The Next 900

Ghost's email arrived on October 30: "Milestone achieved: 100 signups." Member #100 had joined on October 27; by the time the email came, the count was 108. The lag is in the code: Ghost checks for new milestones once after the server starts, following a random wait of zero to three days.

It took about 21 months to reach 31 members. In October alone, 78 people signed up and confirmed by email. Thank you for letting a small publication into your inbox.

Ghost credited 61 of October's 78 signups to one post, the Claude Academy guide, and recorded 66 of the 78 as "Direct". My most-visited page, a guide to free Claude and Codex limit resets, drew 36% of October's visits. It brought 2 members.

On October 13 I published an investigation showing that Claude cites the Academy post and that its citation links strip the referrer. I can't prove any single member's path. My read is that most of October's members arrived through an AI assistant.

If you run a Ghost publication, the box below is the part to copy.

The scoreboard

Measured on October 30. Members and signups come from Ghost, visits from Fathom; "October" means October 1–30.

Signups per visit divides Ghost signups by Fathom visits, two systems that count differently, so read it as a ratio rather than a measured conversion rate. Ghost's milestone counts every member, including the 18 who have unsubscribed from the newsletter.

Where the 100 came from

Ghost credits a signup to the last post the visitor read before joining. For 61 October members, that was the Academy guide; the certification guide came second with 4. The Academy post went up in August and had brought 1 signup before October. In October it had 1,134 visits: 96% arrived with no referrer, and 15 came from Google. It drew 11 visits on October 1 and has held between 28 and 59 a day since October 12.

When I ran the investigation, the post had 8 Search Console clicks in its whole life, and Claude cited it in 3 of 3 incognito test questions that named Claude Academy. Claude's citation links carry rel="noopener noreferrer", so the click arrives without a referrer and Ghost records the signup as Direct. Since that article went up, 48 more people have signed up: 37 credited to the Academy post, 40 recorded as Direct and 3 from claude.ai.

Referrer data barely registers this channel. In October, claude.ai, chatgpt.com, Gemini, Copilot and Perplexity together accounted for 72 visits by referrer or tag. Ahrefs' numbers show a similar gap at a much larger scale. In March 2026, the referral data it tracks across 74,752 websites counted about 2.7 million visitors from ChatGPT and about 102,000 from Claude. In the same month, 2,836 of Ahrefs' own new users said at onboarding that Claude sent them, against 1,978 for ChatGPT. Those are different populations. Still, they point the same way as my data: referrer counts probably undercount Claude.

The likeliest explanation is that Claude sends the Academy post most of its readers, and the post's signup ratio is more than ten times the rest of the site's. It is also a fragile channel. If Claude's search results or answer style change, the curve flattens, and my analytics won't say why.

The page Google loves converts almost nobody

The free limit resets guide had 12 visits on its first day, October 8, and 100 to 175 a day from October 20. In October it had 2,013 visits, 89% of them from Google. Search Console shows 1,685 clicks at an average position of 5.9. The page brought 2 members.

The search queries suggest why. "claude free reset" brought 96 clicks and "claude reset for free" 51. Another 63 clicks, at position 2.0, came from the text Claude shows when its list of limit resets is empty, pasted straight into Google: "no resets right now when you get one, it shows up here with its expiry date." Someone who searches for that wants to know what the message means. The page tells them, and they have no reason to stay.

The Academy reader is in a different situation. Learning a tool or preparing for a certification takes weeks, and a newsletter about those tools fits that. That's my inference, not a measurement.

Quick-answer traffic has a second problem. Using March 2025 data, Pew found that Google users clicked a traditional result in 8% of visits when an AI summary appeared, and in 15% when none did. Summaries appeared on 53% of searches with ten or more words. That pasted sentence has 16.

The page is still the biggest leak by volume. It has only the signup card at the end of the article. I suspect most readers leave before reaching it, though I haven't measured scroll depth. At the 0.4% the site manages without the Academy post, October's traffic would have brought about 8 members instead of 2.

What the March plan got wrong

In March I wrote a growth plan with one goal line: "Scale from 8 to 100+ subscribers and 1,600 to 10,000+ visits/month". It bet on Hacker News, Reddit, newsletter submissions such as TLDR and Ben's Bites, syndication to Dev.to, Hashnode and Medium, and LinkedIn and X.

Member #100 arrived on October 27, seven months later. October had 5,571 visits, not 10,000. What delivered wasn't on the list: Google search on one reference post brought most of the new visits, and, by my inference, an assistant citing another brought most of October's members.

