Most GEO marketing rests on a premise nobody checks: that AI assistants are pulling people away from Google, so the budget should follow them. A survey published by eMarketer on July 27, 2026 points the other way. The people using AI most heavily are also the people using Google most heavily.
We can test that claim against something better than intuition. We run the weekly search program on eight live domains, and we hold the crawler logs and the session data for all of them. So we went and looked.
What did the eMarketer survey actually find?
eMarketer reports that 82% of weekly AI users search on Google every day, against 71% of internet users overall. The heavier the AI habit, the heavier the Google habit. The figures come from its "AI in the Consumer Journey: H1 2026" survey of 1,570 US internet users aged 15 to 80, fielded between April 14 and 30, 2026.
The explanation offered is that AI is not replacing search, it is splitting the job in two. 63% of people who search with an assistant reach for it mainly to get a direct answer to a specific question. When they want to browse, compare options, or see who else is saying what, they go back to a search engine.
Does heavy AI use mean people search less?
No. They search more. If assistants were substitutes for search engines, adoption would show up as displacement: heavy AI users would search less. That is the assumption behind every pitch to move budget out of SEO and into GEO, and the data points the other way.
What a complement looks like in data is the pattern above. Someone who asks an assistant fifteen questions a day is someone with fifteen questions a day, and a good share of those turn into a follow-up search. The behaviour that makes a person a heavy AI user is the same behaviour that makes them a heavy searcher.
What do our own crawler and session logs show?
Two independent measurements sit in our data, and they say different things. AI systems are crawling these sites hard. AI systems are barely sending anyone.
Which crawler hits these sites hardest, ChatGPT or Googlebot?
Across the eight domains, ChatGPT's fetcher is the single heaviest crawler on four of them. On the two busiest sites it is not close.
| Domain | ChatGPT User | Googlebot | Ratio |
|---|---|---|---|
| hyperliquidguide.com | 1,885 req/day | 336 req/day | 5.6x |
| carsnipe.com | 306 req/day | 56 req/day | 5.5x |
| lighterpedia.com | 45 req/day | 41 req/day | 1.1x |
| asterpedia.com | 38 req/day | 29 req/day | 1.3x |
| immlawcenter.com | 48 req/day | 74 req/day | 0.7x |
| drcritter.com | 52 req/day | 230 req/day | 0.2x |
Read on its own, the top half of that table is the case for abandoning SEO. An engine crawling a site five times harder than Google looks like where the audience is going.
Where do the sessions actually come from?
Almost entirely from Google. Line the same domains up against where the sessions actually came from over the same thirty days.
| Domain | Google sessions | chatgpt.com sessions | Ratio |
|---|---|---|---|
| hyperliquidguide.com | 21,734 | 173 | 126x |
| carsnipe.com | 760 | 287 | 2.6x |
| drcritter.com | 530 | 13 | 41x |
| immlawcenter.com | 101 | 79 | 1.3x |
Hyperliquid Guide is the clearest case in the set. ChatGPT crawls it 5.6 times harder than Google does, and sends it one visitor for every 126 that Google sends. The crawl is not a leading indicator of traffic. It is a different thing entirely, and treating the two as the same number is how a budget decision goes wrong.
CarSnipe is the exception worth watching. There, chatgpt.com is the largest named referrer after Google, at 287 sessions against Google's 760. That is a real share rather than a rounding error. It is also a site about comparing specific used cars, which happens to be the exact shape of question people take to an assistant instead of to a results page.
Why are heavy AI users also heavy searchers?
Our numbers and the survey fit together once you stop treating the two systems as competitors for the same moment. An assistant is good at collapsing a question into an answer. A search engine is good at showing you the field so you can judge it yourself. Those are different jobs, and most real decisions need both.
That is also why the crawling is so heavy while the referrals are so light. The assistant reads the page in order to answer without sending anyone. The visit arrives later, if at all, through a search for the brand the answer named or a direct return. This is the same pattern we wrote about in what Google's AI-search click number actually means for small sites: impressions and reads hold up while clicks thin out.
How should you split a budget between SEO and GEO?
On this evidence, GEO reads less like a new line item and more like a second use for work you are mostly doing anyway.
Keep organic funded. On every domain we measured, Google organic delivered more sessions than every AI referrer combined, so moving that budget elsewhere means moving it away from the only channel currently producing volume. Then do the GEO work that overlaps with it: crawlability, clean HTML, schema, entity clarity, and passages that stand on their own serve both systems at once. Very little of the GEO layer is genuinely separate spend, which is the argument we make at length in GEO SEO vs classic SEO.
What we would avoid is pricing GEO on today's referral traffic. On most of these sites it sits under 1% of sessions, and the case for doing it is the citation plus the branded search that tends to follow, not a click you can attribute this month. Which sources actually win those citations is a separate question, and we track it weekly in the AI search citation benchmark. And measure your own ratio rather than borrowing an industry average. CarSnipe gets 12.6% of its sessions from chatgpt.com while Hyperliquid Guide gets 0.5%, and that is the same portfolio, the same year, and the same weekly work on both. Subject matter is what decides which one you are.
What are the limits of this data?
This is a small sample and we are not going to dress it up as more.
- Eight domains, all managed by us, over a single thirty-day window ending July 27, 2026. The three newest sites have 23 to 24 days of data rather than 30.
- The portfolio is not representative. It skews toward niche informational and comparison sites, with one local service business and one wildlife removal company. A retailer or a publisher would likely look nothing like this.
- Crawler attribution comes from Cloudflare's user-agent classification, which can be spoofed and which lumps some fetchers together.
- Session sources come from GA4, where a large share of traffic lands in "(direct)" or "(not set)". Some AI referrals almost certainly hide in those buckets, so the AI numbers here are a floor rather than a full count.
- One domain in the set, drcritter.com, is dominated by Yandex crawling at a volume that has nothing to do with this question. We left it in the crawl table rather than quietly dropping the inconvenient row.
- We did not run a controlled test. This is observational data from live sites, and it agrees with the eMarketer survey rather than proving it.
The analysis pipeline that produced these figures is automated and AI-assisted, which is how it runs every week rather than once a quarter. The numbers are pulled from our own Cloudflare and GA4 records, not estimated, but the drafting had AI in the loop and a human decided what shipped.
Sources: eMarketer, "AI's heaviest users lean on Google more, not less" (July 27, 2026), and first-party Cloudflare bot logs plus GA4 session data across eight ConceptSEO-managed domains, June 28 to July 27, 2026.