For most of 2025, “answer engine optimisation” was shorthand for “get cited by ChatGPT”. The data made that defensible: when one engine drives roughly nine out of ten measurable AI referrals, concentrating there is rational. That maths has now changed — and it changed quickly enough that plenty of AEO roadmaps written six months ago are quietly optimising for the wrong distribution.

The headline number, and who actually owns it

The figure circulating in AEO circles — ChatGPT’s share of AI referrals falling from ~89% to ~63% — comes from Goodie’s 2026 AI Search Traffic Report (published 21 May 2026). It is Wave 2 of a longitudinal study. We fetched and read the full report rather than the summaries, and the precise claims are:

• Wave 1 (Goodie’s October 2025 report): across 41 brand sites and 2,802,519 AI referral sessions from May to August 2025, ChatGPT held 89.1% of B2B AI referrals.
• Wave 2 (covering August 2025 through May 2026): using brand-averaged two-month averages for March–April 2026, ChatGPT’s share of measurable B2B AI referrals fell to 62.6%. Claude reached 18.5%, Gemini 10.6%, Perplexity 7.3% and Copilot 4.0%.
• The “Big 4” (ChatGPT + Claude + Gemini + Perplexity) hold roughly 99% of the cross-brand average — in Wave 1 those four engines already accounted for 96%, with ChatGPT making up 89 points of it.

Source: Goodie, 2026 AI Search Traffic Report. Shares are of measurable AI referrals in a mostly-B2B GA4 panel, brand-averaged.

Reading it like a sceptic before acting on it

We apply the same tests here that we set out in our guide to reading 2026 AEO benchmark reports: whose panel, which months, averaged how, and what doesn’t add up. On those tests, Goodie’s report is unusually transparent — and still has real limits you should carry into any strategy discussion.

What holds up

• The averaging method is stated. Goodie’s own words: “We first compute each brand’s share-of-AI for each source, then average across brands so each brand counts equally regardless of total session volume.” Brand-averaging means one high-traffic site can’t drag the number; it also means the figure describes the typical brand, not total traffic volume.
• The numbers that don’t sum are disclosed. The five headline figures sum to 103.0%, not 100%, because each engine’s cross-source average is computed independently. The report says so explicitly — and the “Big 4 hold nearly 99%” claim uses the cross-source sum (62.6 + 18.5 + 10.6 + 7.3). A vendor showing you the seams of its own maths is a good sign, not a bad one.
• It’s triangulated. Goodie checks its GA4 panel against Similarweb platform-visit data (25.77B visits, January–April 2026) and SensorTower iOS rankings, and the trajectories agree.

What to flag

• Wave 2’s panel size isn’t disclosed. Wave 1 named 41 brand sites; Wave 2 describes a “confidential anonymized panel of business-facing brands, mostly B2B” without a count. Treat the precision of “62.6%” accordingly.
• It’s a B2B lens. If your buyers are consumers, the mix will differ.
• Goodie sells AEO software. “Optimise for more engines” is a conclusion that suits an AEO platform. That doesn’t make it wrong — the independent data below points the same way — but it belongs in your weighing.
• GA4 referrer data undercounts AI traffic structurally. Goodie itself notes that a meaningful share of “direct” traffic in 2026 is AI-driven traffic that lost its referrer, and that Google doesn’t separately attribute AI Mode or AI Overviews referrals in GA4. “Measurable AI referrals” is the floor, not the total.

The second lens: the pie is growing, ChatGPT’s slice isn’t

Share can fall two ways: your traffic shrinks, or everyone else’s grows. Similarweb’s AI search statistics (published 29 July 2026, worldwide traffic panel) resolve which one this is — at the platform-visit level, not the referral level, which is an important scope difference:

• Generative AI platform visits grew about 70% year-over-year, reaching 9.5 billion across June 2025–May 2026, with unique visitors up 57% to 655 million.
• ChatGPT’s share of generative AI website visits fell from about 76% in June 2025 to roughly 53% by May 2026 — “even as its total visit count has stayed roughly flat. The drop reflects the category growing faster than ChatGPT specifically, not ChatGPT shrinking.”
• Gemini climbed to roughly 27–28% of visits; Claude moved from barely 2% to close to 9%, “the largest proportional gain of any platform tracked”.

Two independent datasets, two different methodologies, one conclusion: this is a growing pie with a rebalancing mix, not a ChatGPT collapse.

Why the mix matters: the engines are not interchangeable

If every engine cited the same sources the same way, diversification would be automatic. They don’t, on three counts.

Referral efficiency differs wildly. Goodie’s most strategically interesting number: Perplexity produced 7.3% of B2B referrals on just 1.9% of AI platform visits. A citation-first interface sends users out at a far higher rate than a chat-first one. Small engines can punch well above their visit share in actual traffic to you.

