Is GEO a Distraction? 35 Sites Say AI Traffic Is Tiny, and It Buys More

AI traffic is just 1.3% of Google organic across 35 sites, but it converts better on 8 of 10. The data from both sides of the GEO debate, plus a GA4 check.

Part of the GEO Roadmap › Stage 7: Measure

Short answer: GEO isn’t a distraction, but it isn’t a traffic play either. Across 35 sites I have GA4 access to, AI sends a median of just 1.3% of the sessions Google organic sends. Yet on 8 of the 10 sites with clean conversion tracking, AI visitors convert at a higher rate than Google visitors, a median of 1.6 times. AI is a pre-persuasion channel. Measure it like a conversion channel. Budget it like a brand channel.

Below: who said what in last week’s “GEO is a distraction” debate, my 35-site data, why the Turkish market matters here, why AI visitors arrive half convinced, and a 20-minute GA4 check you can run on your own site this week.

In October 2026, our industry had a quiet civil war. Marie Haynes said GEO is a distraction. Kevin Indig and SEO Works landed in a similar place. Rand Fishkin bundled the three and asked whether AI visibility is a marketing boondoggle. The same day, Mike King fired back: every one of these studies grades a brand channel with a traffic ruler.

I read both sides. One thing was missing: conversions. Everyone was arguing about how much traffic AI sends. A business owner asks a different question: “Do those visitors buy?” So I ran Marie’s method on my own client portfolio, most of it in Turkey, a market nobody in this debate had measured.

The thesis: AI isn’t a traffic channel. It’s a pre-persuasion channel.

Both sides are debating the wrong question. It isn’t “How much traffic does AI send?” It’s “How valuable is the visitor AI sends, and who persuaded them?”

In my data, AI traffic is 1.3% of Google organic (median). Marie is right about that. But on 8 of the 10 sites with clean conversion tracking, AI visitors convert better than Google visitors. The other two are level. The answer persuades the visitor before they ever reach the site.

And the people who never click aren’t wasted. Everyone who sees your brand in an answer has met it one more time, like someone driving past a billboard. So let’s neither bin GEO as hype nor sell it on traffic promises.

1. The map of the debate: Who said what?

Two people didn’t weigh in that week but have a lot to say on it. Eli Schwartz has argued all year that the top of the funnel moves to AI while the value stays on decision-stage pages (AirOps, April). And on Oct. 7, Metehan Yesilyurt of Peec AI showed that 50,000 unverified stats pages, made for $50, pulled more than 10,000 ChatGPT sessions after Google had dropped them (the experiment). A channel that didn’t matter wouldn’t be this easy to game.

Two more studies added data. SISTRIX found across 96,418 URLs that tables of contents, Article schema, genuine FAQs and author boxes with photos go hand in hand with citations, while llms.txt shows no effect (correlation, not causation). Seer ran the same prompt twice, eight minutes apart, and none of the background search queries repeated.

2. My data: 35 sites, 12 months, same method

I excluded large enterprise country properties and test accounts, then ran one report across the GA4 properties I have access to: Google organic sessions vs. sessions from ChatGPT, Gemini, Perplexity, Copilot, Claude and other large language models, Oct. 1, 2025, to Sept. 30, 2026. Of the 35 sites that returned data, 22 sell mainly in Turkey and 13 abroad or across markets (U.S., U.K., Europe). Everything is aggregated and anonymized.

Traffic: Marie is right

Source: GA4 data from 35 sites, Oct. 1, 2025, to Sept. 30, 2026.

Excluding the largest site, the pooled rate is 0.9%, very close to Marie’s 0.73%. And in Turkey, the share is a third of what it is abroad.

Conversions: This is where the story turns

I set two conditions: working key event tracking and at least 100 AI sessions. Ten sites passed. Because each site defines a conversion differently, I compared every site with its own Google organic traffic.

Source: GA4, session key event rate, Oct. 1, 2025, to Sept. 30, 2026.

On volume, the picture flips: excluding the largest site, Google drove 57 times more conversions than AI. So two sentences are true at once: AI visitors convert better. And there are very few of them.

