One widely-shared 2026 benchmark puts Claude at the top of the table: best-converting AI source there is, close to 17%. Another puts it near the bottom, around 5%. Both are real, published, cited numbers. Both can be right at once, and here's the uncomfortable reason: on almost every site, Claude sends about thirty sessions a month, and you cannot build a conversion rate out of thirty sessions. The "best-converting AI source" is the one you can't actually measure. That's the whole problem with the question in one line, so we'll start there and work outward.
If you searched "which AI traffic converts best," you want a ranking. You'll get one below, the honest version. But the ranking matters less than the thing nobody selling you a benchmark will admit: for most stores, only one of these four sources has enough traffic to trust a number at all.
The honest ranking, and why every list disagrees with the next one
Split the question in two, because the AI sources behave completely differently on each half.
Volume: ChatGPT owns it, and it's not close. Trakkr's index, updated mid-July 2026 across 1,429 anonymized GA4 properties, has ChatGPT at 91% of AI referral sessions, Gemini at 4.5%, Claude at 1.9%, and Perplexity at 1.4%. So roughly nine in ten AI visits are ChatGPT, and the other three combined are a rounding error on a rounding error. And the share itself won't sit still. Trakkr's 91% sits next to another mid-2026 index at 77% and a third that has ChatGPT down at 53% with Claude tripling its slice. Same platform, same year, a thirty-point spread. When the market share of a channel is that unsettled depending on who's counting and which slice they count, that's your first warning about the conversion numbers stacked on top of it.
Conversion rate: this is where the lists start shouting at each other. First Page Sage studied 150-plus companies across 32 industries from May 2025 through April 2026 and concluded that ChatGPT's and Perplexity's rates run higher on average than Gemini's and Claude's — except Claude tops the table in Higher Education (5.9%) and Healthcare (5.4%), while ChatGPT wins Hotels (7%) and Legal (5.6%). Other roundups flip the order entirely: Perplexity beating ChatGPT on intent, or Claude leading everyone.
They're not wrong so much as measuring different things. A 5.9% "conversion" in higher education is a lead or an application, on a considered B2B-ish decision. That is not the same event as an ecommerce purchase, and you should never lay the two side by side. When your own AI channel is a store checkout, the honest purchase number is low single digits — in one twelve-month look at 94 ecommerce brands, ChatGPT traffic converted around 1.8% against 1.4% for non-branded organic. A real edge. Nowhere near 17%. Anyone quoting you a double-digit "AI conversion rate" for a store is either measuring a softer goal or riding a sample too small to hold weight.
Strip the noise and the defensible synthesis is boring, which is how you know it's true. ChatGPT is the volume, and enough of it to trust the number you compute. Perplexity and Claude arrive with high intent, since the visitor did their comparing inside the chat, but on a base so thin it barely moves revenue no matter how well it converts. Gemini is the odd one: more reach than Perplexity or Claude, weaker conversion, partly because it bleeds into Google's own AI surfaces and drags general-search intent along with it. That's the ranking. Now here's why you shouldn't take mine either.
Why the benchmarks disagree by 10x: it's arithmetic, not sloppiness
The gap between "Claude converts at 17%" and "Claude converts at 5%" isn't one study being careless. It's what happens when you compute a rate on almost no data.
Take Claude at 1.9% of AI traffic. Say your store does 150,000 sessions a month, healthy. AI is maybe 1% of that on a good day, so about 1,500 AI sessions, of which Claude's share is roughly thirty. Thirty sessions. One order turns that "conversion rate" from 0% into 3%; three orders and it reads 10%. Publish it to one decimal place in a comparison table and it looks like measurement. It's a coin flip wearing a lab coat.
This is the thing the listicles can't do and won't say: a per-source conversion rate needs a floor under it before it means anything, and for most sites three of the four AI sources never clear it. Set the floor at a few hundred sessions in the window you're reading. ChatGPT usually clears it. Perplexity, Gemini, and Claude clear it only at real scale, or if you widen the window to a quarter instead of a month. Below the floor, the number isn't small-and-reliable, it's just noise with a decimal point.
So the real question isn't "which AI source converts best" in the abstract — that's a question about someone else's sample. It's two questions about yours: which of my AI sources has enough volume to even rank, and of those, which one earns the intent it arrives with.
