analytics.”
The mechanism behind that split isn't fully explained by either report, so treat this as a read, not a sourced fact: usage share and click-through behavior aren't the same lever. Gemini's usage gains are concentrated inside Google's own surfaces — Search, AI Mode, Workspace — where a synthesized answer frequently satisfies the query without sending a visit anywhere, including to Gemini's own cited sources. ChatGPT's interface, by contrast, is built around a conversation that regularly pushes the user out to a source link, and its userbase has spent three years developing that habit. A rising usage share inside an interface that discourages clicking doesn't translate into referral traffic the way a shrinking-but-still-dominant usage share does inside an interface that encourages it. If that's right, the gap between these two numbers isn't a data artifact. It's a structural difference in how each engine's interface handles the moment after it decides to cite you.
Claude's 64x climb and Perplexity's collapse
Claude's 64x referral-traffic growth is the number worth watching longer than any other in this dataset, because it's compounding off a small base and it hasn't leveled off. We flagged the same pattern from the usage-share side in AI visibility tracking and the market-share swings underneath it: Claude's gains concentrate in enterprise and technical workflows rather than general consumer search, which is a different growth mechanism than Gemini's default-placement advantage or ChatGPT's incumbency. A referral-traffic curve growing 64x in 19 months, off that kind of base, behaves differently than a mature engine defending share. It's the number in this report most likely to look dated in six months, in either direction — and the one a GEO program should be re-checking most often, not the one it should assume is settled.
Perplexity and Copilot are the cautionary numbers in the same dataset. Perplexity spent 2025 as the engine most GEO advice told teams to prioritize second, right after ChatGPT — a search-native product with a citation format built for exactly this kind of attribution. A 61% decline from its own referral-traffic peak says that positioning didn't hold, at least not on this metric. Copilot's 96% collapse from August 2025 levels is starker still: whatever integration advantage Microsoft has across Windows and Office hasn't translated into people clicking through from Copilot answers to the open web. Both numbers argue against building a GEO program around any single engine's current citation quirks — including ChatGPT's — because the two engines that collapsed here were, a year ago, each somebody's safe second bet.
Where AI referral traffic lands, by industry
Previsible's dataset also breaks referral traffic down by where it lands on the site, segmented by industry, and the pattern is different enough across categories that a single GEO playbook won't serve all of them. SaaS sites captured 34.6% of their AI referral traffic on search and category pages, the pages that answer "what tools exist for X" rather than "what does this specific product do." Publishing sites saw 54% of their AI traffic land on news content — the highest concentration of any category in the dataset, and a signal that AI engines treat time-sensitive reporting as a primary source, not a supplementary one. Ecommerce sites got 43% of their AI referral traffic on product pages directly, meaning the engines are increasingly skipping the category-browse step and citing the specific SKU.
Share of AI referral traffic landing on this page type, by industry (Previsible, 2026 State of AI Discovery Report)
That spread matters for where you put structural work. A SaaS site optimizing only its product docs, while its search and category pages sit unstructured, is defending the wrong 34.6%. A publisher treating news content as disposable once it ages out of the homepage is ignoring the exact format driving the majority of its AI traffic. We've made the extractability argument before in the anatomy of an AI citation — clear structure, a direct answer near the top, a table an engine can lift — and this data adds the missing layer: which page type gets structured first depends on your industry's actual landing pattern, not a generic checklist. The per-engine GEO framework we use with clients starts from exactly this kind of segmentation before it starts prescribing tactics.
What this means for your GEO budget allocation
Here's the position this data actually supports, and it cuts against the instinct a lot of teams will have after reading Indig's usage-share numbers. Seeing ChatGPT's usage share drop 22 points, the obvious move is to shift GEO budget toward Gemini and Claude and treat ChatGPT as a shrinking priority. The referral-traffic data says that's wrong, or at least premature. ChatGPT sends 92.4% of the clicks. A citation you win on Gemini, inside an interface with heavier zero-click behavior, is currently worth less in actual visits than a citation you win on ChatGPT, even as Gemini's usage share climbs. That's not an argument for ignoring Gemini — 30% usage share is real demand, and it will eventually pull more referral traffic behind it as the interface matures or as measurement improves. It's an argument for not moving budget away from your best-performing channel because a different metric, measuring a different thing, moved first.
None of this works without measuring both numbers side by side, on a recurring basis, against your own site's data rather than an industry aggregate. That's the gap between reading a report like this one and running a program off it. Our generative engine optimization work builds citation and referral tracking into the same dashboard from week one, specifically so a swing like the one in these two reports shows up as a line moving on your own chart, not a headline you read a month later. Reporting built around pipeline, not just traffic, is what turns that swing into a budget decision instead of a talking point.
Do this next
Pull your GA4 referral data segmented by LLM source for the last two full quarters. Check what share of your AI-driven sessions actually came from ChatGPT versus Claude, Perplexity, and Copilot, and compare that against how your GEO effort is currently allocated across those four engines. If your content and outreach work is still weighted toward Perplexity because that's where the playbook pointed a year ago, this data says redirect it. If Claude is a rounding error in your current program despite 64x growth in the underlying channel, that's the gap to close next, before it compounds further. Segment your own landing-page data by content type the way Previsible segmented by industry, and confirm the pages actually getting AI traffic are the ones getting the structural investment. We ran this exact reallocation with Zenity, moving budget off a single engine's quirks and onto the metric that was actually moving. Do the same audit on your own data before next quarter's plan gets written.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.