AI Referral Traffic 2026: Gemini Is Catching ChatGPT

AI referral traffic in 2026 is consolidating around more than one assistant — and the platform pie is being redrawn faster than most GEO playbooks assume. According to SimilarWeb, Gemini’s share of generative-AI website traffic more than doubled over a year while ChatGPT’s slid into the mid-sixties, then lower still in the most recent reading. These are estimates from a single panel-based vendor, not audited industry totals.

That distinction matters, because the temptation is to treat a viral traffic-share chart as settled fact and re-plan an entire content strategy around it. The honest read is narrower and more useful: the direction of travel is clear and consistent across independent data sets, even if the exact decimals are not. Gemini is rising sharply, ChatGPT is still the largest single source but shrinking as a share, and at least two other assistants are growing off small bases.

This guide turns that picture into a plan. We lay out what SimilarWeb actually measured, present a proprietary 16-month trajectory table, explain the structural reasons behind Gemini’s climb, and then translate it into a multi-assistant GEO approach — because the strategic conclusion (don’t optimize for one assistant) survives even if you discount every specific number on the page.

The chart that circulated in June 2026 came from SimilarWeb’s official account and was reproduced verbatim by Search Engine Roundtable. It is a two-snapshot comparison of generative-AI website traffic share. The “12 months ago” reading (around June 2025) put ChatGPT at 76.4% and Gemini at 8.9%; the “6 months ago” reading (around December 2025) put ChatGPT at 65.2% and Gemini at 20.3%, with DeepSeek and Grok tied at 3.8%.

Two framing points keep that data honest. First, this is a single vendor’s panel-based estimate of traffic to generative-AI platform websites — not query volume, not intent, and not confirmed by any independent auditor. Second, the “6 months ago” column is December 2025, not the present; the most recent SimilarWeb reading (data through May 26, 2026, reported by ppc.land) shows the trend continuing, with ChatGPT at 52.7% and Gemini at 27.3%. We reference that latest snapshot as the newest data point, not as the viral tweet’s figures.

It also helps to keep the absolute scale in view. Per SE Ranking, all AI platforms combined accounted for roughly 0.24% of global internet traffic in January 2026, up from about 0.15% a year earlier — and traditional organic search still drove on the order of 345 times more traffic than ChatGPT, Gemini, and Perplexity combined. AI referral traffic is the fastest-growing channel on the web, but it is growing from a small base. The strategic urgency is about positioning early, not about abandoning organic search.

Most coverage shows two snapshots. The table below stitches together four SimilarWeb data drops — January 2025, June 2025, December 2025, and the May 2026 reading — into a single trajectory view, so you can read velocity and direction rather than a frozen moment. Every cell is SimilarWeb-attributed. Where the December 2025 snapshot did not break out a platform, the cell reads “n/a in source” rather than an interpolated guess.

The Δ column is computed against each platform’s earliest available SimilarWeb reading, which is why ChatGPT (−34.0 pp from January 2025) and Gemini (+21.6 pp) carry the longest baselines while DeepSeek, Grok, and Copilot are measured from June 2025. Two stories beyond the ChatGPT-versus-Gemini headline are worth pulling out. DeepSeek’s slide from 5.3% to roughly 4% is best read against its late-2024 launch spike rather than as a fresh decline. And Claude jumping to 8.9% in the May 2026 reading is striking — a newer, not-yet-widely-reported development that strengthens the case for planning across more than two assistants.

Read forward, the direction is more reliable than any endpoint. If Gemini continues to compound while ChatGPT’s share keeps easing, the gap narrows further through the back half of 2026 — but the May 2026 reading still has ChatGPT at roughly double Gemini’s share, so a crossover is a scenario to monitor, not a forecast to bank on. The planning implication does not depend on the crossover happening: once the second-largest source is large enough to matter, the cheapest mistake to avoid is building for only the first.

Two metrics get conflated in coverage of Gemini’s rise, and keeping them apart is the whole game. Share is Gemini’s slice of the gen-AI traffic pie — the 8.9% to ~20% climb above. Referral volume is the absolute number of clicks Gemini sends out to external websites. SimilarWeb’s widely cited 388% year-over-year figure describes the latter: Gemini referral visits, not share. Over the same comparison, ChatGPT referral visits grew about 52%. Both can rise together; only one describes the pie.

The volume surge tracks a clear product trigger. Per SE Ranking’s analysis, Gemini’s acceleration began in November 2025 alongside the Gemini 3 rollout — Gemini 3 Pro in mid-November, Gemini 3 Flash in mid-December — with reported month-over-month referral growth of roughly 51% in December 2025 and 42% in January 2026. By January 2026 Gemini had surpassed Perplexity in global referral traffic, reportedly sending about 29% more referral visitors worldwide and 41% more in the US, after Perplexity had sent several times more than Gemini as recently as the previous summer.

The structural driver behind the share climb is distribution. Google keeps embedding Gemini into surfaces people already use — for example, Gemini’s announced expansion into Chrome across the Arab World on June 12, 2026. When an assistant is the default in the browser, the search box, and the phone, its traffic does not have to be won one user at a time. That is the difference between a product people choose to visit and a product people are routed into. It is also why we treat Gemini’s rise as a trend with momentum rather than a spike likely to reverse.

