1. How It Started: Can AI Predict What Goes Viral? 🤔

Michael Taylor set out to answer the question "Can AI predict what will go viral online?" — running an experiment where 1,903 AI personas, built from real Hacker News users' comments, were asked to identify which headlines would become popular.

"They got it right 60% of the time — 20% better than flipping a coin."

This result is more than a fluke. Real-world marketing executives say they'd use AI for market research even if it only agreed with humans 70% of the time. 60% isn't perfect, but it's close enough to be genuinely useful.

That said, the experiment also revealed why it's so hard for AI to do better.

2. The Experiment: Building a Hacker News Simulator 🧑‍💻

• 1,147 headlines were collected for a single day, with popular and unpopular posts mixed 50:50.
• 1,903 AI personas (built from real Hacker News users' comments) evaluated each headline and decided whether to "upvote" it.
• The entire process was automated at scale using a Jupyter Notebook.

"If AI could reliably predict viral headlines, you could keep testing until you hit the jackpot. AI lets you do this much faster and cheaper than traditional market research."

Marketers remain skeptical that a machine can replace a real focus group — which is exactly why Taylor set out to prove it himself.

3. Results: Not Perfect, But Useful! 📊

• Accuracy: 60%
  • AI personas correctly identified which headlines would go viral 60% of the time.
  • Clearly better than a coin flip (50%), but far from perfect.

"When I looked closely at the headlines I got wrong, I understood why this is such a hard problem."

For example, the headline "Gemma 3: Google's New Multimodal Model" was predicted to be a hit by the AI — but only got 4 upvotes in reality. Meanwhile, "Gemma 3 Technical Report [pdf]," covering the same topic, received 1,324 upvotes.

"Is the second headline really 1,000 times better than the first?"

Similarly, "TSA Finds Live Turtle Hidden in Man's Pants" was predicted to be popular, but completely flopped.

4. Why Did AI Get It Wrong? Virality Is About Social Dynamics 🌀

Analyzing the failures revealed that the core of virality is luck and social momentum.

• AI personas evaluated headlines without knowing how others reacted, but
• Real Hacker News users are influenced by upvote counts, page position, and the day's trending topics.

"A single early upvote can change everything — the same content can take completely different paths in parallel universes."

Princeton University research illustrates this clearly. 14,341 people were given the same song list; some groups could see others' choices. The same song was a massive hit in some groups and a complete flop in others. 70–80% of success depended on early luck.

"This 'rich get richer' phenomenon is exactly why simulators struggle."

In other words: even if you perfectly model individual taste, you can't fully predict virality if you miss the social physics — early reactions, competition, and cascade effects.

5. Practical Insights: How to Use 60% Accuracy 💡

The key question isn't "is it perfect?" but "does it help you make better decisions?"

• 60% accuracy isn't enough to reliably hit a home run on the first try, but it's more than sufficient to filter out obvious failures and identify promising directions among many ideas.

"AI market research firms claim 90%+ accuracy, but they're predicting what people say they'll like — not what actually goes viral."

The bigger lesson from this experiment: Even if AI perfectly models individual preferences, viral success still depends on unpredictable social dynamics.

Practical Tips

• Use it for iteration, not prediction.
  • Rather than betting everything on one home run, test 10 versions with AI, filter out the obvious failures, and then actually run experiments with the promising candidates.
• Run multiple simulations.
  • Succeeding once out of eight tries might be luck, but six out of eight suggests genuine potential.
• Focus on relative ranking, not absolute values.
  • AI is stronger at distinguishing "clearly good" from "clearly bad" than at predicting exact outcomes.

6. Try It Yourself: A Prompt for Cloning a Hacker News User 🧑‍🔬

Clone a Hacker News user in ChatGPT or Claude!

1. Copy the comments from a Hacker News user's public comment page.
2. Paste the following prompt:

"You create a detailed persona representing a Hacker News user based on their list of comments. Using only the comments this user has actually made, create a unique persona who would respond the same way. Infer their background, experience, interests, and history as richly as possible and describe it in one paragraph. Make demographic information realistic and credible. The Hacker News user id is {user_id}. Build the profile using only this user's comments."

This produces a fictional character that reflects that user's tendencies. Ask this persona about your ideas and you have a virtual focus group you can use instead of traditional market research!

"AI may invent names or specific details, but in practice these fall within a reasonable range inferred from the actual comments."

7. Conclusion: The Future and Limits of AI Market Research 🚦

• AI makes market research insights — once exclusive to large corporations — accessible to everyone.
• But perfect prediction is nearly impossible because of the chaos of social dynamics.
• AI is best used not as a marketing "crystal ball," but as a tool for smarter experimentation and iteration.

"AI is not yet a magic eight ball for your marketing content — at least, not yet."

Key Keyword Summary

• AI market research
• Hacker News simulator
• AI personas
• 60% accuracy
• Social dynamics
• Early upvotes / luck
• Iteration
• Relative ranking
• Prompt engineering
• Virtual focus group

Michael Taylor

• CEO of Rally (virtual audience simulator)
• Co-author of Prompt Engineering for Generative AI

I hope this summary helps you understand the real promise and limits of AI-powered market research! Feel free to ask any questions 😊