Original research
ChatGPT Referral Traffic on an llms.txt Generator
Ryan KingsFounder & CTOPublished August 2026 · 8 min read
This is a dated field note on llmstxtgen.com, a small site we built to generate llms.txt files. The public site went live in mid-June 2026. The window here runs from that launch through 17 August 2026 — about nine weeks.
The question is ordinary: once a thin llms.txt site is live, where do visits actually come from, and what do the other instruments show beside that?
Every figure names its instrument and date. This is first-party observation on one site we operate. It is not a sector census, and it is not a citation guarantee. Other measured papers live under Original research.
What we recorded
| Instrument | Window | What it counts |
|---|---|---|
| Vercel Analytics (llmstxtgen project) | mid-June – 17 August 2026 | Unique visitors, page views, named referrers |
Google Search Console (sc-domain:llmstxtgen.com) | first impressions 9 June; figures through 17 August | Search impressions, clicks, average position |
Citation ledger (citation_evidence) | first row 18 June – 17 August 2026 | Direct answer-engine observations we logged for llmstxtgen URLs |
| Vercel Analytics (AEOForged) | 10–17 August 2026 | Named referrers onto aeoforged.com after we started pointing the generator at the product |
| AEOForge generation runs | same span | Completed llms_txt jobs that hit our API |
A referrer click is not a citation-ledger row. The two sit next to each other; they are not the same claim.
Launch clock
The live site is mid-June, not May. Search Console first recorded impressions on 9 June 2026 and a first click on 13 June. Vercel first showed visitors around 16 June. The citation ledger’s first row is 18 June; the first ChatGPT cite is 23 June; the first generic ChatGPT cite is 25 June.
We only started pointing the generator at AEOForged around 10 August. The handoff numbers below are one week of that, not the whole nine weeks.
Named referrers on llmstxtgen.com
Vercel Analytics, mid-June through 17 August 2026: 245 visitors, 475 page views, bounce around 71%.
Among named referrers, chatgpt.com was the largest: 67 visitors and 87 page views. That is about 27% of all visitors, and about half of the named-referrer table. Direct and unattributed traffic is still the largest bucket overall. ChatGPT is the largest named source, not most of the traffic.
| Referrer | Visitors |
|---|---|
| chatgpt.com | 67 |
| google.com | 17 |
| vercel.com | 7 |
| baidu.com | 6 |
| youtube.com | 5 |
| msn.com | 4 |
| claude.ai | 3 |
| bing.com | 2 |
| perplexity.ai | 2 |
ChatGPT sent about 3.9× the visitors Google did in this window. That is a referrer comparison. It is not a claim that Google cannot see the site.
Country split in the same Vercel view: United States 73, India 56, China 21, United Kingdom 18. Vietnam showed 3 visitors and 60 page views — that pattern looks more like a crawl than a handful of people reading.
An earlier lifetime page mix on the same project: homepage 167 views, the generator 68, llms-full.txt explained 32, how-to-write-llms-txt 19.
The ChatGPT percentage in Vercel’s Referrers UI is a share of named referrers. It is easy to read that chart as “half of all visits came from ChatGPT.” That is not what the 245-visitor total says.
Search Console
Property sc-domain:llmstxtgen.com, figures through 17 August 2026 (live volume from 9–16 June onward):
2,985 impressions, 13 clicks, 0.44% CTR, average position 27.4.
Desktop: 2,827 impressions / 10 clicks. Mobile: 154 / 3.
| Page | Clicks | Impressions | CTR | Avg. position |
|---|---|---|---|---|
/llms-full-txt-explained | 7 | 288 | 2.4% | 25.6 |
/how-to-write-llms-txt | 3 | 1,433 | 0.21% | 14.1 |
/ | 2 | 1,213 | 0.16% | 43.9 |
The query llms.txt generator had 380 impressions, 0 clicks, average position 48.5. Comparison queries that did click sat closer to positions 7–14 (llms.txt vs llms-full.txt, and similar “difference between” wording).
URL Inspection on 17 August: the homepage, /how-to-write-llms-txt, and /llms-full-txt-explained were submitted and indexed. The homepage’s last crawl in that check was 16 August. Referring URLs for the homepage included aeoforged.com.
Search is present. It is not the main named referrer in this window.
What the citation ledger shows
We logged 582 citation-evidence rows for llmstxtgen URLs between 18 June and 17 August 2026 (575 via API, 7 manual). 199 of those URLs carried ?utm_source=openai — ChatGPT appends that parameter when it links out — and 107 of those 199 were cited.
Most ChatGPT rows in this ledger are branded self-queries (the prompt already names the site). A blended ChatGPT citation rate would mostly describe those prompts. We do not headline that blend.
On generic prompts — the query does not contain “llmstxtgen”:
| Engine | Generic checks | Cited |
|---|---|---|
| ChatGPT | 14 | 14 |
| Perplexity | 16 | 16 |
| Gemini | 17 | 0 |
| Google AI Overviews | 7 | 0 |
The ChatGPT generic wins are one wording, first seen 25 June 2026: benefits of llms-full.txt for content-heavy sites, landing on /llms-full-txt-explained?utm_source=openai. Perplexity’s 16 cited rows span four wordings. Distinct queries in the whole ledger: 23 (12 branded, 11 generic). Only a handful of generic-and-cited rows are usable.
That sits beside a wider observation in ChatGPT often answers without sources: in a July AEO-category panel, ChatGPT returned a source list on 25 of 96 answers. A source list and a named referrer are still different events.
The live first-party board for aeoforged.com is a separate instrument: live AI citation rates.
One week of handoff to AEOForged
We started pointing the generator at AEOForged around 10 August 2026.
Vercel Analytics on the AEOForged project, 10–17 August: 31 visitors / 359 page views. Named referrers: llmstxtgen.com, 4 visitors / 4 views, all landing on /. Next named sources were vercel.com (3) and accounts.google.com (2). Direct and empty referrer is the large remaining bucket — a lot of that is operator use.
Four referrals in a week is small. It was the largest named referrer in that week (4 of 31 unique visitors, about 13%). Mid-June through 17 August still shows exactly those four llmstxtgen referrals. The handoff is this week’s work, not a nine-week funnel.
Generations we can see
AEOForge recorded 19 completed generation runs and 4 timeouts in this span. Twelve completes were on eight distinct third-party sites; four completes were aeoforged.com itself. Five lead rows are tagged as coming from the generator.
The ledger is empty from 17 June through 10 August — about eight weeks with no recorded run. Completes then cluster 11–14 August (peak 12 August: seven), a hole on 15–16 August, and two more on 17 August.
Treat 19 as a floor. A generate click that never reached our API does not appear here. We are not publishing customer emails or a named client list.
What this note is not
It is not “all traffic comes from ChatGPT.” Direct and unattributed visits are still the largest bucket.
It is not a ChatGPT citation rate for the category. Branded self-queries dominate the ledger.
It is not “Google cannot see the site.” Search Console shows impressions; they convert poorly at the positions we hold.
It is not a claim that other sites will see the same mix. This is one proof-of-concept we operate.
It is not most of AEOForged’s traffic. Four named referrals in the first week of a push is a start, not a majority.
If you want the format itself, start with what llms.txt is and the spec at llmstxt.org. The generator is at llmstxtgen.com/generator. Packages that use the same measurement stack are on Pricing. More measured papers: Original research.