For two years the advice was simple: get cited by ChatGPT and the rest will follow. That advice has quietly become a liability.
ChatGPT still sends more AI referral traffic than every other engine combined. But its share of that traffic is falling — fast — and not because it is shrinking. It is falling because everything around it is growing faster. When one engine holds ninety per cent of a channel, optimising for it is a strategy. When it holds sixty-three per cent, it is a bet.
We call this the ChatGPT cliff: the moment a channel leader stops being the channel. Brands that tuned their entire generative search program to one engine are about to find out how much of their visibility was borrowed.
GEO.thejourney.hkThe Cliff, In Numbers
Two independent 2026 studies tracked the same shift from different vantage points — a B2B brand panel of GA4 referrer data, and a global web-traffic analysis across the major AI surfaces. They disagree on magnitude and agree completely on direction.
| Engine | Share of AI referrals, 2025 | Share of AI referrals, 2026 | Direction |
|---|---|---|---|
| ChatGPT | 89.1% of B2B referrals / 79.7% of global AI traffic | 62.6% / 76.0% | Share falling, absolute volume still growing |
| Claude | 1.4% of B2B referrals | 18.5% / 2.6% of global AI traffic | Fastest riser — traffic up ~320% year on year |
| Gemini | Low single digits | 10.6% / 11.6% | Traffic up ~231% year on year |
| Perplexity | 19.7% of US AI traffic in early 2025 | 6.85% of US AI traffic | US share down; global volume flat |
| Copilot | ~4% | ~4% | Flat share, ~31% more volume |
Measurable B2B AI referrals
Where the field sits now — 2026
- ChatGPT62.6%
- Claude18.5%
- Gemini10.6%
- Perplexity7.3%
- Copilot4%
Read the middle column carefully. ChatGPT's own referral traffic kept growing through the same period — BrightEdge measured a 101% rise between January and August 2026 — so the engine is not in trouble. One caveat worth holding as you read: a third dataset, BrightEdge's referral panel, reads the share trend the other way, with ChatGPT climbing back to 95.1% of AI referrals by late 2026 on a much narrower sample of referral events. Which engine leads depends on what you count — which is exactly why you should count it yourself. The problem is structural: the four largest AI surfaces now split close to 99% of measurable AI referrals, and the displaced share has moved to engines with fundamentally different retrieval logic, citation behaviour and user intent.
Put plainly: a strategy built for ChatGPT's citation logic now covers about a third less of the AI traffic landscape than it did twelve months ago.
A Share Shift Is Not Just a Numbers Story
The instinct is to treat this as a marketing problem — reallocate attention, spread the bets, move on. That reading misses the point.
Each engine is not a window onto the same index. They retrieve differently, weigh sources differently, and answer at different levels of synthesis. A page engineered to be quotable inside a ChatGPT answer is not automatically quotable inside Claude's, and it is almost certainly not the page Google AI Overviews decides to lift a paragraph from.
GEO.thejourney.hkThis is the fragmentation we mapped in The Great Fragmentation: Why Your Brand Needs 7 GEO Strategies in 2026 — but with a new consequence. When one engine dominated, a narrow strategy was rational. As the field splits, that same strategy becomes a single point of failure.
The Claude Surge Is an Authority Story
Claude's rise from 1.4% to 18.5% of measured B2B AI referrals is the single largest share movement in the dataset, and it was not driven by marketing spend. It followed product capability: web search with inline citations rolled out broadly through 2025, turning every Claude session into a potential referrer to the open web.
What matters for your strategy is how Claude is used. It holds 1.3% of AI platform visits but 18% of B2B referrals — the signature of a research and source-finding audience rather than one completing tasks inside a product. That audience is looking for material it can check: reference works, professional documentation, published data, named sources. Not keyword density. Verifiability.
That is a different production line from the one most teams have built. It means original data, clearly dated, with named authors and explicit methodology. It means statistics you can point to, not statistics you assert. If your content strategy has been optimising for coverage, Claude is telling you it rewards evidence.
GEO.thejourney.hkThe Perplexity Puzzle Is a Warning About Single-Engine Bets
Perplexity was supposed to be the durable second place. Instead it is the clearest case study in the fragility of platform-dependent visibility. Its US share of AI traffic fell from 19.7% in early 2025 to 11.42% across that year, then to 6.85% in 2026 — while its global traffic stayed almost flat, during a period when the overall AI traffic pool grew by more than a third. Then the panel data inverts it: on Goodie's B2B referrals, Perplexity's share is up roughly 2.4 times since that first wave. Two credible datasets, opposite directions, one engine.
No amount of content quality explains that. Distribution changed. Apple signalled plans to add AI-powered search options to Safari, and reported that Safari search volume fell for the first time as users moved toward AI. When the surface your buyers use is decided in someone else's product roadmap, the only defensible position is presence across several surfaces — not excellence on one.
GEO.thejourney.hkThe engine you are optimising for is not a channel you own. It is a tenancy, renewable at someone else's discretion.
The Measurement Trap
Here is the uncomfortable part: most teams cannot see this shift in their own data, which is precisely why it catches them out.
