Four layers, kept apart
Source, observation, measure, interpretation
Every item PostMoney shows belongs to exactly one of these layers. Interpretations cite the observations that support them and, where possible, surface counterevidence.
Source artifact
The filing, article, job post, API record, review, research paper, or user-uploaded document. The thing itself, retrieved and retained as the license allows.
Observation
A plain fact the artifact supports: "Competitor A posted eight enterprise sales roles," or "the regulator published a proposed rule." No judgment attached.
Derived measure
A count, growth rate, segment total, topic cluster, or calculated TAM scenario. Built from observations, with method and assumptions exposed.
Interpretation
"Competitor A may be moving upmarket," or "regulation could expand the serviceable market." Cites the observations that support it and, where possible, the counterevidence.
The interface labels these layers visually. "Search interest rose" is an observation; "Demand is accelerating" is an inference. A reader should never have to guess which one they are looking at.
Provenance worth preserving
What travels with every observation
A citation is not enough. These are the fields the research says are worth keeping on each piece of evidence, so a claim can be inspected months later and a report can be reproduced as of its date.
- Original source and canonical URL
- Publisher, jurisdiction, and source type
- Observed, published, effective, and retrieved dates where relevant
- Entity-match rationale and confidence
- Verbatim evidence excerpt or structured source location when permitted
- Extraction method and confidence
- Units, currency, geography, segment, and time period
- Whether a value is reported, estimated, calculated, or inferred
- Source freshness and expected update cadence
- Corrections, supersession, and user decisions
- Report snapshots that preserve the evidence available at the time
Two confidences and a materiality
Three separate questions, never one score
Company monitoring fails when a common name produces a confident stream of unrelated results. Matching combines multiple independent clues, and confidence answers two separate questions. Importance is a third question again.
Entity confidence
Is this observation about the right company?
Built from independent clues: exact legal or regulatory identifier; canonical domain or a link to or from it; official handle or verified account relationship; founder and executive names; product names and branded phrases; headquarters, geography, sector, customers, investors, and known aliases; article links, images, and organization structured data; and negative evidence such as a different city, industry, or namesake domain.
Claim confidence
Does the evidence support the summarized event?
A post from the official company account has high entity confidence but may still be an unverified company claim. A government filing tied to the exact legal identifier can have high confidence for the fact of filing, while any business interpretation remains an inference. A local article using only a common brand name may require review even if its claim sounds plausible.
Materiality
Does it matter?
Kept separate from both confidences. A perfectly matched holiday post is high confidence and low importance. A weakly matched lawsuit is potentially high importance but must remain quarantined until reviewed.
Source hierarchy
Which evidence wins a tie
The product could favor evidence roughly in this order, while allowing domain-specific exceptions.
| Rank | Source type |
|---|---|
| 1 | Official regulator, filing, or original company artifact |
| 2 | Direct structured record or licensed primary dataset |
| 3 | Reputable reporting with named sourcing |
| 4 | Specialized trade reporting or analyst research |
| 5 | Public community and review evidence |
| 6 | Modeled estimates and AI inference |
Source rank is not truth. A company press release is authoritative about what the company announced, not whether the claim is accurate. A review is authoritative about one reviewer's statement, not the whole customer base.
Corrections and coverage health
Evidence changes, and monitors break
Corrections are first-class events
Sources edit headlines, funding amounts, people, dates, and claims. The system should detect meaningful modification when possible, retain history, and propagate a correction to alerts, summaries, and report drafts. Deleted social posts are marked deleted rather than silently disappearing, and retention follows platform policy. When a page changes or disappears, PostMoney preserves the allowed evidence and retrieval metadata, not implying that a snapshot is the current page.
Coverage health is visible
"No news" is meaningful only if the monitors are working. The company page could show when each source was last checked, last succeeded, last produced an item, and whether authorization expired. A LinkedIn gap should say "not connected," not appear as an empty LinkedIn feed. A site blocking crawlers should remain a gap, not trigger evasion.
