Is Traffic From ChatGPT and Perplexity Worth Anything Yet?

Is Traffic From ChatGPT and Perplexity Worth Anything Yet?

As of September 2026, yes, but not in the way most dashboards suggest. AI referrals are still a low single-digit share of visits for most sites, and the widely quoted conversion multiples come from a handful of studies with narrow samples. The traffic is small, unusually intent-heavy, and badly measured.

We see the same conversation in almost every client kickoff now. Someone has read that AI visitors convert twenty-three times better than organic search, and wants to know what to build. The number is real. The context around it is usually missing.

So here is the honest version, with every figure attached to the study it came from and the date it was measured.

How Much AI Traffic Is Actually Arriving?

Very little in absolute terms, growing fast in relative terms. Similarweb's analysis of worldwide desktop and mobile web traffic from June 2025 to May 2026 put AI platform referrals at an average of 770.7 million visits per month, up 117.4 percent year over year, while noting that AI referrals remain a low single-digit percentage of most sites' total traffic.

Ahrefs measured the same shape from a different angle. In a June 2025 study of 81,947 sites, Louise Linehan and Xibeijia Guan found AI traffic up roughly 9.7 times year over year while sitting at 0.25 percent of average site traffic. Over the same window, average search traffic was down about 21 percent.

Both numbers can be true and both matter. A channel can triple and still be a rounding error. The question is not whether it grew, it is whether it grew into something you can plan a quarter around.

Does AI Traffic Convert Better Than Search Traffic?

On the best public evidence, yes, by a lot. Patrick Stox published Ahrefs' own funnel data in June 2025: AI search accounted for 0.5 percent of visitors over the previous thirty days and 12.1 percent of signups. That is the 23x figure everyone quotes.

The mechanism is easy to believe. Someone who asks an assistant a specific question, reads an answer, and then deliberately clicks through to one source has already done the comparison work. They are not browsing. They arrive at the bottom of the funnel, having skipped the middle of it.

In our work we see this reflected in the kinds of questions AI visitors ask on arrival. They tend to be specific and late-stage, closer to "does this integrate with our stack" than "what is this category".

Why Should You Be Careful With That 23x Number?

Because it is one company's funnel, in one month, for a product whose buyers live on the technical web. Stox says so directly in the piece. He also reports that users from AI search click links about 75 percent less than they do in traditional organic search, and adds a warning most people skip: these click-through rates are probably the highest they will ever be, as the novelty of the format wears off.

That last sentence is the most useful line in the whole study. A high conversion rate on a tiny sample of unusually motivated clickers is not a stable channel economics. It is the early-adopter curve.

Our position is that you should treat the 23x as evidence that AI visitors are worth designing for, and not as a forecast. If you build a business case on that multiple holding at scale, you have built a business case on novelty.

Do AI Visitors Behave Differently Once They Land?

Yes, and not entirely in your favour. Ahrefs' study of 81,947 sites from May to June 2025 found AI visitors viewed 4 pages per visit against 5.2 for search visitors, and bounced 67.8 percent of the time against 63.7 percent for search.

Read together with the conversion data, that paints a sharper picture than either number alone. AI visitors arrive with clearer intent, do less exploring, and either act or leave. They are not going to wander into your case studies and talk themselves into a demo.

That has a direct design consequence. The page an AI visitor lands on has to be self-sufficient, because it will often be the only page they see.

Why Is So Much AI Traffic Invisible in Your Analytics?

Because a large share of it never carries a referrer. Similarweb notes in the same analysis that some AI visits arrive without referrer data and land in direct traffic. Anything pasted from a chat window, opened from a desktop app, or routed through a client that strips the header shows up as somebody typing your URL.

This cuts both ways. Teams undercount AI traffic when they only count referrals from known AI hostnames, and overcount it when they attribute every unexplained bump in direct traffic to ChatGPT. Both errors are common and both are confident.

We cover the practical setup in more detail in our guide to measuring AI search traffic. The short version is to segment what you can identify, label the rest as unknown, and never let the unknown bucket quietly become a success story.

What Is AI Traffic Actually Worth to a B2B Company Today?

Enough to influence what you publish, not enough to replace a pipeline channel. For a B2B software company doing a few hundred thousand sessions a year, a low single-digit percentage of AI referrals converting at several times your organic rate is a real but modest contribution.

The bigger value is indirect and harder to put in a dashboard. Being the source an assistant quotes puts your framing in front of buyers who never click anything at all. That influence looks like zero-click search, and we wrote about how to think about it in building a zero-click search strategy.

So the honest answer to "is it worth anything" is that the clicks are worth a little today, and the citations are worth more than the clicks. Most teams are measuring only the clicks.

What Should You Change on the Page an AI Visitor Lands On?

Assume a single-page visit from a late-stage reader who has already been given a summary of you by something else. That reader needs three things quickly: confirmation that the summary was right, the specific detail the summary could not carry, and an obvious next step.

In practice that means the page should state plainly what the product does and who it is for near the top, carry the concrete details an answer engine cannot summarise well such as pricing structure, limits and integrations, and put a real action within reach without a scroll marathon.

It also means being careful about pages that only make sense as part of a series. A visitor from an assistant did not read the previous three posts. Anything that assumes they did will bounce, and the bounce data says most of them do.

How Should You Measure This Without Fooling Yourself?

Track three things separately and never blend them. First, citations: which prompts surface you, on which engines, and how you are described. Second, referral sessions from identifiable AI hostnames, reported as their own channel rather than lumped into referral or direct. Third, what those sessions do, measured against your normal organic benchmark rather than against zero.

Then hold the whole thing to a rule we apply to every new channel. A conversion rate on fewer than a few hundred sessions is a story, not a measurement. Report the volume alongside the rate, every time, or the rate will be misread by someone senior who only sees the multiple.

If you are trying to work out what this does to next year's numbers, our piece on forecasting SEO traffic when AI answers take the clicks covers the modelling side.

Where Does This Go From Here?

The direction we would bet on is more volume and worse economics per visit. Growth of 117.4 percent year over year on Similarweb's numbers means the channel gets bigger. Stox's warning about novelty means the visitors get less exceptional as the format becomes ordinary.

That combination argues for a specific posture. Build for AI visitors because the work overlaps almost entirely with writing clear, self-contained, well-sourced pages, which helps everywhere else too. Do not restructure your acquisition plan around a multiple measured on one company's signup flow in one month of 2025.

If you want a second opinion on what your own AI traffic is actually doing, or a look at whether your landing pages survive a single-page visit, we are happy to walk through it. Come and find us at phoenix.studio.