How to verify an AI-generated company report before you rely on it

Automated research tools — including this one — can assemble a company dossier faster than any analyst. What they cannot do is decide what is true. A language model reads retrieved pages and organises them; if the retrieval missed a source, matched the wrong company, or picked up a page that was already wrong, the output will still read fluently and confidently. Verification is therefore not an optional final polish. It is the step that converts a summary into evidence.

Step 1: check that every claim carries a citation

Start with structure rather than content. Scan the report for sentences that assert a fact — a score, a fine, a settlement figure, a headcount, a recall — and check whether each one links somewhere. A report where the sourcing is decorative (a few links at the bottom, none attached to specific claims) cannot be checked, which in practice means it cannot be used. No link, no claim.

Step 2: match the legal entity, not the brand

This is the error that survives every other check. Brands are not entities: a hotel chain's reviews may belong to independent franchisees, an enforcement action may name a subsidiary that was later sold, and two unrelated firms can share a trading name in different states. Confirm the entity through a filing, a state registry or the company's own investor page before you attribute anything to it.

Step 3: date everything

Reputation data decays. A review average moves with volume, a docket entry can be superseded within a week, a consent order can be terminated, and a recall can be expanded. Any figure worth quoting needs a retrieval date alongside its URL. If a report gives you a number without a date, the number is a snapshot of an unknown moment.

Step 4: distinguish status from score

• Accreditation is membership, not quality — a Better Business Bureau accreditation status is a paid relationship, not a rating, and should never be rendered as stars.
• A filed complaint is an allegation, not a finding. Look for the disposition: dismissed, settled, judgment entered, or still pending.
• A regulator inspection is not automatically a violation; read the citation and whether it was contested or vacated.
• An analyst rating or media ranking is an opinion with a methodology attached. Read the methodology before you cite the rank.

Step 5: read the company's own words

Primary sources include the company. Its newsroom, its 8-K disclosures and its litigation statements tell you what it has committed to publicly — which is often more useful than commentary about it, because a company can be held to its own filings. Where a report shows a company response next to a criticism, read both and note where they actually disagree.

The point of a sourced report is not that you trust the summary. It is that you no longer have to.

OpenWebReview editorial note

Step 6: settle disputes in the primary databases

• EDGAR full-text search — confirm figures, risk-factor language and legal-proceedings notes in the filing itself.
• CourtListener — confirm that a case exists, which court it sits in, and what the latest docket entry actually says.
• The relevant sector regulator — NHTSA for recalls, OSHA for workplace inspections, NLRB for labour cases, CFPB or FTC for consumer enforcement.
• The company's investor-relations newsroom — for its own statements and the dates it made them.

Step 7: apply a published standard, not a vibe

There is now external guidance for exactly this problem. The NIST AI Risk Management Framework asks whether an AI system's outputs are valid, reliable and traceable to their inputs. The EU AI Act pushes transparency obligations onto providers and deployers of general-purpose systems. And the FTC's enforcement sweep against overstated AI claims establishes the practical point: presenting unsubstantiated automated output as fact is a consumer-protection issue, not a technical footnote. If a tool cannot show you its inputs, that is the finding.

None of this takes long. Twenty minutes of verification on a report you intend to act on is the difference between research and repetition — and the habit generalises to every automated tool you will use after this one.