Skip to content
Software & SaaS Chrome extension 2026

Veridex

Company research that shows its working.

Overview

A Chrome extension that turns a company name into verified, source-linked data — every value with its source, confidence and date, using on-device AI.

Veridex turns a company name into structured, source-linked data — every value carrying the URL it came from, a confidence score and the date it was found.

The problem with asking a chatbot

Ask a language model for a company's CEO and revenue and you get a fluent paragraph that reads identically whether it is right, out of date, or about a different company with a similar name. For prospect lists, diligence or anything you have to stand behind, that fluency is the problem: it hides uncertainty instead of surfacing it.

Veridex inverts the design. The AI is never allowed to be the source of a fact. It plans the search and suggests where to look; every suggestion is then fetched and checked against the live page. Anything that cannot be confirmed resolves to Not found — enforced in the pipeline, where every requested field maps to a value or to null.

Locate with AI, verify with live data

The model is genuinely useful for one thing rules can't do: guessing that a company's site might be at an abbreviated domain no name-slugging rule would derive. So it suggests domains — and every suggestion is then probed and token-checked against the fetched body. A hallucinated domain that doesn't resolve, or resolves to a page that never mentions the company, is discarded. The AI supplies leads, never facts.

Same-name disambiguation is most of the work

A free-form company name is a weak key; dozens of firms share one. Veridex resolves to the official domain instead, because a domain is unique and checkable, and layers four defences on top: a name and subject guard, a domain guard that rejects unrelated sites sharing a single word with the name, anchor-weighted reconciliation that gives the company's own site decisive weight over same-name aggregator records, and a final semantic check that can only ever remove a suspect value.

Getting this wrong doesn't produce an error. It produces a clean, confident, completely wrong report — which is why it gets four layers.

Confidence follows method

Values are scored by how they were found, because how you found a fact predicts how likely it is to be right: a structured API answer scores 0.96, the site's own JSON-LD 0.92, microdata 0.86, OpenGraph 0.80, meta tags 0.72, and text read heuristically out of page copy 0.55. Two independent sources agreeing promotes a field to verified, and up to four alternatives are retained per field rather than quietly discarded.

The honesty extends to what isn't scored: deep links to registries are surfaced without a confidence value at all, on the reasoning that a patent-search link is not evidence a company holds patents.

On-device AI, and no server

The AI is Chrome's built-in Gemini Nano via the Prompt API — there is no cloud AI endpoint anywhere in the extension. Sessions are pooled per system prompt and cloned per call so one extraction's context can never bleed into the next, and every prompt is serialised through a promise chain because the single on-device model contends and times out if called in parallel.

Veridex has no backend of its own. Searches, results, exports and settings live in IndexedDB and local storage; requests go from the browser straight to the source without cookies. What does leave the device is the query itself — the company name, sent to public search engines and open data APIs. Providers are gated: one is only contacted when a field it can answer was requested and the category matches.

What shipped

Around 17,678 lines across 160 files, with 17 test suites. 62 built-in fields plus custom ones, 96 countries, 40 industry categories, and 18 data providers spanning Wikidata, GLEIF, SEC EDGAR, FDIC, ClinicalTrials.gov, openFDA, PubMed, OpenAlex, UK Companies House and France SIRENE. Bulk research up to 500 companies per file, running in an offscreen document so it survives the panel closing and resumes rather than restarts. Export to PDF, Excel, CSV, JSON or Markdown. Scheduled watching with change notifications, diffed to ignore whitespace and case noise so it doesn't cry wolf.

Highlights

  • Every value carries its source URL, a confidence score and a timestamp — and unconfirmed fields resolve to Not found rather than a guess
  • The AI locates, live data verifies: suggested domains are probed and token-checked, and a hallucination that doesn't resolve is discarded
  • Four-layer entity resolution against the official domain, because a company name is a weak key and a wrong match produces a confident wrong report
  • Confidence follows method — structured API 0.96 down to heuristic text 0.55; two agreeing sources promote a field to verified
  • On-device Gemini Nano with pooled, cloned sessions so context can't bleed between extractions; no cloud AI endpoint exists in the source
  • 18 public data providers, gated so one is only called when a matching field is requested; bulk to 500 companies with export to 5 formats

More software & saas.

View all