How we verify, grade, and index company research.

The three-tier system

Every company record in the database sits in one of three tiers, assigned automatically at build time from measurable properties of the record itself.

Flagship profiles (170 records) are the pages we submit to search engines. A flagship profile must cite at least three public sources that are genuinely independent of the company, carry at least 400 words of original editorial analysis, have no unresolved verification flags on material claims, and either score at or above our research-priority threshold or be an established public company, defense prime, or government-owned company whose significance does not depend on our scoring.

Listed profiles (the balance of the database) meet our publication standard — they are browsable in the Startup Database and sector listings, and they feed the research tools — but they carry a noindex directive. Most fall short of the flagship bar on independent sourcing: our own research shows that hundreds of real, operating Israeli companies simply have not accumulated enough third-party public coverage for us to publish an externally-verifiable deep profile. We would rather list them honestly as research leads than dress them up as verified intelligence.

Unlisted records are entries whose existence or operating status we could not verify at all. They do not appear in listings, are never linked from other pages, and exist only so that our pipeline remembers what it has already investigated.

What counts as an independent source

Source counting is stricter than it looks. A company's own website never counts, no matter how detailed. LinkedIn pages never count, because they are self-authored. Duplicate links to the same URL count once. What remains — regulatory filings, press coverage, government procurement records, conference materials, index constituent lists, investor disclosures — is what we mean by an external public source. Each source on a flagship profile also carries a grading note describing what the source can and cannot support: a regulatory filing is treated as disclosed-transaction evidence, not independent customer verification; a founder interview is treated as a company claim with a date attached.

The editorial floor

Indexable profiles must contain at least 400 words of editorial analysis across the description, dual-use assessment, strategic rationale, competitive assessment, and risk factors — and the words are counted after our language sanitizer strips promotional phrasing, so padding does not help. In practice, published profiles average well above 1,000 editorial words. The analysis follows a consistent argument structure: what the technology actually is, who demonstrably uses it, what the commercial evidence supports, who else does the same thing, and what a diligent reader should verify next.

What the scores mean — and do not mean

Each record carries comparative research scores (technology, market, team, dual-use, strategic alignment, and a composite potential score on a 0–100 scale). These are structured research aids for ranking a large corpus, not investment ratings, price targets, or recommendations. The flagship gate uses the composite score only as a curation filter — a way of asking "is this one of the records where our analysis is deepest and most decision-relevant?" — and deliberately exempts public companies and defense primes, whose importance to the ecosystem map is independent of any score we assign.

The entity taxonomy

"Startup" is often the wrong word. The database classifies every record as one of: an operating startup, a public company, an acquired asset (a company absorbed by an acquirer, tracked because its technology and team still matter), a defense prime, a government-owned company, a fund, an ecosystem reference, a defunct or wound-down company, or an unverified record. The classification is applied automatically from evidence in the record and corrected editorially, and it changes how a profile is written: an acquired asset's profile focuses on what survived the acquisition, and a defunct company's profile focuses on what its failure teaches.

Verification flags

Records whose public web presence is limited carry a visible verification note, and cautious language replaces specific claims. A profile that makes material claims — funding events, customer relationships, government contracts, operational deployments — while carrying a verification flag is automatically excluded from indexing until the claims are sourced or removed. Material claims on indexable pages must trace to a linked public source; a build-time audit fails the deployment if they do not.

Enforcement: the build refuses to ship violations

These rules are not a style guide; they are code. The tier policy lives in a single shared module consumed by the site build and by two independent audit scripts that re-derive the numbers from the raw data on every deployment. If an indexable page falls below the word floor, cites too few independent sources, links to an unlisted record, appears in the sitemap while marked noindex, or contains banned boilerplate or promotional language, the deployment fails and nothing ships. The same pipeline enforces performance budgets and accessibility test suites.

Dataset methodology for the research tools

The interactive research tools are built from documented datasets with their limitations stated. The Dependency Atlas uses public trade data whose known blind spots we publish rather than hide — Israeli fuel-import origins, for example, are materially understated in public records, and the Atlas says so. The TASE Index Explorer uses hand-curated constituent weights, because official index weights are not freely published; our curation method, cap rules, and refresh dates ship with the dataset. Company-level tool data comes from the research-grade slice of the database — records meeting the two-source and editorial-floor standard — so the tools never present unverified records as analyzable companies.

Cadence and corrections

Every profile displays its last-updated date, sourced from the underlying record rather than the build timestamp. Database refreshes are shipped as reviewed batches with commit-level provenance. Corrections follow the process in our editorial policy: send the page URL, the proposed correction, and a public source through the Contact page; material corrections update the page's modification date, and better information can qualify, demote, or remove an entry entirely — including our own.