AI search is overwhelmingly zero-click. A user receives an answer, never visits, and never appears in your analytics. What does surface — referral traffic in GA4, citations in an answer — is a narrow slice of the real exposure. If your reporting is built around sessions, the engines that are growing fastest look smallest, because they are the ones answering most completely inside the interface.
GEO.thejourney.hkSo the first fix is not a content fix. It is a visibility fix: track citation share per engine, per prompt, on a fixed question set, and review it monthly. You cannot rebalance toward engines you are not measuring.
The Three Engines You Are Probably Under-Indexed On
| Engine | What it rewards | Fastest lever |
|---|---|---|
| Gemini | Editorial depth, entity clarity, Google-ecosystem signals | Long-form thought leadership on established editorial platforms; clean entity markup on your own pages |
| Claude | Institutional and data sources — reference works, documentation, published statistics | Original research with named authors, dated figures and stated methodology |
| Copilot | Professional and organisational sources, Microsoft-ecosystem context | Presence in professional publications, association content and business directories |
Google AI Overviews sits across all of this as a fourth surface, and it rewards something else again — being discussed in the places it trusts, Reddit most of all. That is a distribution problem, not a page problem.
Rebalancing in 90 Days
1. Baseline per engine. Take your twenty most valuable questions and record where your brand is cited today on ChatGPT, Gemini, Claude, Perplexity and AI Overviews. Not mentioned versus cited — cited, and at what position. 2. Fix the eligibility layer first. Crawlability, clean structure, entity clarity. Nothing downstream works if the retrieval step fails, and it fails silently. 3. Pick two engines to over-index on. One where you are already visible and can consolidate, one where the growth is and you are absent. Claude and Gemini are the obvious candidates in 2026. 4. Build one engine-specific asset per month. Not a rewrite of an existing page — a new asset shaped for that engine's evidence profile. Data for Claude, editorial depth for Gemini, professional presence for Copilot. 5. Re-measure the same question set. The only honest scoreboard is the same questions, asked the same way, compared month over month.
What the Cliff Actually Means
The story of 2026 is not that ChatGPT is failing. It is that the assumption of a single dominant AI surface — the assumption most GEO programs were built on — no longer holds.
The brands that come through this well will not be the ones that chased every engine. They will be the ones that kept an unshakeable technical foundation and then deliberately fragmented their investment across the surfaces their buyers actually use.
→ Read more in Why GEO is not just SEO with a new name, and see the full side-by-side in SEO vs GEO: a list of links against a cited answer — where we unpack why technical SEO still underwrites every AI surface.
The cliff is not a fall. It is a widening of the field — and the brands still standing on one peak are the ones who will feel it first.
Sources
Every figure in this piece traces to one of the studies below. Where two credible datasets disagree, we say so in the text rather than quoting whichever number flatters the argument.
• Goodie (May 2026). 2026 AI Search Traffic Report: ChatGPT's Grip Slipped, Claude & Gemini Are Surging — the primary B2B referral panel behind the headline table: ChatGPT 89.1% → 62.6%, Claude 1.4% → 18.5%, Gemini 10.6%, Perplexity 7.3%, Copilot 3.2% → 4.0%, and the Big 4 at close to 99% of measurable referrals. Triangulated against Similarweb traffic data and SensorTower adoption signals.
• Goodie (October 2025). AI Search Market Share Report — the Wave 1 baseline: 2,802,519 AI referral sessions across 41 brand sites, May–August 2025, when ChatGPT held 89.1%.
• SE Ranking (2026). AI traffic grew 16x from 2024 to 2026: top AI search engines driving visits to websites — the year-on-year growth figures: Gemini +231%, Claude +320%, Copilot +31%, and ChatGPT's share of all AI referral traffic falling from 79.74% to 75.96%.
• SE Ranking (May 2026). Referral traffic from ChatGPT hit an all-time high — the US AI-traffic series behind the Perplexity figures: 19.73% in the first four months of 2025, 11.42% across 2025, 6.85% in 2026, with global traffic flat.
• BrightEdge (2026). ChatGPT referral traffic more than doubles in 2026 — the dissenting dataset: referral traffic up 101% between January and August 2026, with ChatGPT's share climbing back to 95.1%.
• Octane11 (Q1 2026). State of B2B AI Search — an independent B2B panel: ChatGPT at roughly 72% in February 2026, down 14 points from its October 2025 peak of 86%.
• Peec AI, reported by Search Engine Land (2026). AI search engines cite Reddit, YouTube and LinkedIn most — an analysis of 30 million cited sources across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews. Reddit is the most-cited domain overall.
• Similarweb (2026). AI Search Stats: Market Share, Referral and Citation Data — platform-level traffic shares: ChatGPT falling from about 76% to roughly 53%, Gemini climbing to 27–28%, Claude approaching 9%.
• Reuters (May 2025). Apple's plan to offer AI search options on Safari a blow to Google — Eddy Cue's testimony that searches on Safari fell for the first time in 22 years, which he attributed to AI.
• Anthropic (2026). Web search tool — Claude API documentation — the inline-citation capability behind Claude's referral surge.