I traced every ChatGPT-sourced enrollment back to its first click. None of them started on a blog.
Between April and September 2026, Google organic traffic to Kraftshala, the online business school where I run SEO, was 43% lower than in the same six months of 2025. Non-branded blog traffic was down 78%.
Enrollments from organic and AI search went up 17%.
The difference came from one channel. AI search, almost entirely ChatGPT, produced more than half of those enrollments, five times what it had produced in the six months before.
Traffic charts could not tell me why, so I opened the CRM record for each AI-search enrollment and read the journey from first visit to payment. A couple had no usable first-touch data. This is what the rest showed.
The channel against Google organic
| Apr–Sep 2026 | AI search, relative to Google organic |
|---|---|
| Visits | 7% of Google's |
| Leads | 44% of Google's |
| Enrollments | 1.25x Google's |
| Lead to enrollment | 7.8% vs 2.7% |
AI search sent a fraction of the visits Google did and produced more enrollments. A lead from ChatGPT was about three times as likely to enrol as a lead from Google.
Finding 1: every one of them landed on the homepage or a programme page
Just over half landed on the homepage. The rest landed on a programme page. None landed on a blog.
That is not where the AI traffic goes. Over the same six months, blogs received 37% of all AI-search visits and programme pages 32%. Blogs got the most visits and none of the traceable enrollments.
The reading I trust most: when ChatGPT links to your homepage or programme page, it has usually just named you as an answer to a buying question. When it links to a blog, it is citing you as a source for information. The first visitor is choosing a course. The second is reading a footnote.
Finding 2: they apply almost immediately
Just over half applied on the day of their first visit. About three in four applied within three days.
One person landed on the homepage from ChatGPT, applied, and submitted the screening test the same day. These visitors had done their comparison inside ChatGPT before they arrived. The site's job was to confirm, not to persuade.
Finding 3: applying fast is not the same as paying fast
Time from first visit to enrollment ranged from 3 days to 78. Close to half enrolled within nine days. About a quarter took more than 75.
The slow ones stalled at different points: before applying, before the screening test, or after it. So the channel delivers intent quickly, but it does not remove the follow-up work. A team that judges an AI-search cohort after 30 days will undercount it.
Finding 4: ChatGPT creates branded Google searches
More than a third went to Google after their first AI visit and searched the brand name, or the brand name plus "course fees", then came back through a branded result or a branded ad.
In most analytics setups that second visit gets the credit. If you see branded search holding up while non-branded traffic falls, part of that demand may be coming from AI answers, and your attribution is handing it to brand.
Finding 5: a live session sat in the middle of many journeys
Close to half registered for or attended one of our live sessions between first visit and enrollment. For a considered purchase, the AI answer starts the journey and a human touchpoint often closes it.
Then the channel slowed down
It would be easy to stop here. The monthly trend says otherwise. Each row is indexed to its peak month, May, at 100.
| Index (May = 100) | Apr | May | Jun | Jul | Aug | Sep |
|---|---|---|---|---|---|---|
| AI-search visits | 36 | 100 | 73 | 57 | 65 | 70 |
| Leads | 27 | 100 | 52 | 42 | 42 | 25 |
May was the peak. By September, visits were down about 30% from May and leads were down 75%. Enrollments followed the leads.
Findings 1 and 2 explain it. The visits that fell were the valuable ones: the homepage accounted for 71% of the lost AI traffic. In September, blog visits from AI rose 33% while programme-page visits fell 27% and homepage visits fell 34%. Total traffic looked stable. Its mix had moved from buyers to readers.
When I looked at what ChatGPT was retrieving, I found four problems on our side:
- Placement numbers on the programme page were rendering as zero in the code the crawler read.
- Tracking-parameter and staging URLs were being indexed and cited in place of the clean programme URL.
- The brand had been renamed, and the site did not state the connection. AI engines were at risk of treating the old and new names as two entities.
- Our own listicle blogs named competitors, and ChatGPT was using those blogs to recommend them.
The first three are fixed, and the listicles have been restructured. AI answers now connect the old and new brand names. Whether leads recover is the next thing to measure; I do not have that result yet.
What I would take from this
- Measure AI search by landing page, not by total visits. Homepage and money-page visits are the leading indicator. Blog citations are not.
- Read the journeys. A handful of CRM records told me more than six months of dashboards.
- Check what the AI engine retrieves from your money pages. If the facts a buyer asks about are in an image, a PDF or broken markup, you will be cited less for buying questions.
- Expect branded search to carry part of the credit. Look for it before you conclude the channel is small.
- Give cohorts time. Some of these enrollments closed 11 weeks after the first click.
Limits of this data
This is one brand, one category and a small sample; absolute figures are withheld for confidentiality, which is why the findings are given as rough proportions. AI search is defined here by first-touch referral from an AI engine, so people who asked ChatGPT and then typed the brand name directly are not counted. Enrollments are counted in the month the lead applied, and recent months may still rise. Treat the patterns as hypotheses to test on your own data, not as benchmarks.
cd ~/blog