What this looks like on a real store
Here's a mid-size store's AI channel over 30 days, broken out by source. Illustrative figures, not benchmarks — the point is how you read the shape, not the specific numbers.
| AI source | AI sessions, 30d | Purchase CVR (illustrative) | Rev/session | Read it as |
|---|---|---|---|---|
| ChatGPT | 1,240 | 2.1% | $1.90 | Rankable. Converts fine, above site organic. |
| Gemini | 240 | 0.9% | $0.70 | Rankable, and the real problem. Weak. |
| Perplexity | 180 | 3.3% | $3.10 | Promising, but 6 orders. Can't trust it yet. |
| Claude | 30 | one order | — | Noise. Watch it, don't optimize it. |
A benchmark listicle, handed this store, would have said "your best AI sources are Perplexity and Claude — highest conversion, go chase them." The store's own data says something almost opposite. Only ChatGPT is genuinely measurable, and it's doing its job. Gemini is real, countable volume converting at less than half the ChatGPT rate, and that's the actual thing to open up this week, because it's costing money quietly and at a size you can prove. Perplexity looks great and might be great, but 3.3% is six orders; one refund and it's 2.8%, so it goes on the watch list, not the action list. Claude is thirty sessions and a single lucky order. Nothing to do but let it grow.
Two completely different to-do lists off the same channel, depending on whether you read your own numbers or somebody's benchmark table. That's the entire case for doing this yourself.
How to run it on your own GA4
You need per-source visibility before any of this works. If your AI traffic is still lumped into one channel — or worse, hiding in Direct — start with how to track AI traffic in GA4, which covers the native AI Assistant channel and the custom channel group that splits ChatGPT, Perplexity, Gemini, and Claude apart by referrer host. Come back once you can see them as separate sources.
Then the cut is straightforward. In Reports → Acquisition → Traffic acquisition, switch to your AI channel, break it down by Source, and add session conversion rate and revenue per session. Now apply the floor in your head: cross out any source under a few hundred sessions before you read its rate at all. What's left is your real comparison. For anything the standard report won't segment cleanly, an Exploration with Source as the row and the ecommerce metrics as columns does the same job with more control.
What you do next depends on which bucket each surviving source lands in. A source that clears the floor and converts well: leave the traffic alone and make sure the pages it lands on don't waste the intent; that's the landing-page work, since AI sends people to deep product and comparison pages built to rank, not to greet a mid-decision buyer. A source that clears the floor and converts badly, like the Gemini row above, is the same landing-page diagnosis pointed at a leak you can actually prove. And a source below the floor gets nothing but a calendar reminder to look again next quarter, when it might have grown into a channel with a number worth reading.
One caution that runs in your favor, so you don't undersell the whole thing: GA4's last-click model probably undercounts AI, because plenty of assistant-prompted buyers go search your brand on Google before they check out, handing the credit to branded organic. Your AI numbers are closer to a floor than a ceiling. Hold both errors at once: the vendor benchmarks overstate on tiny self-selected samples, last-click understates on yours, and the truth sits in the uncomfortable middle where you have to measure to know.
The line the numbers won't cross
Two things this comparison can't do for you, and both matter.
First, rate isn't revenue. Volume times conversion is what shows up in the bank, and a source can win the rate and lose the money by a mile — Perplexity at 3.3% on 180 sessions makes you six orders; ChatGPT at 2.1% on 1,240 makes you twenty-six. Rank your sources by sessions times conversion, not by the prettiest percentage, or you'll spend the week chasing a channel that can't move your total even if you double it.
Second, this measures outcome, not visibility. GA4 can tell you whether Perplexity's traffic buys once it arrives. It can't tell you whether Perplexity recommends you in the first place — that's a generative-engine-optimization question, share of model and citation rate, answered by tools that query the models directly, not by your analytics. Keep the two apart. And keep this piece apart from a different one people conflate with it: this is about which source sends the better buyers, not which assistant you'd wire up to do the analysis. For that, Claude vs ChatGPT for analytics is the comparison you want, and the answer there is mostly "use both."
The shortcut: ask instead of build
Everything above is an Exploration you assemble, apply a floor to by hand, and then rebuild next month. Most teams do it once and never reopen it.
