Claude Cites My Blog. My Analytics Call It Direct.
One old post brought in 24 of my 30 new subscribers in two weeks, every one labeled Direct, while Google sent almost nothing. Claude cites it, through links that drop the referrer. The evidence, a check you can run, and what it can't prove.
On October 1, a six-week-old post on this site started behaving like a launch. Claude Academy: 289 Resources, 22 Courses, and What Completion Certifies had drawn one to three visits a day through September. In the first thirteen days of October it drew 322, about a quarter of all visits to the site, and the site went from 31 members to 61. Ghost credited 24 of the 30 new signups to that one post. Before October it had produced exactly one.
Ghost lists the source of all 24 as Direct, and Fathom recorded no referrer on 304 of the 322 visits. The best explanation I found is that Claude keeps citing the post when people ask it about Claude Academy, through links built so that the click carries no referrer. I can show that route works. I cannot link any of those signups to a Claude conversation, and the difference is most of this article.
| 🧭 | Start here: is an AI assistant sending you readers? 1. Find the spike. In Fathom, click the post in Pages and read the Referrers panel; referrer-less visits appear as Direct / Unknown. In other tools, look for that page's direct or no-referrer visits. A sustained jump on one page, from many countries, is worth investigating. That is a rule of thumb, not a standard. 2. Check the signups. In Ghost Admin, open the post's analytics and choose Growth. "Top sources" lists where its new members came from. A member can be credited to your post while their source is Direct. 3. Check Google Search. In Search Console open Performance → Search results, add a Page filter with the post's URL and compare the same days. Google's AI Overviews and AI Mode are part of that data. Near-zero reported clicks over the same dates make Google Search an unlikely explanation for the spike. 4. Ask what your readers would ask. Open claude.ai, start an incognito chat and ask in your readers' words, without telling Claude to search. Add one broader control question, then repeat both in a ChatGPT temporary chat. Write down the date, the question and every cited link. 5. Inspect the citation. If your page appears, right-click the citation and choose Inspect. 6. Ask the readers. Put one question in your welcome email: "How did you find this site?" A reply gives you the reader's own account of how they found you. |
What Direct means
Ghost decides a member's source at signup. Its attribution script keeps a 24-hour history of the pages the visitor viewed on the site, and the server looks through that history for any entry that arrived with a referrer or a ref, source or UTM value. If there is none, the code falls back to a fixed label, with the comment "Fall back to Direct if no matches in any history entry".
So Direct does not mean someone typed the address or opened a bookmark. It means Ghost found no referrer or campaign value in the visitor's recorded pageview history from the previous 24 hours. Apps that hand links to a browser, email clients, chat apps and any link marked noreferrer can produce it. Ghost can still name the article a member signed up on. It just has no evidence of what brought them to it.
The October signups were unusual in a very specific way:
The last row divides Ghost's signups by Fathom's visits, two systems with their own counting, so read it as a ratio rather than a measured conversion rate. Ghost signups are double opt-in: each of the 24 clicked a confirmation email. 22 are still subscribed, none has bounced or complained, and Ghost records 42 opens across the 90 newsletters sent to them. Opens can be automated, so I count that as a weak signal, not proof of engaged readers.
Not Google
Search Console shows 8 clicks to the post since it went up on August 20, one of them in October; its data runs to October 12. Daily impressions in October were lower than in September. The post ranks around position 25 for "claude academy" and around 5 or 6 for certificate questions. Discover and Google News show zero.
Google's AI features in Search live inside the same numbers. Its help page for the generative AI performance report says the report "includes data from the Web search type in the Performance report". Google's search results and its AI features together sent this post almost nothing.
From October 1 the post drew 10 to 51 visits a day, from more than 60 countries, at every hour, about 63% of them on desktop: Windows with Chrome first, then iPhone Safari, Mac Chrome and Android Chrome. The October 1 increase in referrer-less traffic was confined to this page. Elsewhere on the site it stayed flat and rose only when new posts went out a week later.
Asking Claude
I started with Anthropic's web search tool, the one Claude Code uses. It returned the post first for "Claude Academy courses list" and "is Claude Academy worth it", second for a question about what completion certifies, and sixth for plain "Claude Academy". Google has it near position 25 for that last query.
Then I asked Claude the way a reader would, in incognito chats on claude.ai with the default model, Opus 5.5, on October 13:
- "What is Claude Academy, and is the certificate worth anything? Please search the web." 5 links; 3 were this post, including the first.
- "I heard about the Claude Frontier Academy. Can I join it as an individual, or should I do Claude Academy instead? What would I actually get?" Searched without being asked; 2 links, one of them this post.
