Two credible datasets say opposite things about AI search traffic, and both of them are right.
A 973-site ecommerce study found ChatGPT referrals converting worse than organic search. Adobe's retail data found AI referrals converting 54% better. The tiebreaker is not methodology — it is the calendar: the ecommerce window closed in July 2025, and AI traffic quality inverted after that. Which means any "AI traffic converts at X%" figure older than roughly nine months is describing a different internet, and you should not plan a quarter around it.
We run the analytics stack for clients whose AI referral traffic went from a rounding error to a line item in about eighteen months. Here is what the conflicting research actually says, why it conflicts, and what we do with the traffic once it lands.
What the two headline studies actually found
The pessimistic one is a working paper titled ChatGPT Referrals to E-Commerce Websites: Do LLMs Outperform Traditional Channels?, covered by Search Engine Land. Its scale is the reason it gets quoted:
- 973 ecommerce websites, with roughly $20 billion in combined revenue.
- Twelve months of first-party GA data, August 2024 through July 2025.
- More than 50,000 ChatGPT-driven transactions compared against 164 million from every other channel.
Its findings were blunt. ChatGPT referrals were about 0.2% of sessions, roughly 200 times smaller than Google organic. Organic search converted about 13% better than ChatGPT, affiliate converted 86% better, and the only channel ChatGPT beat was paid social. On revenue per session it trailed both paid and organic search.
The optimistic one is Adobe Analytics' ongoing AI traffic research, reported by Digital Commerce 360. Different picture entirely:
- AI-referred traffic to retail sites grew 138% year over year in May 2026, and 1,324% since October 2024.
- AI traffic converted 54% better than traffic from non-AI sources.
- AI-referred visitors spent 53% more time on site and viewed 23% more pages per visit.
- Consumers were 15% more engaged once they arrived from an AI referral.
Same broad question, same broad industry, opposite answers.
Why they don't actually contradict each other
Read the Adobe figure that nobody quotes. Alongside that +54% conversion advantage, Adobe noted what the same metric looked like a year earlier: conversion rates from AI sources were then nearly half those from non-AI sources.
A year before May 2026 is mid-2025 — squarely inside the ecommerce study's August 2024 to July 2025 window.
So the two datasets are not disagreeing. They are describing the same curve at two different points, and the curve crossed zero somewhere in late 2025. The ecommerce paper measured AI traffic while it was still worse. Adobe measured it after it got better. Both are accurate reports of their own window, and neither is a description of what your traffic will do next quarter.
This is the practical takeaway: AI referral quality is a moving target, and the direction of travel has been consistently upward. A benchmark from a study window that closed fourteen months ago is a historical artefact.
The 15.9% number is a case study, not a benchmark
The other figure in circulation comes from a Seer Interactive case study: ChatGPT referral traffic converting at 15.9% against 1.76% for Google organic on the same site. It gets repeated as though it were an industry average.
It is one site. The gap is real and the direction matches everything else in the recent data, but a 15.9% conversion rate is a lead-gen form fill on a site with a small AI session count, not an ecommerce checkout across 973 properties. Quoting it next to the ecommerce paper is comparing two different conversion events on two different scales.
We've had clients bring us that number and ask why theirs looks nothing like it. The answer is usually that they're measuring a purchase and the case study measured a form.
The denominator problem nobody warns you about
Here is where most agencies get this wrong in a client report.
AI referrals are still small. In the properties we manage, AI Assistant sessions typically sit in the low single-digit percentage of total sessions for service businesses — a finding consistent with the 0.2% the ecommerce paper measured a year earlier, allowing for growth since.
Small denominators produce unstable rates. If you get 60 AI sessions in a month and 5 of them convert, you have an 8.3% conversion rate and a headline. Two conversions fewer and you have 5%. Nothing changed about your site. Reporting that swing as a performance change is how a channel gets budget it hasn't earned, or gets killed before it's had a chance.
Set a floor before you read the rate. We don't treat an AI channel conversion rate as a signal until there are a few hundred sessions behind it, or a full quarter of data. Until then it goes in the report as a volume trend with the rate flagged as provisional.
There's a second measurement trap on top of that one. A real share of AI-driven visits never gets attributed to AI at all, because some assistants and AI browsers strip the referrer and the session lands in Direct. We wrote up the mechanics of that in what AI browsers are doing to ad attribution, and the tracking setup we run instead of relying on GA4's native channel in the GA4 AI Assistant channel and what it misses. If you haven't built the custom channel group, your AI number is low and your Direct number is carrying the difference.
Why AI traffic converts well when it does
The mechanism is not mysterious, and understanding it tells you what to fix.
Someone who lands on your page from an assistant has already done the research step. They asked a question, read a comparison, got a recommendation, and clicked. A cold organic visitor arrives in the middle of that process; an AI visitor arrives at the end of it. Adobe's engagement numbers — more time on site, more pages per visit — are what that looks like in the data.
That produces a specific, fixable failure mode: the assistant makes a claim about you, and your page doesn't confirm it.
An assistant tells someone you do same-day service in their city. They click. Your homepage leads with a decade of experience and an award badge, and the service-area detail is three clicks deep. The visitor arrived pre-sold on one thing and got a page arguing something else. That's not a design problem or a speed problem. It's a context mismatch, and it converts badly no matter how good the page is on its own terms.
What we change on the page
Once a client's AI channel has enough volume to read, the work is narrow and mostly copy:
- Pull the AI channel's top landing pages and read them as the assistant's reader would. For each page, write down the question that would have caused an assistant to send someone there. If the page doesn't answer that question above the fold, that's the edit.
- Put the specific claim where the eye lands. Service area, turnaround, pricing model, integration support — whatever the assistant is likely asserting about you. Vague positioning survives cold traffic and loses pre-qualified traffic.
- Shorten the path for people who already decided. Pre-qualified visitors don't need the full nurture ladder. A visible, immediate booking path near the top matters more for this segment than for any other.
- Make the page quotable in the first place. The same structure that gets you cited is the structure that confirms the claim when someone arrives. We covered the citation side in why your content gets crawled but never cited.
- Fix the fundamentals underneath all of it. Pre-qualified traffic still bounces off a slow page or a nine-field form — the same items in our seven high-impact CRO changes apply here, they just cost you more per visitor because these visitors were worth more.
So what should you actually plan around?
Three things, in order.
Stop importing benchmarks. The published numbers span from "worse than paid social" to "converts at 15.9%", and the spread is explained by dates and conversion definitions, not by one study being wrong. Your own AI-versus-non-branded-organic comparison, same event and same window, is the only number that should inform a decision.
Treat AI traffic as a quality segment, not a volume channel. At a low single-digit share of sessions, it will not replace search this year. But it is the segment arriving latest in the decision, and the marginal return on optimising for it is high because the visitor is already most of the way there.
Expect the number to keep moving. Every dataset from the past two years points the same direction, and there is no sign of the curve flattening. Build the measurement now so you have a trend line when the volume arrives, rather than a start date.
The clients who are getting real revenue out of this didn't do anything exotic. They built honest tracking, waited for a denominator worth reading, and fixed the gap between what an assistant says about them and what their page says about them.
Want to know what your AI traffic is actually doing? We'll audit your tracking setup, show you where AI-driven sessions are hiding in Direct, and flag the landing pages that break the handoff. Get a free automation audit — no pitch deck, just the findings.