For most of 2025, "AI search visibility" effectively meant "ChatGPT visibility". A brand panel analysis covering 2,802,519 AI referral sessions across 41 sites found ChatGPT accounted for 89.1% of all AI referral traffic between May and August 2025. If you optimised for one assistant, you had covered the market.
By March-April 2026, that number was 62.6% — a fall of 26.5 percentage points in roughly seven months. ChatGPT did not shrink. The market around it grew.
Where the share went
- Claude: 1.4% to 18.5% (+17.2pp) — now the number two AI referral source outright.
- Gemini: 2.4% to 10.6% (+8.2pp), carried substantially by Google surface integration.
- Perplexity: 3.1% to 7.3% (+4.2pp), still small but consistently growing.
- Copilot: 3.2% to 4.0% (+0.8pp), essentially flat.
The strategic consequence is not subtle. A year ago, a single-assistant visibility programme covered nine out of ten AI referrals. Today it misses nearly four in ten. Any agency still selling "ChatGPT optimisation" as a complete product is selling a 62% solution at a 100% price.
Why fragmentation makes the work harder, not easier
Different assistants ground their answers in different places. Some lean heavily on live retrieval; others weight what they absorbed in training; others privilege a specific set of high-trust domains. A page that gets you cited in one may do nothing in another. That is why credible AI search agencies have moved from rank tracking to prompt testing: define a set of real buyer prompts, run them across several assistants on a schedule, and record whether you were named, in what position, how you were described, and — critically — which sources were cited.
That last column is the entire job. It tells you which pages, on which domains, are actually feeding the answer. Everything else is a proxy.
The engagement data complicates the "AI is killing traffic" story
The same analysis found AI-sourced sessions average roughly 58.5 seconds of engagement time at a 67.8% engagement rate, against 44.2 seconds and 61.8% for Google organic. Fewer visits, better visits. Someone arriving from an assistant has usually had the basic questions answered already and is further down the decision.
This changes what a visit is worth, and therefore what you should be willing to pay to earn one. Treating an AI referral as equivalent to an organic click understates it.
What to do this quarter
- Segment AI referral sources separately in analytics. Aggregating them hides exactly the shift described above.
- Run the same ten buyer prompts across ChatGPT, Claude, Gemini and Perplexity. Log who gets named and what gets cited.
- Prioritise the cited sources you can actually influence — directory entries, comparison pages, review profiles.
- Stop treating Claude and Gemini as rounding errors. Between them they are now roughly 29% of AI referrals.
To compare providers who do this work as their core service, see our AI search agency listings, or get matched with one directly. For the fundamentals, start with our guide to answer engine optimisation.
The share shift in one table
| Assistant | May-Aug 2025 | Mar-Apr 2026 | Change |
|---|---|---|---|
| ChatGPT | 89.1% | 62.6% | -26.5pp |
| Claude | 1.4% | 18.5% | +17.2pp |
| Gemini | 2.4% | 10.6% | +8.2pp |
| Perplexity | 3.1% | 7.3% | +4.2pp |
| Copilot | 3.2% | 4.0% | +0.8pp |
The baseline wave analysed 2,802,519 AI referral sessions across 41 brand sites, with the later figures drawn from a refreshed anonymised panel. Panel studies have real limitations — the site mix shapes the result, and a panel weighted toward developer tooling would flatter Claude, while one weighted toward consumer retail would flatter ChatGPT. Treat the exact decimals as indicative and the direction as solid, because the direction is corroborated everywhere.
Why the number two slot changed hands
Three forces compounded. First, assistant choice stopped being a novelty decision and became a workflow decision, and workflow decisions get made per task rather than once. Second, frontier pricing roughly halved during 2026, which pulled a large tier of cost-sensitive usage into models that had previously been evaluated and shelved. Third, and least discussed, integrations matured: once an assistant sits inside the tools people already use, referral volume follows the integration rather than the brand preference.
None of those forces is finished. Anyone building a visibility strategy on today's share table is making the same mistake as the people who built one on 2025's.
What a real prompt test looks like
The word "test" does a lot of work here, and most of what gets sold under that name is a screenshot. A defensible prompt test has five properties:
- Fixed prompts. Written once, in buyer language, never edited between runs. Changing the prompt to get a better result is the AI-search equivalent of moving the goalposts, and it destroys comparability.
- Multiple assistants, same day. Answers drift as indexes update. A test spread across a fortnight measures the calendar, not the strategy.
- Named, positioned, described. Record all three. Being mentioned ninth with a lukewarm description is a materially different outcome from being the first recommendation, and a binary mentioned/not-mentioned metric hides it.
- Cited sources captured every time. This is the column that generates the work. Everything else generates the report.
- A schedule. Monthly is enough for most categories. Weekly is theatre unless you are in a fast-moving consumer space.
Run that for three months and you have a trend line. Run it once and you have an anecdote — which is, to be fair, what most of the market is currently paying for.
What this does to pricing
Because AI-sourced sessions engage longer and convert further down the funnel, the value of an AI referral is higher than an equivalent organic click, and the volume is lower. That combination supports outcome pricing rather than volume pricing. An agency charging by traffic delivered is structurally mispriced for this channel; one charging for citation share in a defined prompt set is aligned with what the buyer actually gets.
Frequently asked questions about AI referral traffic
How do you actually track AI referral traffic? Most analytics platforms now surface assistant referrers, but the defaults often bucket them together. Split them by source so a shift between assistants is visible rather than averaged away — that separation is what turned a 26-point market change into something anyone could see.
Is AI referral traffic worth more than organic? On engagement, yes: about 58.5 seconds and a 67.8% engagement rate against 44.2 seconds and 61.8% for Google organic. The visitor has usually already had their basic questions answered, so they arrive further along.
Which assistant should you optimise for? The question is already outdated. A year ago one assistant covered nine in ten referrals; today it covers under two thirds. Build a measurement habit that survives the next reshuffle instead of picking a winner.
How often should you re-test? Monthly for most categories. The share table moved 26 points in seven months, so an annual review is not a review.
The bottom line
The single-assistant era lasted about eighteen months. Anyone who built a visibility programme around ChatGPT alone now has a measurable and growing blind spot, and the correct response is not to pick the next winner — it is to build a measurement habit that survives the next reshuffle. Read the primary write-ups from Similarweb and Anthropic for context on how fast this moved.