Limitations: My sample is smaller than Marie’s, and the conversion comparison covers 10 sites. These sites aren’t random. They’re a portfolio I advise or have data access to, so they invest in SEO. Conversion definitions differ (forms, calls, sign-ups, purchases), which is why I compared each site with itself. Some AI traffic lands in GA4 as “direct,” so the real share may be higher.

I also turned this analysis into a step-by-step deck in NotebookLM. You can open it here.

3. The Turkey lens: Smaller channel, same premium

This is the part you won’t find in the U.S. studies. In Turkey, AI traffic is even smaller: a median of 1.09% of Google organic, a third of the share on the international sites in my sample. ChatGPT is widely used here, but buyers still go to Google when it’s time to act.

Yet the conversion premium survives the move. Of the six Turkey-market sites in the conversion comparison, five converted better from AI, from 1.21 times up to 7.9 times, and one was level. The channel is smaller in a non-English market. The visitor isn’t worse.

That matters for anyone running SEO outside the U.S. If you judge AI by U.S. traffic benchmarks, you’ll write it off twice: once because the channel is small, and again because it’s even smaller in your market. Judge it by what each visitor is worth.

4. Looking at it through conversions

I’ve done SEO and conversion optimization since 2009, and the first thing I saw in this table was this: we didn’t persuade these visitors.

The visitor from Google is at the start of their research: 10 tabs open, comparing, bouncing back. The visitor from AI has already compared. The answer told them, “Look at these three. This one fits you.” Whoever clicks through has half made up their mind.

Three conclusions:

1. What we’re measuring isn’t traffic. It’s pre-persuasion. Conversion starts in the answer, not on the site. If the answer didn’t recommend you, that visitor never arrives. If it did, half the job is done.
2. The traffic ruler hides the value. Deleting a small channel whose every visitor is worth more, because of its session count, is like judging email or referral programs by their traffic share. We never did that there.
3. Pre-persuasion can break on the page. On 23 of 63 sites in the SEO Works sample, AI traffic converted worse. My read: AI sends the visitor with a promise, and the page doesn’t keep it. The answer said “transparent pricing,” your page says “contact us.” The answer said “quick setup,” your form has 14 fields. The conversion gap measures the distance between the promise in the answer and the promise on the page.

That brings us back to Eli Schwartz: AI eats the top of the funnel and sends you the moment of decision. If the page isn’t ready for that moment, you’re sending your best visitor to your worst page. That gap between promise and page is exactly what we look for in conversion optimization work.

5. Nobody measures a billboard by clicks

There’s a half the table can’t see. Someone asks ChatGPT to recommend an SEO consultant for an enterprise site. They see three brands and click none. Two weeks later, the budget is approved and they type your brand into Google. GA4 records it as Google organic, branded search. AI’s share in that sale shows as zero.

Nobody measures a highway billboard by how many people clicked it. It’s hard, indirect and incomplete. But no marketing director says, “I can’t measure it, so it doesn’t work.” An AI answer is a step ahead of a billboard: it reaches the user at the exact moment they ask, as a recommendation written for their situation.

Now let’s look in the mirror. I think there’s another reason “GEO is hype” spread so fast, and nobody says it out loud: GEO is hard to explain to clients. Telling a board “ChatGPT recommends you for these questions, your competitor for those, and describes you in the wrong category” takes far more work than “we moved from seventh to third for this keyword.”

Rank reports have been our industry’s comfort zone for years. They mean less every month (in October, Ahrefs said Google is showing rank trackers different results on 5% to 20% of queries), but they still fit on a slide. That’s why “GEO is a distraction” felt good to so many people: it says we don’t have to change how we work. I think the opposite. Dismissing a channel because it’s hard to measure is like dismissing social media because “likes don’t sell.”

The book I started reading that week touches exactly this point: Generative Engine Optimization: The Definitive Guide to AI SEO by Ross Hudgens, founder of Siege Media, published by Wiley. I haven’t finished it yet. What I value for this debate is that it explains GEO through real use cases: according to the publisher, there are case studies from Zapier, Bluevine and Siege Media itself. One chapter, on the GEO-first organization, covers getting stakeholders on board. In other words, the “how do I explain GEO to a client” question.