To be fair to the plan, I worked little of it. In April, one Hacker News submission brought 26 visits, and three Dev.to cross-posts and one Ben's Bites submission went out. After that, each article got a LinkedIn and X post drafted for me to paste, and the daily distribution sessions mostly didn't happen.

Its premise was also too narrow. It treated growth as distribution: push each post to where people gather. What worked was being a complete answer where people, and their assistants, go looking. The one first-hand comparison I found isn't kind to aggregators either. Gergely Orosz reported in 2022 that about 1% of his Hacker News front-page visitors subscribed, against 15% of visitors from his blog. Volume may change that: Sean Goedecke reached the front page of Hacker News or similar aggregators 33 times in 2025, added email in May and had just over 2,500 email subscribers by the new year.

Ghost's guide, stage by stage

The milestone email links to Ghost's guide How to get your first 1,000 email subscribers. Ghost published it in September 2021 and last updated it that December. It never mentions AI assistants. It also predates most of Ghost's own growth features, and several of those are off on this site: Recommendations, growth data for Ghost Explore, the free welcome email and llms.txt.

The guide assumes your first 100 came from "your expanded circles of influence". Here, most came through one post. The principle still holds: "providing value in public in new ways and for new groups of people."

100 → 250: a lead magnet and niche communities.

• Lead magnet. Ghost's rules are that it should be something people would pay for, offer a quick win "in 10 minutes or less", and "must lead somewhere". Verdict: later, and never by gating a post an assistant cites. I'd expect a gate to hide the content from the assistant too.
• Contribute to a niche community. Verdict: sound, but slow.

250 → 500: directories and one social platform.

• Directories. Ghost's 2024 list has 26 entries, with Ghost Explore first. I haven't checked which ones still operate. Verdict: cheap. Do it once, tag every link and expect little.
• Go all-in on one social platform. This is 2021 advice written with Twitter in mind. LinkedIn sent 21 visits in October. The site's social web profile has 2 followers, and followers don't count as members anyway. Verdict: not the bet; it needs the daily routine I won't promise.

500 → 1,000: a pillar page and contests.

• A pillar page readers share. Ghost says a great pillar "becomes the thing to read about a given subject". Verdict: yes. The Academy guide works like one, and it apparently spread through an assistant, not search rankings. The share ask moves up to now.
• Contests, giveaways, referral programs. Ghost's own caveat is that some people "will only be interested in the prizes and not your content." Verdict: no.

What the guide misses

Assistants as a channel. SE Ranking measured AI platforms at 0.32% of all website traffic across 101,574 sites, against 42.75% for organic search. That figure is counted by referrer, so it misses clicks like the ones above. For a niche reference post, I think this undercounted channel can be the main one.

What gets a page cited. The original GEO paper (KDD 2024) tested ways to make a page show up more in AI-generated answers. Citing sources, adding quotations and adding statistics raised visibility by 30 to 40% on its main metric. Keyword stuffing gave "little to no improvement". A 2026 survey of 45 studies adds that those gains are "conditional on a source already being present in a fixed context". In other words, they were only shown to help once the engine had already found your page.

An October 2026 study of 158,099 answers from generative search engines looked at what decides citations. For ChatGPT, which pages its own search retrieves explained far more of the variation than the properties of the pages themselves (32% against 2.8%). It also found that "pages whose title and paragraphs match the questions people ask are cited more." In practice: be findable first, use the asker's own words in titles and paragraphs, and back claims with sources and real numbers.

What llms.txt can and can't do. Ghost's GEO feature serves an llms.txt file, a summary of the site meant for AI tools, and Markdown copies of posts. It's off by default for existing sites. The evidence for it is thin:

• Ghost says it makes content "more likely to be referenced", but gives no data.
• Google says no "AI text files" are needed to appear in its AI features.
• Chrome's Lighthouse audit treats the file as optional.
• I found no statement from Anthropic, OpenAI or Perplexity that their assistants read it.

As far as I know, the feature has never been on here, and Claude cited the Academy post anyway. I'll switch it on in Settings → Meta data, because it's one setting, and I won't credit it with any growth.

The door that does matter. This site's robots.txt blocks only Ghost's admin paths, so Claude-SearchBot, Claude-User, OAI-SearchBot and PerplexityBot can all crawl the site. Anthropic says blocking Claude-SearchBot "may reduce your site's visibility". OpenAI says sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers". If you want to opt out of AI training, block the training crawlers, such as ClaudeBot and GPTBot, by name, and leave the search bots alone.