Retrieval differs. Gemini can ground answers in live Google Search results — Google documents this as “Grounding with Google Search”, connecting “the Gemini model to real-time web content” — so classic Google organic visibility carries over to Gemini more directly than to any other engine. ChatGPT’s browsing behaves more like an agent fetching a handful of pages per query, which rewards being the page that fully answers one question. Optimising for one does not automatically optimise for the other.

Volatility differs. Goodie’s unsmoothed data shows Claude hitting 27.2% of B2B referrals in April 2026 alone, up from 1.35% eight months earlier. As the report puts it: “The volatility is the point. Any AEO strategy that relies on one or two surfaces staying dominant is planning for a market that no longer exists.” The Big 4 of mid-2026 may not be the Big 4 of mid-2027.

This is the same reason our own outbound platform never bets on a single channel — Zian’s SmartReach AI™ orchestrates message, channel and timing per country, industry and profile precisely because contact channels shift under you. Discovery channels are now doing the same thing.

The diversification playbook

1. Measure per engine before you optimise per engine

You cannot rebalance what you can’t see. Segment ChatGPT, Gemini, Perplexity, Claude and Copilot referrals separately in your analytics — our guide to measuring AI referral traffic in GA4 covers the exact mechanics, so we won’t repeat them here. The one addition this data forces: report the engine mix, not just the AI total. A flat AI total can hide ChatGPT falling while Claude rises.

2. Track visibility per engine, not just traffic

Referrals are a lagging indicator; whether each engine cites you for your money prompts is the leading one. That’s share of model — and it has to be polled per engine, because a prompt you win on ChatGPT can be invisible on Gemini and vice versa. Our own polling shows exactly this divergence between engines on identical prompts.

3. Weight effort by referral efficiency, not visit share

Perplexity’s 7.3%-of-referrals-on-1.9%-of-visits ratio means visit-share headlines systematically understate citation-first engines. Rank engines by referrals delivered to your site per unit of your effort, and revisit quarterly — eight months was enough for Claude to go from #5 to #2.

4. Keep ChatGPT — drop ChatGPT-only

At 62.6% brand-averaged, ChatGPT is still the largest single source by a wide margin. The correction isn’t abandoning it; it’s ending the implicit assumption that ChatGPT visibility equals AI visibility. Entity consistency, clean structure, answerable pages and verifiable claims travel across all engines; per-engine tuning sits on top.

5. Report the undercount honestly

When you present AI-referral numbers internally, carry Goodie’s caveat with them: referrer-based counting misses AI-driven traffic that arrives as “direct”, and Google’s AI Mode and AI Overviews referrals aren’t separately attributed in GA4 at all. The trend is trustworthy; the absolute total is a floor.

Sources & ownership

Frequently asked questions

Did ChatGPT’s referral traffic actually decline?

Not in absolute terms, as far as the owner data shows. Goodie reports ChatGPT’s share of measurable B2B AI referrals falling from 89.1% to 62.6% brand-averaged, and Similarweb reports ChatGPT’s total visit count staying roughly flat while the category grew about 70% year-over-year. The share fell because rival engines grew faster, not because ChatGPT shrank.

Is AI referral traffic to websites still growing overall?

Yes, on every dataset we could verify at its owner. Similarweb’s AI search statistics report generative AI platform visits growing about 70% year-over-year to 9.5 billion between June 2025 and May 2026, with unique visitors up 57% to 655 million. Goodie’s panel likewise shows four engines now delivering meaningful B2B referrals where one dominated a year ago.

Should we stop optimising for ChatGPT?

No. At 62.6% brand-averaged, ChatGPT remains the largest single AI referral source in Goodie’s B2B panel by a wide margin. The change is dropping the assumption that ChatGPT visibility equals AI visibility: measure and optimise Claude, Gemini and Perplexity as separate surfaces alongside it.

Why don’t Goodie’s engine shares add up to 100%?

Because each engine’s cross-source average is computed independently across brands, the five headline figures sum to 103.0% rather than 100%. Goodie discloses this in its methodology — each brand’s own shares sum to 100%, but the cross-brand mean doesn’t preserve that constraint. It’s a property of brand-averaging, not an error.

Which engine should a B2B team add first after ChatGPT?

On Goodie’s March–April 2026 data, Claude is the #2 B2B referral source at 18.5%, ahead of Gemini at 10.6% — but Gemini’s grounding in Google Search means existing Google organic strength transfers there with the least extra work, and Perplexity converts a small visit share into an outsized referral share. Let your own per-engine GA4 segmentation decide; the mix varies by brand.