Yes. On one of our own sites over six months, AI tools sent 361 sessions against 7,196 from Google organic, equal to about 5 percent of the Google organic total.
That site is in travel services, a category where customers often research remotely, so do not treat the figure as a benchmark for other businesses.
The actual numbers
Data period: 1 February 2026 to 31 July 2026. Source: Google Analytics 4, one site we own and run, in travel services. We update this dataset every six months, so the article does not become a snapshot of an old market.
| Source | Sessions | Bounce rate |
|---|---|---|
| Google organic | 7,196 | 28.1% |
| Direct | 1,542 | 65.3% |
| ChatGPT | 312 | 51.6% |
| Gemini | 28 | 25.0% |
| Perplexity | 10 | 41.7% |
| Copilot | 8 | 37.5% |
| Claude | 3 | 33.3% |
ChatGPT accounted for about 86 percent of tracked AI referral sessions in this dataset.
We are publishing our own numbers because almost nobody does. Plenty of agencies will tell you AI search is changing everything. Very few will show you what it looks like on a real site.
The engagement finding, and what it does not tell you
51.6 percent of ChatGPT sessions were not engaged sessions in GA4, compared with 28.1 percent from Google organic.
Worth being precise about what that means. In GA4 an engaged session is one lasting more than ten seconds, or with two or more page views, or with a key event. So a non engaged session is not the same thing as someone leaving immediately.
And here is what this dataset cannot tell you. Key events were never configured on this property during the period, so there is no conversion data for any channel, including Google organic. On this site, ChatGPT referral sessions showed lower engagement than Google organic. We cannot say whether they were more or less commercially valuable, because the measurement was not there.
The standard pitch says AI referrals arrive better qualified because the AI has already answered the easy questions. On this data they arrived less engaged. Whether they bought anything is unknown.
Traffic going up is not the same as traffic that buys
Worth separating out, because it applies well beyond AI referrals.
A rising traffic line feels like progress. What matters is whether the people arriving can actually buy from you.
An example from our own portfolio
One of our sites generates a large volume of search impressions from all over the world. The traffic chart looks healthy.
The business it serves sells to buyers in Thailand. Over eleven months it produced eleven enquiries.
The impressions and visits were real, but over eleven months the site produced only eleven enquiries. High visibility did not translate into much commercial activity. Somebody in another country reading about a Bangkok service they will never use is a visit, not a lead.
So when anyone shows you a traffic increase, ask two questions. Where are these people, and what did they do after arriving. A number going up answers neither.
how we approach SEO and AI visibility
Why your category probably affects the opportunity
Here is what almost nobody selling AI optimisation will discuss.
Category is likely to affect this a great deal, because different buying decisions create different research behaviour.
A travel business naturally answers many of the questions people research before arriving somewhere they have never been. A local purchase involving site visits, physical samples, quotations and a company to hold responsible creates a different customer journey, and an AI recommendation plays a smaller part in it.
Our data comes from the first type of business. So it should not be treated as a benchmark for every category.
Working hypotheses, not measured facts
Higher AI referral likelihood: travel, tourism, restaurants and venues, activities, hotels, anything a stranger researches remotely, general information and comparison questions.
Lower AI referral likelihood: local trades, construction and installation, anything needing a site visit, a quotation or a licence check, and B2B where buyers use tenders, referrals and existing suppliers.
If your business is in the second group, be sceptical of anyone selling AI optimisation as a budget line.
Which pages received AI referral traffic on this site?
More useful than the totals.
The pages picking up AI referrals were not the service pages. They fell into three groups:
- Process answers. How a specific procedure works, how long something takes.
- Comparisons and lists. Best of pages naming actual venues and options.
- Price and timing pages. Real figures rather than “contact us for a quote”.
Meanwhile the plain service pages, the ones describing what we offer and inviting an enquiry, got two or three sessions each.
On this site, informational pages received far more tracked AI referral traffic than plain service pages. The strongest examples answered specific questions, compared options, or contained concrete prices and timings.
That is an observation from one site, not a rule for every AI system or every category. But it points the same way as normal search and the same way as a human reader, which is reassuring rather than surprising.
One risk worth knowing about
AI answers can still surface outdated information.
Search systems recrawl and refresh sources at different times, and an answer may also draw on information from another source that has not been updated.
We know this directly. Our best of list included venues that have since closed, and we removed them. That does not guarantee every AI answer reflects the change immediately.
So an AI can recommend a business that shut some time ago, on the strength of information that is no longer current.
That is your reputation carrying someone else’s error. Anything with a shelf life, prices, opening hours, venue lists, product availability, needs regular review. It is no longer only your visitors reading it.
So what should you actually do?
Nothing new, if you are already doing it properly
Publish useful specifics. Answer important questions clearly. Keep dated information current.
That works for Google, for AI tools, and for the person reading. If someone is selling AI optimisation as a separate monthly service, ask what it includes beyond normal SEO foundations, crawl access, useful content and keeping information current.
Check you are not blocked
Before anything about how content is written, one technical thing matters.
ChatGPT search uses OpenAI’s search crawler to discover and surface websites, and it can also draw on third party search providers. OpenAI says public websites can appear in ChatGPT search, and allowing that crawler access to the relevant pages is an important eligibility check. A blocked page can still show up as a bare link and title in limited circumstances, but it will not be summarised or quoted.
Google’s AI features work differently. They use its Search index and core ranking systems, so the usual technical and content foundations apply there. Google states that no special formatting or AI files are needed for them.
So the honest position is not that SEO has been replaced. The same fundamentals still matter, plus one crawler setting that often goes unchecked.
How to check your own site tonight
Open Google Analytics, go to your traffic acquisition report, and look at session source.
Then look for:
- chatgpt.com
- gemini.google.com
- perplexity.ai and perplexity
- copilot.com
- claude.ai
OpenAI appends a utm_source parameter to ChatGPT search referral URLs, which makes those clicks measurable in analytics.
Two things to know about the data. Sources sometimes appear more than once under different medium labels, so check the totals rather than individual rows. And tracked referral data is never a perfect measurement of every AI influenced visit, so treat the totals as what GA4 can attribute rather than the complete universe of AI influence.
Then compare AI referrals with your own Google organic traffic and watch the direction over time. Your own category and customer behaviour matter more than somebody else’s benchmark.
Our honest position
AI referral traffic is real, measurable and currently small on this site. It is heavily concentrated in ChatGPT.
It does not require a new service, a new acronym or a new budget line. It requires what it required before: pages that answer the question clearly, with facts, kept up to date.
We will publish updated numbers every six months. If the picture changes, we will say so.