Boundaries
What the product refuses to do
Public does not mean consequence-free, and a citation does not cure a bad ingestion model. These four boundaries come straight from the research.
Public does not mean consequence-free
PostMoney collects only what serves a legitimate company-intelligence purpose. Monitoring founders can feel invasive even when posts are public, so the default is professional, company-relevant content. No private groups, circumvented access controls, personal-life profiling, sensitive-trait inference, face recognition, or tracking of non-public individuals. Users see why a person is monitored and can remove them.
Rights, source by source
Many valuable databases and publications restrict API use, retention, training, redistribution, or customer-facing display. Rights are evaluated source by source. Citations do not cure an unlicensed ingestion model. Direct quotation is limited, and article text is not redistributed when the license permits only links and snippets: store the title, URL, timestamp, compliant excerpt, observation, and provenance rather than full text.
No false precision
PostMoney does not convert job postings into headcount, social engagement into revenue, website traffic estimates into actual usage, or article sentiment into company health without clear labeling. Higher-order observations are valuable as hypotheses. "Hiring mix shifted toward enterprise sales" is an observed pattern. "The company is successfully moving upmarket" is an interpretation requiring more evidence.
No contamination of portfolio truth
External observations and estimates must not silently become canonical founder-reported datapoints. They can contextualize, challenge, or suggest a review, but they carry distinct provenance and semantics. A founder's reported ARR and a third-party traffic estimate never share a column.
Risks and guardrails
Nine ways this goes wrong, and the answer to each
| Risk | Guardrail |
|---|---|
| False precision | TAMs, private-company revenue, traffic, downloads, employee counts, and sentiment are frequently estimates. Display ranges, methods, confidence, and sensitivity rather than a single authoritative-looking value. |
| Entity-resolution mistakes | Company names, product names, founders, and acronyms collide. Every watch profile needs aliases and exclusions; low-confidence matches enter a review queue instead of silently contaminating the dossier. |
| Popularity and coverage bias | English-language, U.S., venture-backed, and digitally visible companies are overrepresented. Bootstrapped, local, services-heavy, and non-English competitors may be invisible. Coverage indicators show where the system is blind. |
| Narrative herding | News volume and funding can reflect hype rather than customer value. Deliberately include disconfirming evidence, failures, quiet incumbents, and non-venture substitutes. |
| Licensing and copyright | Many valuable databases and publications restrict API use, retention, training, redistribution, or customer-facing display. Evaluate rights source by source. Citations do not cure an unlicensed ingestion model. |
| Privacy and ethical people monitoring | Avoid personal surveillance, sensitive personal profiling, private groups, or inferred employee behavior. Focus on public professional events relevant to company and market analysis, and provide correction and deletion paths. |
| Regulatory and scientific overclaiming | Do not turn an adverse-event record, proposed rule, patent, or clinical trial into legal, medical, or scientific advice. Preserve the official status, jurisdiction, caveats, and source link. |
| Alert fatigue | More collection can make the product worse. Let users define materiality, learn from dismissals, batch weak signals, and explain why each alert matters. |
| Contamination of portfolio truth | External observations and estimates must not silently become canonical founder-reported datapoints. They can contextualize, challenge, or suggest a review, but carry distinct provenance and semantics. |
Ten principles
The product principles, verbatim
- Research is a living evidence system, not a generated essay.
- Start from the investor's market definition and thesis, not a generic industry keyword.
- Separate source, observation, calculation, and interpretation.
- Show assumptions, ranges, counterevidence, confidence, and freshness.
- Prefer meaningful change over exhaustive feeds.
- Treat competitors as an ecosystem that includes incumbents and substitutes.
- Keep external estimates distinct from company-reported portfolio data.
- Make every important claim inspectable and every point-in-time report reproducible.
- Respect source rights, platform boundaries, privacy, and domain-specific caveats.
- Keep watchlists, annotations, thesis judgments, and report usage organization-scoped.
Where the evidence comes from, and what each source family cannot tell you, is on Sources. The smallest coherent product that honors these principles is on Foundations.