Share of model: the distribution metric AI just invented

There's a new question that should terrify and excite anyone who sells software: when a buyer asks ChatGPT or Perplexity "what's the best tool for X," does your product come up?

Ask it for your own category right now. Actually do it. If you're in the answer, congratulations, you have distribution you're not paying for. If you're not, one of your competitors does, and neither of you can see the traffic in any dashboard.

We started calling the underlying metric share of model: of the times an AI assistant answers a question in your category, what fraction of answers include you? It's share of voice, except the "voice" is a model's memory, and the audience never visits a website you can instrument. The broader practice is getting a name too, GEO, generative engine optimization, as the successor habit to SEO.

Where model recommendations actually come from

Models don't invent opinions about products. They compress what people wrote. And what people wrote, for software categories, is disproportionately: Reddit threads, Hacker News discussions, Stack-adjacent forums, comparison posts, and documentation. When engines do retrieval (Perplexity always, ChatGPT often), the live citations skew even harder toward community threads, because that's what ranks for "best X for Y" queries after years of review-site distrust.

This has a strange implication that took a while to sink in for me. The same thread is now two assets. A "what do you all use for social listening" thread on Reddit is, today, a live buyer you could win this week. And for the next several years, it's a training document and a citation source that shapes what every model recommends to every buyer who never posts at all.

Active return this week. Passive return for years. One thread.

Why GEO isn't "SEO but for robots"

SEO taught everyone to build content on domains they control. GEO mostly doesn't work that way, because models discount vendor self-description almost as aggressively as human Redditors do. What carries weight is what other people say about you in the places models read and cite.

Which means the levers look less like publishing and more like participation:

• Be present in the threads that get cited. When a category question is being answered in public, a helpful, disclosed reply from you becomes part of the permanent record models learn from and point to.
• Win the comparisons that already exist. Threads comparing you to competitors get cited constantly. If the top reply is three years stale and wrong about your pricing, that's what the model repeats.
• Feed the specific pains you solve. Models connect "problem language" to "product names" from co-occurrence in real discussions. If nobody ever describes the problem next to your name in public, the association doesn't exist.

• Measure it like a channel. Run your category's buying questions through the major engines on a schedule. Track who appears, who gets cited, which threads the citations point at. Those threads are your target list.

The honest caveats

Nobody, including the people selling GEO dashboards, fully knows the weighting functions. Model refreshes move rankings unpredictably. Anyone promising "we'll get you into ChatGPT's answers in 30 days" is selling weather control.

And there's a line: seeding fake organic mentions is astroturfing, models and platforms are both getting better at detecting it, and getting caught poisons the well permanently. The durable strategy is identical to the durable Reddit strategy, which is convenient: show up in real conversations, be genuinely useful, disclose who you are. The gray-hat shortcuts get lasered eventually. The helpful reply from 2024 keeps getting cited in 2029.

Why we care

Half of ThreadSonar's job description is finding you exactly these threads while they're alive: the category questions, the comparison discussions, the pain descriptions. We built the product for the this-week return (a buyer you can answer today) and kept running into the multi-year return (that answer becoming part of what machines recommend). We track them as one motion now. Every thread does two jobs, and the companies that internalize that earliest are going to look, in three years, like they got impossibly lucky with AI recommendations.

They won't have been lucky. They'll have been present.