Unit economics

The math of a conversation: cost and revenue per session

30 June 2026 · 8 min read · Monetzly team

Monthly aggregates hide the thing that decides whether an AI app works: the economics of one conversation. If a session costs more to serve than it earns, no amount of growth fixes it — growth makes it worse.

This is the model we use, with every assumption stated so you can substitute your own numbers.

The cost side

Cost per session is turns × tokens per turn × price per token, plus whatever you spend on retrieval, tools and hosting. Two details dominate.

First, context grows. Turn eight of a conversation re-sends turns one through seven, so cost per turn rises through a session rather than staying flat. An 8-turn conversation costs meaningfully more than twice a 4-turn one.

Second, model choice swamps everything else. Frontier-tier pricing versus small-model pricing is an order-of-magnitude difference on the same conversation. Routing simple turns to a cheap model is usually the single largest cost lever available.

Lever	Effect on cost per session	Effect on quality
Route easy turns to a smaller model	Large reduction	Neutral if routing is accurate
Trim or summarize old context	Grows with session length	Risk of losing detail
Cache system prompts and retrieved docs	Moderate reduction	None
Cap response length	Small reduction	Noticeable if capped too tightly

The revenue side

For an ad-supported session, revenue is: sessions with a qualifying match × placements per matched session × effective rate per placement. Three numbers, and only one of them is about ad pricing.

Match rate — the share of sessions where the conversation expresses intent a campaign can legitimately serve. Driven by your app's subject matter more than by anything you configure.
Placements per matched session — deliberately low, because the format degrades quickly with repetition.
Rate per placement — set by advertiser demand for that intent. Commercial categories price far above general chat.

Notice that a general-purpose assistant and a shopping assistant with identical traffic can differ several-fold in revenue, entirely through match rate and category value. Vertical apps with commercial intent are the strong case for this model; a pure coding-help tool is a weaker one, and it is worth knowing that before you build a plan around it.

The number to watch

Contribution per session — revenue per session minus cost per session — is the metric worth putting on a dashboard. If it is positive, growth compounds. If it is negative, growth is a countdown.

Track it split by cohort. Long sessions cost more but also carry more intent; short sessions are cheap but rarely match. Those two curves rarely move together, and the split tells you where to spend engineering time.

One caveat: attribution lags. Clicks and conversions land after the session ends, so a same-day read of revenue per session will understate it. Compare complete weeks, not hours.