In an earlier post, SEO Is Not Dead, but GEO Is Not Optional, I wrote that ChatGPT sends about 2% of the traffic that Google sends. I did not check that number carefully enough.
Marie Haynes has now published something much better. She looked at 12 months of GA4 data from 100 sites, 145.8 million sessions in total, and wrote it up as GEO is a distraction. In her data, ChatGPT sent 0.62% of the sessions that Google organic sent. All LLMs combined sent 0.73%.
So my 2% was about three times too high. It was also wrong in the direction that made AI search look bigger than it is, which did not help the point I was trying to make in that post.
What the Numbers Look Like
Marie sums up the study in one graphic. Out of every 1,000 visits that came from search engines and AI tools, 993 came from Google organic and 7 came from all the LLMs put together.
In actual sessions, that is 144,788,763 from Google organic against 1,062,833 from LLMs. That works out to about 136 Google visits for every visit from a chat tool.
Within the 0.73%, the split looks like this:
- ChatGPT: 904.7K sessions, 0.62% of Google organic
- Gemini: 53.3K sessions, 0.04%
- Perplexity: 39.9K sessions, 0.03%
- Copilot and the rest: 63.3K sessions, 0.04%
ChatGPT makes up about 85% of all LLM referral traffic. So when people talk about "AI search" sending traffic, they are mostly talking about one product.
I wrote before that the GEO spam emails promising AI visibility were not worth worrying about, and these numbers back that up. If a channel brings you 0.62% of what organic search brings you, it is worth keeping an eye on, but it is not something to reorganise your marketing around.
Higher Conversion Rate, Far Fewer Conversions
There is one number in the study that I expect people will quote on its own, so I want to put it next to the other one.
55 of the 100 sites tracked conversions in GA4. On those sites, visitors from Google organic converted at 3.02%, and visitors from ChatGPT converted at 13.8%, which is 4.6 times higher.
But Google still brought in 96.5% of all conversions across the channels she looked at. That is 27 times more conversions than all the LLMs combined.
Both of these are true. ChatGPT visitors convert well, but there are so few of them that the total stays small.
In Traffic Is Nice, but Trust Is What Converts, I argued that conversions matter more than raw traffic. I still think that, but this study shows a gap in how I put it. A conversion rate is not the same as the number of conversions. If your dashboard sorts channels by conversion rate, ChatGPT will end up at the top, and it would be easy to move budget towards your smallest channel because of it.
The 13.8% does tell us something useful, though. People who click through from ChatGPT have usually asked their questions already, and they have partly made up their minds by the time they land on your site. It is good traffic. There is just not much of it.
ChatGPT Traffic Stopped Growing in August 2025
Marie's data also shows ChatGPT sessions growing until GPT-5 launched in August 2025. After that they flattened out, and they have been slowly going down since.
Back in 2023, some of us (me included) looked at the early growth and assumed it would keep going until ChatGPT caught up with Google. That has not happened.
To be fair, this is a sample of 100 sites, and I would not read too much into one turning point. But if someone pitches you a GEO service based on how fast AI traffic is growing, it is worth asking them which year's numbers they are using.
45 of the 100 Sites Were Not Tracking Conversions
The number that surprised me the most is this one: 45 of the 100 sites had no key events set up in GA4 at all.
That is nearly half. These are businesses with websites, traffic and revenue targets, and they cannot tell which channel brought in a sale, a booking or a signup.
I think this is a bigger problem than anything to do with AI. Without a key event, you cannot tell what your own data means. Marie's 0.62% is an average across 100 sites, and your own number might be 0.2% or 4%. You will not know until you set up conversion tracking.
The fix is not complicated:
- Decide what a conversion is for your business and write it down in one sentence. It could be a purchase, a completed enquiry form, a booked call or a trial signup.
- Set it up as a key event in GA4 and check that it actually fires.
- After that, compare your channels against each other.
If you use Kafkai, it reads your GA4 data directly once your domain is verified, so you can see traffic and conversions inside the project without exporting anything. The GA4 setup guide explains how to connect it.
I have also written about what to look for in your website traffic, which covers the metrics that still matter when your traffic sources change. Browsers have also been restricting third-party cookies for years, and I went through what actually happened in the disappearance of third-party cookies. Tracking your own first-party events is now the minimum.
Most AI Search Traffic Still Comes Through Google
Marie's conclusion is that the AI search that actually sends people to businesses today is Google. That means AI Overviews, AI Mode, and a ranking system that now relies heavily on AI.
This lines up with the AI Overviews outbound clicks experiment I wrote about. When researchers removed AI Overviews from a set of search results, outbound clicks went up 68%, and the quality of the visits did not drop. AI Overviews show up on 53% of informational searches and around 15% of transactional ones. So the clicks are being lost mostly on informational content, which is where most content teams spend their budget.
Google answers more questions directly on the results page than it used to. It is also still where nearly all search traffic comes from. Both are true at the same time.
Looking back at SEO Is Not Dead, but GEO Is Not Optional, the part I still agree with is the writing advice: give clear answers, back them up with evidence, show who wrote the page, and structure it so the information is easy to pull out. But all of that is just good SEO. It does not need a separate name or a separate budget. I would also not pay for a GEO retainer that is sold on traffic, because going by Marie's data, it is chasing 0.62%.
The same goes for citations inside AI Overviews. In the AI Overview citation clicks analysis, only 7.1% of outbound clicks on AI Overview pages came from links inside the AI answer itself. Being cited puts your name in front of people, but it does not bring many visitors.
What the Replies Get Right
Marie's post got a lot of replies, and many of them were thoughtful. Several people agreed with her numbers but disagreed with her conclusion, and I think some of their points are fair.
