TAM, SAM, SOM With AI: Market Sizing Without a Consultant

Market sizing has a reputation problem on both ends. Founders over-trust big numbers — "$50B market" on a slide, sourced from a report nobody opened. Investors under-trust all of them, correctly assuming most bottoms-up math is a wish wearing arithmetic. The honest version is smaller and more useful: a triangulated estimate with visible assumptions, where you can say which inputs are solid and which are guesses. An AI skill can now produce that version in an afternoon. It cannot make your assumptions true, so this post spends as much on where estimates break as on how to build them.

What the three numbers actually mean

TAM is the total addressable market — everyone who could conceivably buy the category. SAM is the slice your product and geography can actually serve. SOM is the share you can realistically capture in a given horizon, constrained by competition, distribution, and capacity. The common failure is presenting only TAM, which answers a question no investor is asking: how big is the pond matters less than can you catch fish in it.

The skill and its three methods

biz-market-sizing estimates all three numbers using three methods, then triangulates. Each method fails differently, which is exactly why you run all three.

Top-down starts from an industry report or category estimate and narrows it: TAM from the category figure, then apply filters for your segment, geography, and price point to reach SAM, then a reasoned capture rate for SOM. Its weakness is inherited: you inherit the source report's definitions, and category boundaries are drawn by people selling reports. Treat top-down as a sanity ceiling, not a foundation.

Bottom-up starts from units you can count: number of potential customer accounts (or developers, or stores — whatever your buyer unit is) times realistic annual price times plausible penetration. Its weakness is enthusiasm — founders reliably over-count the reachable accounts and under-price the sales effort required per account. Push the inputs until they sting a little.

Analogous markets borrow structure: find a market that already went through your adoption curve — a neighboring category, an earlier geography, a comparable tool cycle — and scale its realized numbers to your differences. Its weakness is the differences themselves; two markets that look identical on paper can diverge on regulation or habit. State the analogy's breaking points explicitly in the output.

Triangulation and sensitivity scenarios

The skill's real contribution is putting the three estimates side by side and marking the disagreements. Top-down says one number, bottom-up says half of it — now the interesting question is which input caused the gap, and usually the gap itself is the most informative result in the whole exercise.

Sensitivity scenarios take this further: vary the two or three inputs your estimate actually hangs on (price point, penetration rate, segment growth) across a pessimistic, base, and optimistic case, and report the range rather than the point. A SOM that survives the pessimistic case is a plan; one that only exists in the optimistic case is a hope. Honest outputs label which case each conclusion depends on.

Where estimates go wrong (evidence grades)

The skill grades evidence behind each input, and the grades tell you where the estimate is load-bearing versus decorative. Public pricing pages and counted account registries are strong evidence. Analyst report figures for adjacent categories are medium — definitional drift is common. Founder intuitions and vendor blog claims are weak, and they are also, inconveniently, where most of the optimism lives.

The systematic errors to watch: using TAM where SAM belongs on the slide (the classic sin), assuming today's growth rate persists (it typically doesn't), ignoring that SOM is bounded by sales capacity rather than market willingness, and double-counting — the same customer appearing in two segments. No tool removes these errors; a tool that makes them visible and labeled is doing the important part. For how this feeds a larger plan, see building a SaaS MVP without a developer. For the surrounding PM workflow, see our product-manager tooling guide.

The meshcode angle

meshcode is a native macOS and Windows app with a multi-pane agent workspace — run the top-down estimate in one pane, the bottom-up unit math in another, and diff the results in a third while you watch. It drives your existing Claude Code or Codex subscription or metered pay-as-you-go credits, and your skills and hooks carry over unchanged. Feed the skill your real numbers, not your pitch-deck numbers; the sensitivity scenarios are only as honest as the inputs, and that part is still your job.

Add the biz-market-sizing skill in the app

👉 Download meshcode — Mac, Windows