The faster version is to wire GA4 to Claude or ChatGPT over MCP and just ask. That's what we built ConvRadar for (and yes, this site is ours): a hosted GA4 MCP server you set up in the browser, no terminal or Python, that lets the assistant run the exact comparison above. "Compare conversion rate and revenue per session across my AI sources — ChatGPT, Perplexity, Gemini, Claude — for the last 30 days, and flag any source with too few sessions to trust." The segment-comparison and traffic tools do the per-source split and the small-sample flag in one answer instead of a report you rebuild monthly, and the prompt library has the rest. Setup runs about five minutes in Claude or ChatGPT, and it's free during the open beta — email signup, no card.
Where it stops is worth saying plainly. It reads GA4, so it knows exactly what GA4 knows: it can rank your AI sources by conversion and tell you which ones are too thin to rank. It can't show you the page through the buyer's eyes, it can't read the wording of the AI answer that set their expectation, and it can't tell you whether the assistant recommends you at all. That last one was never GA4's job.
FAQ
Which AI traffic converts best? For the traffic you can actually measure, ChatGPT — because it's the only AI source most sites get enough of to trust a conversion rate on, and it converts a touch above non-branded organic. Perplexity and Claude often show higher rates in public benchmarks, but on samples so small that for any individual site the number swings on a single order. Gemini sends more volume than those two and converts worse. The honest answer is "measure your own by source," because the ranking that pays your bills is the one built from your data, not a benchmark's.
Does Perplexity traffic convert better than ChatGPT? Sometimes, on paper. Perplexity visitors tend to arrive further along — they did their comparing inside the answer, and Perplexity's inline citations send a click that's already qualified. But Perplexity is around 1.4% of AI referral volume, so on most sites its "conversion rate" is built on a few dozen sessions and isn't stable enough to bank on. ChatGPT converts slightly lower per session but on roughly 60x the volume, which makes its number both trustworthy and the one that actually moves revenue.
Why do AI conversion benchmarks disagree so much? Because they measure different events on different samples. A "conversion" in one report is a B2B lead or a signup, in another it's an ecommerce purchase — and those differ by 5-10x before you compare a single platform. Layer on tiny per-platform samples (Claude and Perplexity each send 1-2% of AI traffic), different industries, and different months, and you get "Claude converts at 17%" in one table and "5%" in the next. Both can be technically true. Neither tells you what will happen on your site.
Is Claude traffic worth tracking? Track it, don't optimize it — for now. Claude sends around 1.9% of AI referral traffic, which for most sites is a few dozen sessions a month: too few to compute a conversion rate you'd trust, too few to justify building anything around. The move is to keep it in your custom AI channel group so it's counted, watch whether it grows, and revisit when it clears a few hundred sessions in a window. High intent on almost no volume is a nice footnote, not a channel yet.
How much AI traffic do I need before the conversion rate is trustworthy? Set a floor of a few hundred sessions in the window you're reading before you believe a per-source rate. Under that, one order or one refund moves the number by whole percentage points, so you're reading noise. If a source can't clear the floor in a month, widen the window to a quarter, or accept that you can only judge it at the blended-AI level, not source by source.
How do I compare ChatGPT vs Perplexity vs Gemini traffic in GA4? Split your AI channel by Source (set up the custom channel group first if AI traffic is still lumped together), then add session conversion rate and revenue per session to the Traffic acquisition table, and ignore any source under the sample floor. For a cleaner cut, build an Exploration with Source as the row and the ecommerce funnel as the steps. Or connect GA4 to an assistant over MCP and ask for the per-source comparison with a small-sample flag in one prompt.
Does more AI traffic mean more revenue? Not on its own — it's volume times conversion that lands in the bank, and the two often point different ways. A high-intent source with tiny volume (Perplexity, Claude) can convert beautifully and still contribute almost nothing, while ChatGPT's larger, slightly-lower-converting stream does the real work. Rank your AI sources by sessions times conversion rate, not by the highest percentage, and spend your effort where the money actually is.
Rank your own, not the listicle's
The comparison tables rank four AI sources for a store that gets meaningful traffic from one. Yours might be different — maybe Gemini is your volume, maybe Perplexity actually clears the floor. You won't know from anyone's benchmark. Connect GA4 to your assistant, break the AI channel down by source, and ask: "Which of my AI sources converts best, and which have too little traffic to trust yet?" The three that don't clear the floor, you're watching. The day one of them does, it's a real channel with a real number — and you'll be the one who saw it coming.