6. An aside: Look at the data, not the name

The most shared link in this debate was the one with the least new information. Rand Fishkin’s post is a five-minute roundup with no data of its own. The headline asks whether AI visibility is a marketing boondoggle. The ending answers: measure it as a brand channel, a digital billboard. People shared the headline, not the conclusion. And the conclusion was Mike King’s point anyway.

Then there’s memory. In August, it was Rand who wrote that new research suggests AI tools often already know which brands they’re searching for. That’s an argument for the visibility that produces no click. August Rand was right. October Rand measured that sentence’s value in clicks.

Credit where it’s due: he founded Moz and taught many of us the language of this industry. But the real problem isn’t any one person. It’s our reflex. A post drops, hundreds of comments say “so true,” and nobody asks, “Where’s the data?” The real contribution that week came from elsewhere: Marie opened up 100 sites’ data, Mike wrote his argument with sources, and Kevin flagged the limits of his own findings. Let’s look at what people base their claims on, not who makes them. This post included.

7. Two distinctions the debate missed

Memory and citation are two different games. As I wrote in ChatGPT Decides Before It Searches, models pick the brands they’ll recommend from memory, before they search. In Suganthan Mohanadasan’s analysis, brands the model writes into its own query appear in the answer 68.9% of the time; brands it only retrieves, 2.1%. All three studies in this debate measure clicks from citations. What drives sales is being recommended, and that happens without a click. Kevin not being able to isolate that effect doesn’t mean “no effect.” It means “we can’t see it with today’s tools.”

Citation is a diagnostic, not a success metric. My suggestion: track your AI conversion rate next to your right-identity rate, the share of your AI visibility that lands on the topics you actually sell. When I built a free GEO funnel from Search Console and Bing data, I measured it for stradiji.com: 56% of my AI impressions sat on three pieces of content outside my core business. Being visible for the wrong topic doesn’t bring sales on the right one.

8. Our own test: U.S. press release, week 2

We’re running a 12-week U.S. press release test on stradiji.com, which started with I Asked AI for the Best GEO Agency. Week 2:

• Links (Semrush, Oct. 19 to 25): About 77 new referring domains, almost all automated “checker” sites. The first week’s wave has faded, with no new editorial coverage.
• U.S. Google impressions: 645 (Oct. 5 to 11), 728 (Oct. 12 to 18), 465 (Oct. 19 to 23, five days). Same band.
• GeoGenie (U.S. workspace): From Oct. 20 to 25, 139 answers carried 533 citations, and two were ours (0.4%). We’re 102nd of 261 cited domains. A week earlier it was 1 of 425 (0.2%), 205th. A small twitch, but both citations point to the homepage, not the release. Our visibility on the “press releases and earned media” topic is still 0%.

No movement from the release yet, but it’s early: Notified’s study puts 56% of such citations after week one. In light of this debate, I’m widening the test’s question: if citations come, what do the visitors from those citations do? In week 6 (around Nov. 23), I’ll put citations, impressions and form conversions side by side.

The one thing to do this week: Join the debate with your own data

1. In GA4, open Explore > Blank. Dimension: “Session source.” Metrics: “Sessions” and “Session key event rate.” Last 12 months.
2. Filter the source with this regex and add a “google / organic” row:
   chatgpt|openai|perplexity|gemini\.google|copilot|claude\.ai|deepseek|grok|meta\.ai|mistral
3. What share of Google are your AI sessions, and how many times better or worse do they convert?
4. Add “Landing page.” Open the three pages that get the most AI traffic and ask: Does this page keep the promise the AI made?

If you don’t track conversions (45 of Marie’s 100 sites didn’t), that’s step one. You can’t call something unimportant if you don’t measure it.

Closing

Measure a channel with the wrong ruler and you’ll see it as either bigger or smaller than it is. AI visitors are few, but they arrive with their minds half made up. As I wrote when HubSpot lost 140 million visits and posted record revenue, traffic and revenue can part ways. Your traffic can shrink while your sales don’t.

So here’s the question: Does your site keep the promise AI makes on your behalf? If you’d like to find out which AI answers recommend you and whether your pages close the deal, that’s where our GEO consulting work starts.