Jevtown

The Creator’s New Pre-Flight Check: What a Simulated Town of 10,000 Readers Says About Your Post Before You Hit Publish

Every social media operator I know has a private ritual before hitting publish on a high-stakes post: they squint at the draft, imagine their audience, and guess. Will this land with the LinkedIn crowd? Will TikTok scroll past in the first second? Will the Instagram caption feel salesy? That guesswork is the most expensive part of the job, because a failed launch costs more than time — it costs algorithmic momentum. Jevtown, a new experiment from maker Ivan Gabor, tries to replace that guess with something stranger and more interesting: a simulated town of 10,000 residents who read your text first and tell you whether it travels or dies. It is not a scheduling tool, not an analytics dashboard, and not another AI caption writer. It is a stress test for copy — and for creators who ship dozens of posts a week, that framing is worth taking seriously even if the execution still has gaps.

What Jevtown Actually Does (and Why It Isn’t Just Another AI Gimmick)

The pitch, as Gabor describes it on the Product Hunt launch page, is deliberately odd: “Most Jev demos ask the model for one decision… I wanted to see what ten thousand decisions about the same text look like, so I built a town and put Jev in every house.” You paste in a post, a listing, a product description, or a headline. The town’s 600 “first wave” residents read it. If enough of them are glad rather than annoyed, the text travels further, eventually reaching all 10,000. A weak text dies in the first wave for about half a cent. A good one reaches the full town in roughly 14 seconds for about ten cents, according to the maker’s own numbers.

That cost-per-test math is the part that should make a social media manager sit up. In my own experience running paid A/B tests on ad copy, even a cheap Facebook test costs dollars per variant and takes hours to accumulate signal. A ten-cent, fourteen-second read on whether a headline will resonate is a different category of tool — closer to a spell-checker for persuasion than a campaign simulator. Gabor’s own demo is the clearest illustration: he wrote one iPhone listing two ways. The version offering payment on inspection reached 2,100 residents and prompted 142 to write to the seller. The advance-payment-only rewrite stalled at 600, with 204 residents suspecting a scam. That is not a reach prediction — it is a suspicion detector, and for anyone writing product copy, cold DMs, or landing pages, suspicion is the silent killer.

The Three Calibration Findings That Matter More Than the Demo

Buried in the launch post are three empirical notes from Gabor’s testing that I think are more valuable than the product itself, because they apply to any AI-assisted content workflow:

1. Batching doesn’t hurt accuracy. 200 personas in one request answered the same as one asked alone. If you are building any AI evaluation pipeline — for comment moderation, ad variant scoring, or content grading — this suggests you can scale without degrading signal.
2. Order effects are real and large. Reversing the order of options shifted answers by 0.062, which Gabor says is two and a half times the noise between two identical calls. His fix: fix the order and never shuffle. Anyone who has run survey-based creative testing knows this pain; seeing it quantified for LLM-based evaluation is useful.
3. Question framing dominates calibration. Asking “what is the highest price this buyer would pay” turned 90% of simulated people into buyers. Writing the base rate into the question gave 48%, which matched what they actually did elsewhere in the same run. Gabor’s conclusion: “The calibration is real, but it calibrates the question you wrote.” That is a warning label for every AI research tool, not just this one.

How It Compares to the Tools You Already Use

If you are a social media operator, your stack probably includes a scheduler like Buffer or Later, an analytics layer like Metricool, a design tool like Canva, and maybe an AI writing assistant baked into ChatGPT or Jasper. Jevtown does not replace any of them. It sits upstream of all of them, at the moment before you commit a draft to the calendar.

The closest incumbent comparison is not a scheduling tool at all — it is the manual “send it to a friend and ask what they think” workflow, or the more sophisticated version where you post to a private Facebook group or a Slack community and watch reactions. Jevtown compresses that loop from hours to seconds and from a handful of readers to ten thousand simulated ones. The trade-off is obvious: simulated readers are not real readers. Gabor is transparent about this. When asked whether he compared AI reactions to real human reactions, he answered plainly: “No, not against real audiences.” He checked two proxy validations instead — a gardener persona test where 93% stopped at a tomato seedling post versus 15% for programmers, and a rule that separated six weak texts (including spam and a scam listing) from six normal ones. That is not the same as predictive validity against your actual Instagram followers, and he says so.

Why TikTok Creators Should Care More Than LinkedIn Ones

The platform where this kind of pre-testing has the highest potential value is TikTok, and the reason is mechanical. TikTok’s distribution model is famously brutal on the first wave: your video is shown to a small seed audience, and if watch time, rewatches, and shares don’t clear a threshold, distribution stops. That is structurally identical to Jevtown’s “first wave of 600 residents” gate. A tool that tells you whether your hook survives the first wave is directly analogous to TikTok’s own algorithm logic.

