Short answer: whether AI-referred visitors convert better than organic search visitors depends almost entirely on what you sell. In B2B and SaaS, the premium is large and repeatedly measured. In e-commerce, the largest study to date found AI traffic converting worse than organic. But every one of these studies, including the ones with the least flattering conversion numbers, found the same behavioral signature: AI-referred visitors bounce less and go deeper into the site than search visitors do.
That consistency is worth more than the conversion multiplier, because it's the part you can build a page against.
The numbers everyone quotes, and where they came from
The 23x figure that circulates in almost every AEO deck comes from a single company measuring itself. In June 2025, Patrick Stox published Ahrefs' own analytics: AI assistants sent 0.5% of Ahrefs' traffic over 30 days and drove 12.1% of signups. Real data, honestly reported, one site. Stox says so himself and notes a larger multi-site study was still in progress.
The 4.4x figure comes from Semrush's June 2025 analysis of more than 500 high-value topics. Worth knowing before you repeat it: those were digital marketing and SEO topics, not a cross-industry sample. Most roundups describe it as "across industries." It isn't.
Seer Interactive reported roughly 9x around the same time, based on one client website.
Then the picture inverts. In October 2025, Search Engine Land covered research by Maximilian Kaiser and Christian Schulze analyzing twelve months of first-party GA4 data from 973 e-commerce sites with $20 billion in combined revenue: more than 50,000 ChatGPT-driven transactions measured against 164 million from other channels. ChatGPT referrals converted worse than every major channel except paid social, with organic search ahead by roughly 13% and revenue per session well behind direct, email, and paid search. The authors' own summary is blunt: results contradict widespread expectations of LLM superiority.
By early 2026 the e-commerce read had softened. Visibility Labs analyzed 94 seven- and eight-figure brands across all of 2025 and found ChatGPT sessions converting 31% better than non-branded organic, with average order value 14.3% lower and revenue per session slightly higher. Note the methodology difference that probably drives most of the gap: that study excluded homepage and blog traffic to isolate commercial-intent sessions. Amsive, meanwhile, found no significant difference at all.
So the honest range is somewhere between "converts 13% worse" and "converts 23x better," and the spread tracks business model and measurement choices more than it tracks anything about AI.
Why the disagreement isn't going to resolve
The mechanism people cite for the premium is pre-qualification: a model synthesizes several sources, compares options, and the user clicks a citation after the comparison rather than before it. That mechanism is real, and it explains why the effect is strongest in B2B SaaS, where the research phase is long and the purchase is considered.
It also explains why the effect weakens or reverses in e-commerce. Nobody asks Claude for help choosing a $15 phone case. Low-consideration purchases run on browsing and visual discovery, which is still Google's home turf. Kaiser and Schulze also flagged consumer trust as a likely factor, and their own time-trend analysis showed conversion improving month over month across the study window without reaching parity.
There's a measurement problem sitting underneath all of it too. E-commerce organic search conversion rates are inflated by absorbing what is really returning-customer intent, so the baseline AI traffic is being compared against isn't clean either.
If you are building a content strategy on the conversion multiplier, you are building on the least stable number in the entire dataset.
What every study agrees on
Strip out the conversion rates and the datasets converge.
Ahrefs found AI-referred visitors bouncing less than search visitors, viewing 50% more pages, and spending less time per visit. Kaiser and Schulze, whose whole finding was that ChatGPT traffic converts poorly, still recorded low bounce rates and high session depth, and read it as interest without purchase readiness. And in Similarweb's June 2026 clickstream study, AI-influenced visitors averaged 12.0 pages and 11.8 minutes per session against 6.5 pages and 5.6 minutes for everyone else.
Three studies, three different verdicts on revenue, one shared behavioral picture: these people arrive already oriented, move through the site rather than reading one page, and are looking for something specific.
The Ahrefs detail that gets least attention is the most useful one. Stox found that 80% of Ahrefs' AI-referred visitors landed on the homepage, free tools, or product pages, not the blog. The model already did the explaining. What the visitor wants is the thing itself.
The signal most teams are segmenting on is the wrong one
Here's the finding that should change how you build the rule layer.
Similarweb tracked US desktop panel users who asked ChatGPT an industry question, received a specific brand recommendation, and then visited a site within seven days. Recommended brands were 2.5x more likely to get that visit than their direct competitor. But 55.9% of that downstream traffic arrived through a branded Google search, not through a click on a citation inside ChatGPT.
Most of your AI-influenced visitors do not carry an AI referrer. They carry google.com.
Which means a personalization rule built on "referral source matches chatgpt.com" catches a minority of the segment it's aiming at. Referrer is still the cleanest signal you have when it's present, and it's worth using. It just can't be the only one.
Two caveats on that study, because they matter: Similarweb states plainly that the data shows correlation rather than causation, and the panel covers US desktop only across finance, travel, and beauty between July and December 2025. Rand Fishkin of SparkToro contributed analysis showing AI tools return different recommended brands across repeated runs of the same query, which complicates any plan that treats AI visibility as a stable asset.
