BrightEdge published its AI referral data for January through August on September 24. ChatGPT accounted for 95.1 percent of all traffic referred by AI assistants in August. Its referrals grew 101 percent since January, with a 36 percent jump in August alone. Gemini held 2.4 percent, Perplexity 1.1 percent, and Claude 0.8 percent. Total AI referral traffic grew about 73 percent in under a year, and BrightEdge notes it is still small next to Google search.
That last point matters. For most credit unions I work with, AI assistant referrals are a fraction of one percent of sessions. The reason to care is not the volume. It is that these sessions convert at several times the site average, and that the mix behind them is changing in a way most dashboards do not show.
What changed this month
One of my clients started running ads inside ChatGPT in September. The first thing I noticed in the monthly pull was that ChatGPT sessions were up sharply from August. The second thing I noticed was that GA4 filed every one of those paid clicks under the same source as the organic ones: chatgpt.com. The only difference was the medium. Organic citations came through as ai-assistant. The ads came through as cpc.
Once I split them, the picture flipped. Organic ChatGPT referrals per day were flat against August. All of the growth was paid. The paid sessions lasted a fraction as long as the organic ones and converted at a much lower rate. On the site average they were still decent. Against the organic ChatGPT sessions they looked like display traffic.
If I had reported the source total, the story would have been "AI referrals up sharply." The true story was "organic AI visibility is flat, and we bought some clicks that have not proven themselves yet." Those lead to different decisions.
Five things to check before the number goes in your deck
First, open GA4, go to Traffic acquisition, and set the primary dimension to Session source / medium instead of Session source. Type chatgpt into the search box. If you see only chatgpt.com / referral or chatgpt.com / ai-assistant, you have organic only. If you also see chatgpt.com / cpc or a campaign name, someone is buying placement, and you need to report the two rows separately from now on.
Second, compare the two rows on average engagement time and key event rate, not sessions. In every case I have looked at, organic AI referrals run well above the site average on both. Paid ones so far run below organic and sometimes below the site average. If your paid row is converting like your organic row, that is worth knowing too. Either way, sessions alone will mislead you.
Third, trend the organic row per day, not per month. September has 30 days and August has 31. A month with a holiday looks like a dip. Per-day counts are the only way to tell flat from down.
Fourth, check whether your paid ChatGPT campaign is tagged at all. If the ads platform is not passing a medium, the paid clicks show up in the organic row and you will never separate them. Fix the tagging before the next reporting cycle, because you cannot retroactively split a month that was blended.
Fifth, look at the other assistants even though they are small. Gemini, Copilot and Claude referrals are a handful of sessions a month for most credit unions. They still tell you whether your content is being cited outside of ChatGPT. A zero on Perplexity is a data point, not a rounding error.
Earned and bought are not the same channel
BrightEdge's CEO said in the release that ChatGPT is sending substantially more users to the web than it was in January. I believe that. I can see it in client data. What I can also see is that a share of those users are now arriving because someone paid for the placement, and the analytics tools have not caught up to the difference.
Earned citations in an AI answer are the cheapest, highest-intent traffic a credit union can get right now. Paid placement in the same assistant might turn out to be a good buy. But they are not the same thing, and the moment they get reported as one number, you lose the ability to manage either one.
Split the row. Then report the number.