The first point is that clicks undercount influence. AI tools are designed to give answers, not to send people to websites. Someone might get an answer from ChatGPT, then search for the brand on Google, call the business, or just walk into the shop. GA4 records that as direct traffic, a branded search or Google organic. Gaetano DiNardi made this the main point of his reply. In his view the numbers are right, but the conclusion misses that people now do a lot of their research inside the AI tool, and sending clicks was never what the tool was for. I agree with him. I made a similar point before, but less clearly.
The second point is that Google gets the credit for decisions made elsewhere. Someone uses AI to narrow down their options, then searches on Google to find the site, and GA4 counts it as Google organic. John Arndt put it more directly: AI usage keeps going up while referral traffic stays flat, because people simply do not click. That is a gap in measurement, not proof that AI has no effect. The resoneo study I covered in what ChatGPT actually reads showed how selectively these tools read a page, and being read still matters even if nobody clicks through.
The third point is about who your audience is. A 0.62% average can hide a much higher share among B2B buyers, developers and other professionals, who use ChatGPT and Claude more than the general public does. For a local business, the decision can happen inside the answer itself. Someone asks for a good plumber nearby, and they call the first one mentioned instead of opening five websites. So before you rely on any benchmark, including this one, check your own GA4. Filter your traffic by source for chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. You can do this in GA4 explorations without any add-on. If your share turns out to be 3%, your situation is different from the typical site in Marie's study, and you should plan around your own number.
The fourth point is that the work of getting mentioned by AI tools leads into the agentic web, which Marie covers later in her post. AI tools cite what is on your own site, and an AI agent choosing a supplier will most likely pick from the businesses that today's tools already recommend. I agree with that, but I would describe the work differently. Making your site clear, accurate and easy to read for both people and machines is SEO. It is the same work, so I do not think it needs its own budget line.
There was also a practical point that I understand. Telling a client or a marketing director that "GEO is a distraction" is not easy, especially while Google keeps showing AI Mode on more searches. In that situation, I think it is more useful to explain that this is part of SEO, and to show them the AI referral visits in GA4 next to the Google search clicks.
Where I still agree with Marie is the arithmetic. However you handle attribution, 0.7% of sessions is not going to carry next quarter's revenue plan. AI tools probably have more influence than the click numbers show, but that influence is not arriving as visits you can count.
The Agentic Web Is the Bigger Change
I found the second half of Marie's post more useful than the first. She points to what she calls the agentic web, where AI agents shop and book things on behalf of people. Her examples are Universal Cart, the Universal Commerce Protocol, WebMCP and intelligent eyewear.
If you have not come across these yet, here is a quick summary of each:
- Universal Cart puts products from several retailers into one checkout inside Google Wallet. It also watches for price changes and restocks, so a shopper can buy without visiting the retailer's site.
- The Universal Commerce Protocol is an open standard run by Google, Shopify and Stripe. It lets AI agents compare stock and complete a purchase, while your business stays the merchant of record.
- WebMCP lets a website make its own functions, such as forms, calculators and booking flows, available to an AI agent acting for a customer.
- Intelligent eyewear is Google's name for glasses with a camera and Gemini built in (they avoid calling them AI glasses). Google says the first devices, built on Android XR, are coming out around now.
Marie quotes Sundar Pichai saying that search will become an agent manager. I could not verify the quote myself, so I am relying on her for it. Either way, the products she lists all point the same way. Search is moving from somewhere you go to look things up, to something that does tasks for you.
I think this matters more than GEO, because it changes where the purchase happens. If an agent can buy something for a customer inside Google, then what gets compared is your product data, your stock levels and your prices. Your landing page copy may never be read.
What to do about it depends on what you sell.
If you sell products, your Merchant Center feed is now part of your shop. Your prices, stock status and product structured data need to be correct before an agent can include you in anything. Many small retailers still treat the feed as an afterthought, and that is where they will fall behind.
If you sell services, ask yourself what an agent would need to complete a booking with you. If the answer is "download a PDF form and call us", that is worth fixing, and fixing it makes booking easier for your human customers too. WebMCP is still at an early stage, but you can start preparing for the idea behind it now.
I wrote about the MCP side of this in AI data monetization and MCP. WebMCP applies the same idea to websites instead of APIs. I think it is the first agent standard that an ordinary business can start preparing for this year.
What I Would Do This Quarter
- Check your own LLM traffic share before trusting any benchmark, including this one. It takes one GA4 exploration with four source filters, about 30 minutes.
- Set up a key event and make sure it fires. If you are in the same position as those 45 sites, you cannot judge any of the other points on this list.
- Keep investing in Google, and put more of the budget on pages where people still click. Transactional and navigational searches show AI Overviews much less often than informational ones.
- Get one thing ready for AI agents. That could be a clean product feed if you sell products, or one booking or quote tool that an agent could use if you sell services.
- Keep writing pages with clear answers, evidence and a named author. Treat it as part of your SEO work, not as a separate AI project.
Summary
Marie Haynes measured something I had only estimated, and my estimate was off. ChatGPT sends 0.62% of the traffic that Google organic sends, and all LLMs together send 0.73%, not the 2% I quoted. ChatGPT visitors convert 4.6 times better, but Google still brings in 27 times more conversions. ChatGPT traffic stopped growing in August 2025. And 45 of the 100 sites in her study were not tracking conversions at all.
The replies to her post make a fair point that GA4 cannot see everything AI tools do, because a lot of their influence never turns into a click. I agree with that. But it does not change where the visits and the conversions come from today, and that is still Google.
So my advice is to keep your budget on Google search, check your own AI referral numbers in GA4, and get one part of your site ready for AI agents this quarter.
That is also where Kafkai's tools sit: keyword and competitor data, rankings, and your own analytics once your domain is verified. Getting found and getting clicked has not changed as much as the last two years of conference talks suggested.