Behavioral signals that survive the referrer problem
If referrer catches half the segment, the other half has to be inferred from what people do on the page. Three signals do most of the work.
Session depth relative to your site average. A visitor hitting five pages in one session on a site averaging two is either evaluating you seriously or lost. Combine it with which pages: five blog posts reads as research, while homepage plus pricing plus a comparison page reads as a shortlist check. That second pattern is the AI-influenced signature described across all three studies above.
Entry page type. If someone's first page is a product, pricing, or tool page rather than a blog post, they skipped the education phase. Serving them a "new to this topic?" intro is friction. This is the operational version of the Ahrefs 80% finding, and you can implement it with the landing-page dimension you already have.
Return-visit count. The clearest intent proxy available without identity resolution, and it needs nothing but a first-party cookie. First visit, orient. Second, deepen. Third and beyond, convert.
Scroll depth is worth capturing but is weaker than it looks as a trigger. On a page someone is scanning for one specific fact, a shallow scroll can mean they found it in eight seconds. Treat it as a diagnostic for page structure rather than a personalization rule.
Building the rule layer without a CDP
The architecture is signal, then inference, then one content swap. No model required, and a CDP is not the missing piece for most teams.
Start with instrumentation you probably half have. GA4 records the events but doesn't hand values back to the page at render time, so you need something client-side that can read session count, entry page, and referrer and expose them to your CMS or tag manager. That can be twenty lines of JavaScript writing to a first-party cookie.
Then define segments narrowly. "Any return visitor" is not a segment; it lumps a ten-second bounce in with someone who read 1,200 words. Something like "second or third session, entered on a product or pricing page, no conversion event" is specific enough to reason about and rare enough to be meaningful.
Map each segment to exactly one change. The call to action is almost always the right first target, because it's the element where a wrong assumption about visitor state costs you the most and the change is trivially reversible. Reordering sections is possible and considerably harder to maintain.
Then test against a control before you scale, and be honest about whether your segment has the volume to reach significance. Most don't, quickly. A segment producing forty sessions a week will take months to tell you anything, and running it anyway while acting on week-two results is worse than not testing.
Every signal described here is first-party and session-scoped, which is why this approach outlives third-party cookies rather than needing to be rebuilt around their absence.
Where this falls apart
Segments that are too broad produce rules that fire on noise. Segments narrow enough to be meaningful often lack the traffic to validate. That tension is real and there's no clever way around it; you pick the side you can live with and document the choice.
The bigger failure is changing too much at once. If four modules shift based on behavior, you can't attribute a result to any of them and you can't debug the experience when someone reports it looking wrong. One swap per segment.
And the honest limit on the whole approach: you're inferring intent from proxies. A person researching for a competitor looks identical to a buyer. Someone on a shared device looks like a returning visitor who changed their mind. These rules should be explainable to anyone on the team in one sentence, and if a rule needs a paragraph to justify, it's probably too clever.
FAQ
Does AI search traffic convert better than organic? In B2B and SaaS, the measured premium ranges from roughly 4x to 23x, though the largest figures come from single-site analyses. In e-commerce, a 973-site study covering $20 billion in revenue found ChatGPT converting about 13% worse than organic search, while a later 94-brand study that excluded blog and homepage traffic found it converting 31% better. The direction of the effect depends on your business model and on what the study chose to measure.
Why do the studies disagree so much? Sample composition mostly. Studies showing large premiums skew toward B2B software companies with long research cycles. The study showing a deficit looked exclusively at e-commerce. Methodology matters too: excluding blog and homepage sessions raises AI's apparent conversion rate considerably, because that's where a lot of low-intent AI traffic lands.
Can I segment AI traffic using referral source alone? Not completely. Similarweb found 55.9% of downstream traffic from a ChatGPT brand recommendation arriving via branded search rather than an AI referrer. Referrer-based segmentation is accurate for what it catches and blind to roughly half the segment. Behavioral signals fill the gap.
Do I need a CDP for behavioral personalization? Not for rule-based content swaps. Session count, entry page type, and session depth are available from a first-party cookie and your existing analytics. A CDP adds cross-device identity resolution and real-time profile stitching, which matter for orchestrated multi-channel campaigns and are overkill for swapping a call to action.
What's the minimum viable version of this? One rule: if entry page is a product or pricing page and session count is one, replace the top-of-funnel call to action with a direct next step. Run it against a control. That single change tests the entire premise before you invest in the infrastructure.
Should I still invest in AI visibility if the conversion premium is contested? The volume case is independent of the conversion case. AI referrals were 0.5% of traffic in the Ahrefs data and under 0.2% in the e-commerce study, so nobody is arguing this replaces organic search today. The argument for investing is that citation share compounds and branded search lift shows up whether or not the referrer does.
RankSage is being built to join GA4, Search Console, first-party behavioral data, and AI citation data at the page level, so you can see which pages earn citations and how the visitors who arrive from them actually behave. We're pre-launch, which means we have no performance data of our own to show you yet, and we'd rather say that than borrow someone else's numbers. Join the waitlist for early access.