In July 2025, Semrush published a blog post called “We Studied the Impact of AI Search on SEO Traffic.” The sentence everybody’s quoting reads: “We have seen that the average AI search visitor is 4.4 times as valuable as the average visit from traditional organic search, based on conversion rate.”

We have seen. That’s the methodology. That’s all of it.

No sample size, no site count, no window, nothing you could replicate. The “500+ topics” figure that always rides along next to it describes a traffic projection model, not the conversion measurement.

The post does carry a Notes section, but read what it covers. It disclaims the 2028 crossover and the 2027 value-parity estimate as extrapolations. It says nothing at all about the 4.4x figure.

Semrush didn’t do anything especially wrong here. They put a number in a blog post, described it as something they’d seen, and went back to shipping product. What happened next wasn’t their fault. Sort of like how that bat bodega in Wuhan didn’t mean to touch off anything, either.

Stat laundering

A number shows up on a vendor blog with a soft qualifier attached. A second blog cites the first and drops the qualifier. A third cites the second and calls it “research.” By the fifth pass it’s got a year on it, a sample size nobody can trace, and a permanent home in somebody’s board deck under Industry Benchmarks.

Call it stat laundering. Goes in dirty, comes out looking peer-reviewed.

The 23x figure went through the exact same wash cycle. It’s from an Ahrefs post in June 2025 reporting that AI search drove 0.5% of their traffic and 12.1% of their signups over thirty days. One company. One month. Their own product, their own analytics tool, and an audience of SEO professionals who are quite possibly the most likely humans alive to be asking ChatGPT about SEO tools. Patrick Stox wrote it up straight and hedged it in the text.

Everybody downstream kept the multiplier and threw out the hedge.

Both numbers might be pointing in the right direction. Neither one’s a benchmark. And if you put either in front of a CMO who reads primary sources, you’ll spend the rest of that meeting defending your homework instead of selling anything.

What the peer-reviewed evidence actually says

There’s exactly one peer-reviewed study on this we could find, published in April 2026 in Marketing Science, which is about as unimpeachable as marketing research channels get.

Maximilian Kaiser and Christian Schulze went through twelve months of first-party Google Analytics data from 973 e-commerce websites with $20 billion in combined revenue. Over 50,000 ChatGPT-referral transactions against 164 million transactions from traditional channels, with website, device, and month fixed effects and a robustness check for everything you’d think to ask about.

Their headline finding is rough news for the AEO industrial complex. Across the full sample, organic search beat LLM referral traffic on conversion likelihood by roughly 13%. Affiliate beat it by 86%. The only channel LLM traffic outperformed was paid social.

So much for 4.4x.

But! There’s the part that nobody quoting that headline might have read. To wit…

Complexity is the whole ballgame

Kaiser and Schulze split the sample by product category complexity. Sites in high-complexity categories showed 4.6 times higher LLM traffic share.

And in those categories, the whole thing flips. LLM conversion rate beat paid social, referral, email, direct, and organic search.

Their high-complexity list? Heavy industry and engineering, business and consumer services, vehicles, law and government, finance, careers.

That’s not an accidental overlap with B2B software. That’s what happens when a purchase means synthesizing specs, comparing five vendors, and then defending the pick to a buying committee that wasn’t in any of the demos. The model does the synthesis now. Your buyer shows up having already finished the part of the funnel your blog used to handle.

The authors’ own summary line, which I’d frame above my desk if I were the framing type:

In high-complexity categories, LLM referrals combine strong conversion rate and revenue per session with favorable bounce rate, and can work as a high-intent complement to organic and paid search even at small volumes.

They also point out that their attribution’s last-click, which structurally understates what LLM traffic contributed. The channel’s doing better than the numbers say, not worse.

Where it goes the other way

If the mechanism is “the model already did the comparison work,” the advantage should vanish anywhere there wasn’t much comparison work to do. It does.

SALT.agency ran it across 671,694 LLM referral sessions against 188 million organic sessions in forty website categories in early 2025. Organic converted better in most of them. Consumer ecommerce was the ugliest gap: 17.56% from LLM referrals versus 24.12% from organic. Human resources came in at 2.17% against 8.99%. Pharmaceutical, 4.06% against 8.29%. Education, 1.89% against 3.97%. Automotive, 0.24% against 0.41%. Travel converted fine from AI at 24.25% and still lost to organic at 28.97%.

Nobody’s asking Claude to help them pick out a phone case. Low-consideration buying happens by scrolling and looking at pictures, and that’s Google’s turf.

Now the part where I undercut my own argument, because the alternative is being wrong in public later and I’ve done that enough.

SALT’s B2B Services row came in at 2.77% for LLM traffic against 3.01% for organic. SaaS was a dead heat, 6.69% to 6.71%. And Amsive’s study of 54 websites, the only one I’ve found that ran significance testing at all, produced a paired t-test p-value of 0.794. That’s the statistical version of a shrug.

Ahrefs’ bigger follow-up across 82,000 sites found AI visitors looked at fewer pages and bounced more than search visitors.

The trend line beats any single number

Adobe’s been tracking AI-referred traffic to U.S. retail sites since October 2024, and their series explains why half these studies contradict the other half. They’re measuring different years.

• July 2024: AI traffic converted 43% worse than non-AI.
• February 2025: gap down to 9%.
• Holiday 2025: AI referrals crossed over and converted 31% better.
• March 2026: 42% better, with 13% more pages per visit and 48% longer on site.

Engagement moved first. Conversion caught up about a year later, once people got better at using the tools and started trusting what came back. Adobe’s survey found 83% of AI shoppers say they’re more likely to use it for larger or more complex purchases, which is just the buyer-stated version of the same complexity finding Kaiser and Schulze pulled out of the transaction data.

Any study run in 2024 concluded AI traffic was garbage. Hey, I was there, seeing those numbers. We went ahead and optimized clients like Evisort anyway, because why not? Imagine our surprise when we actually got traction from it.

Any study run in 2026 concludes it’s the best-converting channel on the site. They’re both right about their observation window, but neither one is a law of nature. Shit happens.

What to do with this

Quit quoting multipliers you can’t source. If you’re in B2B software, medtech, govtech, financial services, or anything else where a committee builds a shortlist before you know you’re on it, the peer-reviewed evidence backs the claim that AI-referred visitors show up further along. Make that claim, cite Kaiser and Schulze, say the complexity condition out loud. It’s a stronger position than 23x because it holds up when somebody tries to pin your ears back.

Then go find out if it’s true on your own site, because none of these studies is about you. Pull AI referrals out of the Referral bucket in GA4, where they’re currently sitting anonymously next to a Reddit thread from 2023, and watch your own conversion rate for two quarters.

If that’s too much trouble, fine. Just take 4.4x off slide four.