- "Is the Claude Academy certificate worth putting on my LinkedIn?" Searched without being asked; 4 links, this post first.
- "How do I get a Claude certification? Which free courses should I take first?" 8 links, none of them this post.
Three links to one article are one source, not three. Still, the post won every question that named Claude Academy and lost the broader certification question. That is a small sample, and answers change from run to run.
The answers also borrowed the post's framing. One said: "One reviewer described it as a certificate that only shows you attended." That echoed the post's original headline, which ended "a Certificate That Certifies Attendance" until I retitled it on September 20. Anthropic's search results still show the old title.
Why nobody seems to have sent them
Every citation link in those answers was an anchor with target="_blank" and rel="noopener noreferrer". MDN describes what the second keyword does: it "instructs the browser, when navigating to the target resource, to omit the Referer header and otherwise leak no referrer information".
claude.ai's own pages use the browser's default policy, strict-origin-when-cross-origin, which would normally send https://claude.ai/ as the referrer. The attribute on the link overrides it, and the citation URLs carry no UTM parameter. A reader who clicks one arrives exactly like someone who typed the address.
Analytics tools now have AI reports. Fathom has an AI tab beside Referrers, and for the 30 days to October 13 it credits ChatGPT with 16 visitors across the site, Gemini with 11, Claude with 6, Copilot with 3 and Perplexity with 2. Filter it to the Claude Academy post and Claude has 1 of the page's 344 visitors. The report can only count the clicks that say where they came from.
Some do. Ghost recorded 4 visits to this post in October with claude.ai as the source, so some links or surfaces pass it. The citation links I inspected did not. I did not test the Claude desktop and mobile apps.
The other assistants
I asked ChatGPT the same first question in a temporary chat. It cited academy.claude.com and an article on claude.com, not this post. Its two links also carried rel="noopener noreferrer" and no utm_source parameter, so ChatGPT clicks like those can land in Direct too. Some ChatGPT clicks do announce themselves: Ghost counted 9 visits from chatgpt.com across the site in October.
Gemini returned an error twice. Perplexity and Copilot need a signed-in account to answer, so I did not test them. The Perplexity clicks that did reach this post carried utm_source=perplexity, and there were only a few. Claude is the only assistant I saw cite the post. That is not the same as the only assistant that does.
What I can't prove
An independent review by Codex, the second provider in my two-provider setup, raised the key limitation: these tests show that the post can be found and that the clicks lose their referrer, not the journeys of the 24 people who subscribed. The tests ran on October 13, after the surge began, and no signup is linked to a conversation. Private shares, course directories and other assistants can produce Direct too.
The start date is unexplained as well. On October 2 Anthropic announced the Claude Frontier Academy. With a $100 million commitment, Anthropic says it "aims to train 10,000 Frontier Deployed Engineers (FDEs) by the end of 2027", and participation is by nomination. When I asked Claude about it, Claude sent individuals to Claude Academy and cited this post, so the news plausibly added readers. It cannot explain a step that began a day earlier. Anthropic sent me no email about the Academy, and referrer-less traffic elsewhere on the site did not step up. The stale title in Anthropic's results says nothing certain about when the page was last crawled.
One more piece of context: Anthropic publishes an official academy-guide skill that tells Claude to close answers about using Claude with a Claude Academy course when the match is strong. It links only to academy.claude.com, and I did not find it in the claude.ai skills directory. It may raise interest in the Academy, but I have no measurement that ties it to this post.
Why these readers subscribe
This part is my inference. My hypothesis is that readers arriving from an assistant have already asked one narrow question: is this certificate worth anything, or which course should I take first? The post answers it with the full course catalogue and a verdict, and the subscribe box at the end offers more of the same. The 7.5% ratio fits that explanation without proving it.
What I would tell any publisher
- Treat a one-post Direct spike as a question. It may be loyal readers. It may be an assistant that keeps citing you. Ask the assistants before you decide.
- Check Search Console before crediting Google Search. SEO will not restore a missing referrer.
- Inspect the link. When the citation carries
noreferrer, your analytics cannot recover the omitted referrer from that click. There is no UTM to add to someone else's link. - Record the misses. Keep the controls and the questions that did not cite you. One favorable answer shows a possible route, not how often anyone takes it.
- Ask the readers. A reply gives you the reader's own account of how they found you.
- Publish the reference answer, then maintain it. The post went up on launch day with an exact course catalogue. Keep its numbers current, but do not assume assistants will pick up edits immediately: Claude still echoed my old headline three weeks after I changed it.
Analytics were built for links that say where they came from. Claude's citation links don't, so Claude can send readers without leaving a referrer. How many of these 24 subscribers arrived that way remains unknown.
