The setup
The question was simple: if you replace your social writing entirely with AI — from the first prompt to the final post — what actually happens to performance?
Not "does AI help you write faster?" That answer is obviously yes. The real question is whether the output connects with an audience. Whether the engagement is real. Whether the growth is sustainable.
The test: 30 days, 2 posts per day per platform, 100% AI-drafted content on X and Bluesky. Every post was prompted, generated, reviewed for character count, and queued via Buffer. No human rewrites — only rejection and regeneration if a draft was genuinely off-brief.
The tools used: Claude for drafting (via Projects for persistent brand context), Buffer for scheduling and analytics, and the platform native analytics for deeper engagement data. Here is what happened.
The numbers at 30 days
The headline number is the impressions jump. Publishing twice a day consistently — something that was genuinely unsustainable without AI — produced a significant visibility increase just from volume. Consistency matters more than most marketers give it credit for.
But the engagement rate improvement is more interesting. That number went up even as volume quadrupled. The AI content wasn't just being seen more — it was landing with the audience that saw it.
What the AI nailed
Contrarian takes outperformed everything else
The highest-performing post format by a wide margin was the contrarian take — a single observation that challenged a common assumption about AI or marketing. These posts generated 3–5x the replies of any other format.
The pattern: one provocative opening sentence, two sentences of reasoning, no hashtags, no CTA, no emojis. That's it. Claude produced these reliably when prompted with a clear angle and instructed to avoid corporate phrasing.
Key learning: The brief quality determined the post quality. Posts prompted with "write a contrarian take about AI content tools" produced generic output. Posts prompted with "write a contrarian take arguing that most AI writing tools are unnecessary because Claude's Projects feature already does 80% of what Jasper charges $49/month for" produced the actual top performers.
Consistency compounded — slowly, then quickly
Weeks one and two were unremarkable. Same impressions, similar engagement. By week three, the algorithm behaviour shifted noticeably. Posts started reaching accounts that hadn't seen the content before. Follower growth accelerated in week four.
Buffer's 2026 social engagement report found this pattern across 52 million posts: consistency drives more long-term reach than any individual post format or timing hack. The data matched that finding exactly.
Key learning: The compounding effect of consistent posting is real, but it takes three to four weeks to show up in the numbers. Most marketers quit before then.
Bluesky responded better to longer posts
X and Bluesky are different audiences with different expectations. The Bluesky account consistently outperformed on posts that ran closer to the 300-character limit — slightly more developed thoughts, one additional sentence of context.
X performed better with shorter, punchier posts. Often under 180 characters. The audience there responds to hooks, not paragraphs.
Key learning: Treating both platforms as identical is a mistake. Same angle, different format. Always write platform-native rather than cross-posting the identical text.
What flopped
List posts with arrows performed below average
The format that looks great in theory — "3 things AI can do for your email workflow → item 1 → item 2 → item 3" — consistently underperformed. Impressions were fine. Saves were below average. Replies were almost zero.
The problem: list posts feel like content rather than conversation. They don't invite a response. People read them and scroll on. They are optimised for saves, not engagement — and on a small account, saves without engagement don't move the algorithm.
Key learning: Save list formats for accounts with existing audiences. At the growth stage, prioritise replies over saves.
Question posts with no answer given landed flat
The theory: ask the audience a question, they reply, the algorithm sees engagement, the post spreads. The practice: question posts with no opinion attached generated almost no replies. People don't answer questions from accounts they don't know yet.
The posts that generated replies were the ones that stated an opinion and invited disagreement — not the ones that asked a neutral question and waited.
Key learning: Provoke, don't ask. "Is AI making content worse?" gets no replies. "AI has made most content worse, and here's the specific mechanism" gets six.
Article teasers: reach yes, followers no
Posts designed to drive traffic to blog articles got impressions but didn't convert to followers. They made sense as part of the mix — keeping the content pipeline visible — but they shouldn't be more than 20% of your posting volume at the growth stage.
Key learning: Tease articles to serve the algorithm cadence, not as the primary growth mechanic. Followers come from opinions, not links.
The one finding I didn't expect
Replying to comments on AI-generated posts drove more growth than the posts themselves.
Buffer's engagement data found the same thing across 52 million posts: creators who reply to comments outperform those who don't on every platform studied. But I didn't expect the effect to be this pronounced at small follower counts.
Three posts in week four generated follower spikes. In each case, the spike happened after a reply thread, not after the original post. Someone engaged, I responded with a developed take, others joined the thread, and the conversation drove more follows than the original content.
AI can draft your posts efficiently. It cannot replace the real-time response to a conversation. That part still requires a human, and it's still the highest-leverage activity on social media for a small account.
The honest verdict
AI-written social content works. The engagement numbers improved. The consistency was genuinely unsustainable without tooling. The quality of the output, when the brief was specific, was indistinguishable from human-written content in performance terms.
But the ceiling is lower than you might hope. AI produces good content. It produces reliable content. It does not produce the occasional exceptional post that comes from a genuine, specific observation in the moment. Those outliers — the posts that hit 10x normal engagement — came from human-written additions, not AI drafts.
The right model isn't "AI instead of human" or "human instead of AI." It's AI for volume and consistency, human attention for the replies and the real-time moments that actually move the audience.
For a practical workflow that applies this to blog production, read our piece on using AI to write a month of blog posts. For the tools that make this sustainable, see the ChatGPT prompts for social media marketing guide.