“AI Traffic Converts 4.4x Better” — Reading the Benchmarks Honestly

Search for how well AI referral traffic converts and you will find, within about ten minutes, credible-looking claims that it converts 1.3 times better than organic search, 4.4 times better, 8 times better, and 23 times better. All four are published. All four cite data. Several come from organisations with real measurement practices.

That spread is not noise. It is four different questions being answered and reported as though they were one. This post is an audit: what each family of study actually measured, why the answers diverge by more than an order of magnitude, and what number you should plan with.

The spread

Here is the range, plotted on a log axis because a linear one would compress the low end into a single stripe.

An eighteen-fold gap between the lowest and highest credible estimate of the same quantity should make you suspicious of the quantity, not of the researchers. Four things explain nearly all of it.

Driver one: the baseline is not the same baseline

This is the big one, and it is usually stated clearly in the study and dropped entirely from the headline.

Compare AI referral traffic to non-branded organic search and you are comparing two pools of people who did not previously know your company. That is close to a like-for-like comparison, and it produces the smallest multipliers — the Visibility Labs analysis of 94 ecommerce brands found ChatGPT referrals converting at 1.81% against 1.39% for non-branded organic, a 31% lift.

Compare AI referral to all traffic and the denominator now includes direct visits from existing customers, email clicks from your own list, and branded search from people typing your name. Those convert well and drag the baseline up — which should shrink the multiplier.

Compare AI referral to all non-branded acquisition traffic including paid and display, and the baseline collapses, because display traffic converts terribly. Now the multiplier explodes.

The 23x and 8x figures generally involve a broad, low-converting baseline. The 1.3x and 1.4x figures generally involve a tight, like-for-like one. Neither is dishonest. They answer different questions, and only one of them is the question you are usually asking.

Driver two: you cannot see the traffic you are measuring

This one is specific to AI referral and it is under-discussed.

To measure the conversion rate of AI traffic, a study must first identify which visits were AI-referred. Most rely on the referrer header or a campaign parameter. But the majority of AI referrals arrive with neither — the visit lands in Direct and never enters the study's AI cohort at all.

So these studies are not measuring AI traffic. They are measuring the observable subset of AI traffic. The question is whether that subset is representative, and there is good reason to think it is not.

Observable AI visits skew toward assistants with a web interface and toward desktop sessions. Referrer-less visits skew toward mobile app handoffs. Desktop research sessions and mobile app taps are not equivalent behaviours, and the desktop one is more purchase-adjacent.

If the visible subset over-represents higher-intent visits, then studies measuring only the visible subset overstate the channel average. That is a systematic bias with a known direction, and no amount of sample size fixes it.

Driver three: composition, not quality

Conversion rate is computed over sessions, and sessions are not interchangeable.

AI assistants recommend specific pages. Asked "what should I use for X", an assistant names products and links to product pages, pricing pages, and comparison content. It rarely links to your blog's tangential explainer, because that is not what answers the question.

Organic search delivers a much broader mix — informational queries landing on articles, navigational queries landing anywhere, long-tail queries landing on pages nobody planned.

So part of the measured lift is not "AI visitors are better." It is "AI visitors arrive on pages that convert better." Those are different findings with different implications: the first says invest in AI visibility, the second says AI happens to route people to your best pages, and if organic search routed people the same way it would convert similarly too.

Almost none of the published studies control for landing page. The few that segment by page type find the gap narrower.

Driver four: absolute rates that cannot all be channel effects

Some published per-assistant conversion rates are extremely high in absolute terms — figures like 15.9% for ChatGPT referrals and 10.5% for Perplexity appear in vendor analyses.

Treat absolute rates like these with particular care. A 15.9% session-to-conversion rate is extraordinary for any acquisition channel. It is achievable if "conversion" is defined loosely — a newsletter signup, a demo request, any tracked goal — or if the sample is dominated by a handful of sites with unusual funnels.

When you see an absolute rate that would be remarkable for any channel, the definition of conversion is usually doing more work than the channel is.

The one finding that survives everything

Strip out the magnitude arguments and something important remains.

The Adobe Digital Insights analysis found AI referral traffic converting 42% better than non-AI traffic in March 2026 — having converted 38% worse a year earlier. That reversal is a far more useful finding than any single multiplier, for two reasons.

It is a within-methodology comparison. The same measurement approach, the same baseline, two points in time. Whatever biases exist are present in both readings and largely cancel out. This is the one comparison in the entire literature you can trust for magnitude, because it only asks whether the number moved.

And it identifies a real transition. Early AI referral traffic was exploratory — people testing what the new toy could do, poking at recommendations without buying anything. It has become transactional as assistants became a normal way to make decisions. That is a genuine behavioural shift, and it explains why older studies read lower than newer ones independently of methodology.

How to read the next study you see

Five questions, in order of how much they change the answer:

1. What is the baseline? If it is not stated, the study is not usable for magnitude. If it is "all traffic", expect inflation.
2. How were AI visits identified? Referrer-only detection means a biased subset. Ask what fraction of AI traffic they believe they caught.
3. What counts as a conversion? Purchases, signups and newsletter subscriptions differ by an order of magnitude in rate.
4. Is landing page controlled for? Almost never. If not, part of the lift is page mix.
5. Who benefits from the finding? Not disqualifying — this is a young field and vendors have most of the data — but it should set the level of scrutiny.

A study that answers all five clearly is worth more than one with a bigger sample and a bigger number.

What we would actually claim

Being specific about our own position, since this post is published by a company that sells AI attribution:

We believe AI referral traffic converts better than non-branded organic search. Every independent dataset points the same way, the mechanism is obvious — a recommendation is a warmer introduction than a search result — and the effect has strengthened over time.

We do not believe anyone can tell you by how much. Not for your business, not from published benchmarks. The spread is too wide, the detection too incomplete, and the composition effects too large.

We think the second statement is the more useful one commercially. If the multiplier were reliably knowable from public sources, you would not need to measure it. The reason to instrument this channel is precisely that the benchmarks cannot answer the question for you.