The most consequential shift for practitioners is not the traffic split — it is that the rules for earning AI citations have come unstuck from classic SEO. Independent analyses report that the correlation between traditional organic rankings and AI citations dropped sharply in under a year, and that a large majority of AI citations now come from URLs that do not rank in the organic top results. We treat the exact percentages as directional, because they come from aggregated industry sources — but the pattern shows up across enough data sets to be real.

That breaks a comfortable assumption. For two decades the SEO answer to “how do we get more visibility” was “rank higher.” In an answer engine, ranking is one input among many, and not the decisive one. As one widely cited GEO framing put it, the shift from a link economy to an answer economy is structural rather than cyclical — which means waiting for it to revert is not a strategy. The work moves from climbing a results page to becoming the kind of source an assistant chooses to cite.

A June 2026 Search Engine Land study surfaced a finding that reframes the whole problem. Recognition by a model does not guarantee recommendation. In the study’s athleisure example, Nike appeared in 71% of AI recommendations while New Balance — sharing essentially identical Knowledge Graph descriptions — appeared in none. The difference was not on-site optimization. It was co-mention: how often a brand is named alongside category leaders in the external content models learn from.

The study’s co-mention clusters made the point starkly — pairings like lululemon and Alo Yoga showed hundreds of co-mentions, while other plausible pairings showed almost none. AI visibility, in other words, is built on semantic association in the broader web, not on tags and schema on your own pages alone. That is the missing half of most GEO advice, and it connects directly to the platform-diversity argument: different assistants weight that external association differently.

The actionable read is that earned mentions in third-party content — being named in the comparisons, round-ups, and expert lists your category already trusts — now does work that on-page SEO cannot. That is a fundamentally different muscle from keyword targeting, and it is closely related to AI brand visibility: the brands that get co-mentioned with leaders are the ones that get cited.

Here is the practical reason the traffic-share data demands a multi-assistant response: the platforms do not cite the same sources the same way. Optimizing for ChatGPT’s preferences does not automatically earn Gemini or Claude citations, because each model applies different filters. The table below maps the reported citation tendencies to the optimization lever each one rewards. Treat it as a directional guide drawn from industry analysis, not a fixed specification.

The implication is concrete. If Gemini strongly favors official brand websites and applies E-E-A-T as a primary filter, then a clean, authoritative owned-site presence is non-negotiable for Gemini visibility — which is the same thing distribution is rewarding as Gemini’s share climbs. If Claude leans on user-generated content at multiples of its peers, then reviews, forum presence, and social proof do citation work for Claude that a polished landing page cannot. One content asset rarely satisfies all four levers, which is exactly why a single optimization target underperforms across the field.

The reason to care about a channel that is still a fraction of a percent of total traffic is quality, not quantity. Per SimilarWeb, ChatGPT referral traffic converts at about 7.1% — second only to paid search at 7.8%, and ahead of direct, organic, social, email, and display. Adobe Digital Insights reported that AI referrals converted about 31% better than non-AI traffic in January 2026, with revenue per visit up sharply year-over-year. Independent analyses add that AI visitors spend meaningfully more time on site and bounce less.

Read those figures the same way you should read everything on this page — as directional vendor estimates, not audited universals. But the qualitative story is consistent and intuitive: someone arriving from an AI answer has usually been pre-qualified by the assistant, having described intent in natural language before the click. That is a higher-intent visitor than a broad organic searcher, which is why the conversion premium shows up across multiple independent data sets even when the exact percentage varies.

There is also a measurement caveat worth naming. ChatGPT only introduced clickable brand links in responses in May 2026, and SimilarWeb reported that the change drove a large week-over-week spike in referrals — meaning ChatGPT’s referral traffic was structurally suppressed before then. Some of the share movement between platforms reflects shifting product decisions about whether and how to send clicks out, not only shifting user demand. The takeaway compounds rather than weakens: optimize for the assistants, because the rules they each set move the numbers as much as user behavior does.

Pull the threads together and the strategy writes itself. Share is fragmenting across more than one assistant; each assistant cites differently; rankings no longer predict citations; and the visitors who do arrive convert above average. A ChatGPT-only approach optimizes for a shrinking — if still dominant — slice while ignoring the citation mechanics of every other platform. The decision below is how to allocate effort, and it does not require betting on any specific share forecast.

Sequencing matters as much as the list. Start with the owned-site authority work, because it is the only lever that pays off on both of the two largest platforms at once and because it is fully within your control. Layer earned co-mentions next, since they compound slowly and are the hardest to fake. Treat UGC and review cultivation as a parallel track for the assistants that weight it. And measure throughout — for the reasons in our breakdown of why most GEO advice misses the platform-diversity point, and against the broader backdrop of zero-click research showing search traffic is concentrating. If you want a partner to build and measure this across platforms, our agentic SEO engagements are built around exactly this multi-assistant model, and our content engine produces the citation-grade assets each platform rewards.