How it works
Six passes over one deal, in a fixed order, with a hard separation between the parts a computer should do and the parts a model should do. That separation is the whole design.
You give it what you have
A listing URL from any marketplace or broker site — or just the listing text if the page blocks scrapers. Then whatever documents exist: P&Ls, tax returns, the lease, a CIM, payroll registers, bank statements, equipment lists, utility or fuel reports, seller disclosures. More documents means a more confident report, but a listing alone still produces a useful one.
Every document is read and traced
PDFs, spreadsheets, Word files, CSVs and photographs of statements are all parsed, with page and sheet boundaries preserved. That traceability matters: a figure that cannot be pointed back to a page is marked unsupported, and unsupported figures do not count toward earnings.
The arithmetic runs in code, not in a model
Revenue, gross profit, EBITDA, SDE, normalized SDE, add-backs, margins, debt-service coverage, break-even, margin of safety, working capital, every multiple and five valuation methods are computed in a deterministic engine. A language model is never asked to add, divide or amortize — it is given the results and asked what they mean.
The listing’s own history is recovered
Archived versions of the same advertisement are pulled from the Internet Archive and parsed with the same extractor as the live page, then diffed field by field. Price cuts, restated cash flow, a rewritten reason for selling and a moved founding year all surface automatically — with the dates.
The outside world is investigated
Competitors within 1, 3, 5 and 10 miles with ratings, review counts and hours. Census demographics for the trade area against national benchmarks. Traffic counts and planned development. Corporate registrations, licences, inspections, liens, litigation and property records. Each research brief is independent, so a blocked source degrades one section rather than the report.
A model does the judgement — and only the judgement
With the numbers fixed and the research in hand, the model identifies risks, quantifies each in dollars, writes the narrative, and drafts the offer strategy and the questions for the broker. On the Deep tier a second pass audits the finished report against the computed facts for contradictions and unsupported claims.
You choose the model
The default is OpenAI’s flagship at high reasoning effort, which is the most reliable at catching arithmetic that does not reconcile. You can switch to any of the others on any analysis — and because the financial engine is deterministic, the numbers do not change when you do. Only the interpretation does.
| Model | Provider | Context | Best at |
|---|---|---|---|
| OpenAI GPT-5.4 · High reasoningDefault | OpenAI | 400K | Default. Deepest financial reasoning and the most reliable at catching arithmetic that does not reconcile. |
| OpenAI GPT-5.2 Pro | OpenAI | 400K | Slowest and most expensive, and the best at the hardest reconciliation problems. |
| Claude Opus 5 | Anthropic | 200K | Excellent at long-document synthesis and writing the narrative sections in a defensible, hedged voice. |
| Claude Sonnet 5 | Anthropic | 200K | Fast, capable balance for Deal Scan and mid-tier reports. |
| Gemini Pro · Google Search grounding | 1,000K | Best outside-research engine — native Google Search grounding returns real, citable sources for entity, licence and lawsuit checks. | |
| Gemini Flash | 1,000K | High-volume grounded search at low cost. Used for the research fan-out. | |
| Kimi K3 | Moonshot AI | 256K | Very strong long-context reasoning at a fraction of flagship pricing. Good second opinion on the financial reconciliation. |
| GLM-5 | Zhipu AI | 200K | Zhipu flagship. Competitive structured-extraction quality, very low cost. |
What the report will never do
- Invent a record, a case number, a filing ID, a date or a URL. If a search returns nothing, the report says so — an empty result is an honest finding.
- Let a seller claim drift into being described as a fact. Every assertion is tagged documented, seller claim, or hypothesis.
- Speculate about a named individual’s motives or integrity. It describes records, not people.
- Manufacture red flags to look thorough. When something checks out, it says so plainly.
- Give you legal, tax, accounting or environmental advice. It tells you precisely what to take to the professional who can.
Common questions
How long does an analysis take?
A Deal Scan takes about three to five minutes. A Full Deal Report takes eight to fifteen. Deep Due Diligence takes fifteen to twenty-five, because it runs the public-record and environmental research and then audits its own output.
What if the marketplace blocks the scraper?
Several marketplaces block datacenter traffic. When that happens the report falls back to archived versions of the listing and to text you paste in directly, and it tells you plainly in the limitations section that the live page could not be read.
What file types can I upload?
PDF, Excel, Word, CSV, plain text and images. Scanned documents are read with vision OCR, though a native export is always more reliable.