LinkedIn is a different beast. Its feed rewards dwell time, comments, and network effects more than raw stop-rate, and the audience is narrower and more predictable. A simulated town of 10,000 generalists is less useful there because your LinkedIn audience is probably 500 specific people in your industry, not a random cross-section. Gabor acknowledges this limitation: the town is not configurable on the site. Everyone posts to the same 10,000 residents, about 800 of whom are into startups. You can filter reactions by interest after the run, but you cannot build a town of only tech founders unless you self-host. The code is open under MIT, and the personas live in a single file, public/shared/personas.js, where you can edit jobs and interests. For a creator with a niche audience, that self-hosting path is the only one that makes the simulation meaningfully representative.

What Creators and Social Teams Can Borrow From This

Even if you never open Jevtown, the underlying workflow is worth stealing. Here is what I would take from it:

Pre-test the hook, not the whole post. The most valuable output is not a reach number — it is the distribution of reactions. Jevtown returns probabilities for scrolling past, reading, liking, reposting, following, and blocking. That breakdown is more actionable than a single score. If 40% of simulated readers would block, your hook is not weak — it is repellent, and no amount of caption polishing will fix it.

Use buyer questions as a gap detector. In the iPhone listing demo, the tool showed what buyers would ask first: 24% wanted to know if the price was negotiable in the inspection version, while 43% in the prepayment version asked for a safe deal or cash on delivery. That is a content brief for the next version. Gabor notes the tool does not suggest rewrites — “Jev only answers questions and writes no text” — which I actually prefer. Rewrite suggestions tend to flatten voice. Question distributions tell you what is missing without telling you how to sound.

Calibrate your own intuition against a baseline. The /me feature lets you describe a resident (ideally yourself), answer what they would do with 12 posts, then see how Jev’s answers compare. Gabor frames it as a way to check whether the simulation matches one real person. For a social media manager, that is a low-stakes way to audit whether your mental model of your audience is accurate or just familiar. I would use it as a calibration exercise before trusting any single test result.

Where the Math Breaks

The honest limitation is that simulated engagement is not engagement. A resident “liking” a post in Jevtown costs nothing and carries no social risk. A real person liking your post risks their boss seeing it, their ex seeing it, or their feed filling up with similar content. The simulation cannot model social cost, and social cost is a huge driver of real engagement behavior — especially on LinkedIn and X, where professional identity is on the line. Gabor’s own framing is careful: “These are still simulated readers, so the numbers are best used to compare two versions of the same text.” That is the correct use case. Do not read the absolute reach number as a forecast. Read the delta between version A and version B as a signal.

The Verdict: Promising Instrument, Unproven Calibration

My take: Jevtown is the most interesting content pre-testing concept I have seen since Persado started pitching emotional language optimization to enterprise marketers, but it is earlier and rougher. The no-sign-in, no-account, bilingual (Ukrainian and English) approach is refreshing — as one commenter noted, “Finally a tool that does not make me connect all my socials just to try it.” The MIT-licensed code and single-file persona config make it hackable for teams who want a niche town. The speed and cost are genuinely novel.

But the validation is thin. Gabor has not benchmarked against real audiences. The persona set is fixed and generalist. There is no reading-time measurement, no sentiment trajectory, no platform-specific modeling. And the tool is silent on the biggest question for social operators: does a text that “travels” in Jevtown actually perform better on Instagram, TikTok, or LinkedIn? Until someone runs that correlation study, this is a drafting aid, not a decision engine.

Who is it not for? Anyone who needs statistically defensible audience research. Anyone whose audience is highly specific and not represented in the default town. Anyone who wants rewrite suggestions — you will not get them here. And anyone who treats a ten-cent simulation as a substitute for actually publishing and reading the comments.

What I’d Watch / Test Next

If you want to put Jevtown through its paces this week, here is what I would do:

1. Run a head-to-head on a real draft. Take two versions of a post you are genuinely unsure about — same content, different hook. Run both through Jevtown and note not just which reaches further, but which reaction distribution looks healthier. Then publish the winner and compare the first-hour engagement to your last ten posts. You are looking for directional agreement, not precision.
2. Try the /me calibration. Go to jevtown.ivanhabor.com/me, describe yourself as a resident, and answer the 12-post questionnaire honestly. If Jev’s predictions for your own behavior are way off, the town is not modeling your audience well either. If they are close, you have a baseline for trusting the deltas.
3. Self-host a niche town if you have a developer. Clone the repo, edit public/shared/personas.js to reflect your actual follower demographics, and run your next ten posts through it. That is the only version of this tool that could plausibly earn a place in a serious content workflow.
4. Watch for a real-audience correlation study. The maker has openly invited feedback on cases where the town got it wrong. If a creator with a decent-sized account publishes a side-by-side of Jevtown predictions versus actual platform analytics, that is the data point that would move this from curiosity to tool.

Until then, treat Jevtown as a structured way to argue with yourself before you publish — which, honestly, is more than most social media teams do.