The plan for the next 900

At October's pace of 78 members a month, 1,000 is about 11 months away. I don't trust that pace to hold: one post did most of the work, through a channel I can't measure directly. So there's no target date. Instead there are seven moves, each with its evidence, a first action and the number that will show whether it works.

1. Write more posts like the Academy guide

• Evidence: about 5.4% signups per visit there, against about 0.4% on the rest of the site. The citation research favors pages that match the questions people ask.
• First action: list the questions people ask assistants about learning and using the AI tools I cover. Write one complete, sourced answer per question, titled in the asker's words, and keep each one current.
• Number: signups per visit for each new reference post after 30 days. The target by 250 members: at least one more post bringing in members at three times the site's 1.4% or more.

2. Put the signup where the answer is

• Evidence: 2,013 visits and 2 members on the resets page. Both top posts carry only the signup card at the end.
• First action: add labeled signup cards higher up on both posts, each with an offer written for that page's reader.
• Number: members per label, and the resets page moving from about 0.1% to at least 0.4%.

3. Ask every new member how they found the site

• Evidence: 66 of October's 78 signups are Direct. In Ahrefs' onboarding survey, Claude ranks far higher than referrer data puts it. The free welcome email here exists but is switched off.
• First action: switch the welcome email on and ask: "How did you find vanja.io? One line is enough."
• Number: answers by channel (assistant, search, a person, other), always reported alongside how many people replied.

4. Test citations every month

• Evidence: Claude cited the Academy post in 3 of 3 test questions that named Claude Academy. In the citation study above, ChatGPT's retrieval mattered far more than how the page was built. Answers also vary from run to run: one 2026 study saw 15% of citation decisions change when the same question was answered again.
• First action: build a fixed list of questions my reference posts answer. Run each one twice, in fresh chats, in Claude, ChatGPT, Gemini, Copilot and Perplexity, and write down every cited link, misses included.
• Number: the runs that cite vanja.io, out of all runs, reported at each checkpoint.

5. Turn on Recommendations

• Evidence: the feature is off here. beehiiv says newsletters using its recommendations "grew 2.75x faster on average". Substack said in 2024 that recommendations and its app together drove 50% of new subscriptions. These are vendor numbers from other platforms; I found none from Ghost.
• First action: in Settings → Growth → Recommendations, add a few AI publications I actually read. Ghost shows them to new members at signup, so they have to be worth it. It also invites the ones it can notify to recommend vanja.io back.
• Number: free members on the "Recommending you" tab, and recommender domains among signup sources. A click that arrives without a referrer or ref tag will still show up as Direct.

6. Ask for the share, then make sharing one click

• Evidence: Ghost's first-100 guide calls word of mouth "the most effective and trusted form of marketing" and suggests tying share requests to occasions such as a milestone. Ghost's native share buttons can end every email with a link that opens the post's share options.
• First action: the ask below, plus that email link once I've confirmed the theme supports it.
• Number: welcome-email answers that name a person. Forwards are otherwise invisible.

Here is the one ask in this post: forward one article to one person who would want the next one.

7. Collect the cheap listings, tagged

• Evidence: the site is listed on Ghost Explore with growth-data sharing switched off, and Ghost says sharing growth metrics lets Explore "rank your site even higher". Explore tags its outbound links with ref=ghost-explore, so signups from it can be counted.
• First action: turn on growth-data sharing, submit to the directories on Ghost's list that still operate, and give each one, and every link I post on LinkedIn or X, its own ref tag.
• Number: signups per tag. I expect single digits. If that's what comes, these stay in the background.

What I won't do

• Stuff keywords. On Perplexity, the GEO paper found it did 10% worse than leaving the page alone.
• Count open rates as progress. October's 32.5% comes from a tracking pixel and is inflated by mail-privacy prefetching.
• Promise a daily distribution routine.
• Cut the publishing process to publish faster. Every article still gets sourced research and independent review.

How progress will be reported

At 250, 500 and 1,000 members, I'll publish this scoreboard again with the same definitions, plus three lines: welcome-email answers by channel, citation test results, and members per signup label. Each of the seven moves gets a verdict: kept, changed or dropped, with the number that decided it.

Unsubscribes get extra attention. October had 17, about 1.5 per 100 emails delivered, and none after October 21. The site sends one email per article, which meant 21 in October, and beehiiv reports that sending weekly correlates with the strongest growth on its platform. If unsubscribes climb again, sending less is the first thing I'll test.

The test at 250

The first checkpoint has to answer one question before any other: did a second post bring in members at three times the site's 1.4% or more, the way the Academy guide did? If it did, this is a method I can repeat. If it didn't, I had one lucky post, and the scoreboard will say so.