What Squsha is —
and what it isn't

A current Claude model from Anthropic — the same family behind the Claude chat app. We do not name the exact version because we move to a newer one when it does the job better. The point of saying this at all is that the model is not the product: two passes of it run per report, one to derive the search phrase from your brief and one to write the analysis, and both are given only what the sources returned.

Partly. A chatbot with browsing will open a few pages it chose, and you cannot see which sentences came from where. It will not build a roster of competitors from the live results and re-check each pick in code, it will not refuse to print a market size nobody measured, and it will not give you the same twenty sections next month so you can see what changed. With enough prompting and patience you could approximate a good deal of it by hand — that is roughly the two days of work a report replaces.

Yes, in one case: when the live roster does not yield five genuine competitors, it may add well-known names it knows. Those are stamped Unverified by the code — not by the model’s own claim — and sorted after the verified ones. The instruction it is given is explicit that four evidenced competitors beat five with one invented, and that an invented competitor is the single worst failure the report can have.

Two things. The model may only copy dollar and growth figures that appear in the measured data block it was handed, and must print Not measured when there are none — research-firm forecasts are quarantined in their own labelled field with the firm’s name and the broader market they describe. And the headline numbers on the page do not come from the model at all: they are computed in Python from the same government data points, so a figure worth printing in large type never passes through a language model on its way to you.

Yes. A vague brief produces a search phrase about a vaguer market than you meant. A source can be thin or unreachable for a run, and the report will say so but will have less to work with. The model can misread a snippet. And a From live search badge proves a competitor appeared in this run’s results, not that it was the best pick from them. What the checks rule out is a narrower thing: a name or a number with no traceable origin.

Because a chatbot answers from memory, and a report does the work first. A run makes live requests to Google Trends, two Google searches, YouTube, Reddit, Quora, X, Hacker News, Indie Hackers, Product Hunt, GitHub, the App Store, G2, Trustpilot, TechCrunch, the US Census and BLS, Statistics Canada and SEC EDGAR, then opens each competitor’s own pricing page. Then the model reasons through everything it read, section by section, citing what it found, and every quote and price is checked against the page it came from before you see it. That reasoning is most of the eight minutes. A chatbot skips both halves, which is exactly why it is faster.

It is not built to. The opportunity score is arithmetic: four dimensions read from live data, weighted and added, plus a quarter counted straight from the evidence — pain points people actually wrote, competitor weaknesses they reported, how much the community is talking. It uses the whole 0–100 range, so a weak niche scores low and the verdict comes back Avoid. Risks are graded, and a deal-breaker is marked High rather than softened into a caveat. That said, a low score is a reason to look harder, not proof — see the section on what a report cannot tell you.

Yes. Use it to shape the idea, thin twenty half-ideas down to three, and argue with yourself about them — it is free, instant and it talks back, none of which a report does. Then measure the three here. When the two disagree about a competitor or a figure, ask which one can show you where it came from.