Is an evidence gap in an AI-generated summary the same as an error?

No. A gap means a claim could not be traced to a source during the audit. Some gaps turn out to be fabrications, some are paraphrases that drifted from what the source said, and some are true statements that simply were not sourced. The audit's job is to find them; classifying them comes after.

Why not just ask the AI tool for its sources?

Asking is the first step, but a cited source is only evidence after it has been opened and the specific passage located. Tools can cite documents that exist while attributing claims those documents never make.

Does this audit tell me whether the summary's conclusion is right?

No. The audit is descriptive: it separates traceable claims from unsourced ones. What remains after the audit is a cleaner view of what the public record supports, not a judgment about the stock.

How to spot evidence gaps in an AI-generated stock summary

An AI-generated company summary has a distinctive failure mode: it is fluent everywhere and sourced only somewhere. The prose quality is uniform, so nothing about a sentence tells you whether it came from an audited filing, a stale secondary source, or nowhere at all. The risks of relying on such summaries are covered on the hub page; this guide is the working procedure, how to audit one summary, claim by claim, until every sentence is either traced or flagged.

The SEC's investor education arm has warned in an investor alert that AI tools can produce inaccurate or wholly misleading information about companies and that fraudsters lean on AI branding precisely because it reads as authoritative. The audit below is the descriptive counter to that: it does not judge the summary's conclusion, it tests its traceability.

Step 1: itemize the claims

Rewrite the summary as a numbered list of factual claims, one per line. Split compound sentences: "revenue grew 20% driven by the cloud segment while margins expanded" is three claims (growth rate, growth driver, margin direction), and they can fail independently.

Two kinds of sentences drop out at this stage. Pure framing ("the company has a long history") carries no checkable content. And any sentence that reads as an instruction or a judgment about what an investor ought to do is not a claim at all; it is out of scope for an evidence audit and, in ThesisCheck's methodology, out of scope entirely.

Step 2: demand a named, dated source per claim

For each claim, write down the source the summary gives, if any: document name, publisher, and date. Most summaries fail here quietly, in one of three ways:

No source: the claim is simply asserted.
Vague source: "according to filings" without naming which filing or period.
Undated source: a real document with no way to tell whether it is current.

A claim without a named, dated source is not yet wrong. It is untraced, and it goes on the audit list.

Step 3: open the source and find the passage

This is the step that separates auditing from skimming. A citation is only evidence once the cited document has been opened and the specific passage located. For SEC-reporting issuers, Investor.gov's EDGAR guide covers pulling the actual filings, and EDGAR full-text search can locate a quoted phrase across filings when the summary names a document you cannot immediately place.

Three outcomes are possible for each claim:

The passage exists and says what the summary says. The claim is traced.
The passage exists but says something narrower, older, or more conditional. That is paraphrase drift, and it is the most common finding: a hedged disclosure became a confident assertion.
No passage in the cited source contains the claim. The citation is decorative.

Step 4: mark what cannot be traced

Every claim still standing after step 3 without a located passage is an evidence gap: a statement the summary needs that the reviewed record did not supply. Gaps are findings, not accusations. As the missing evidence guide explains, a gap does not disprove the claim; it names an assumption that is currently resting on the summary's fluency instead of on a source.

The discipline is to keep gaps visible instead of resolving them optimistically. A summary with twelve claims, seven traced, two drifted, and three unsourced, is a much more useful object than the same summary read as a single trustworthy paragraph.

Step 5: rebuild the summary as a ledger

The finished audit is effectively a small source ledger: claim, source, date, locator, and status for every line of the original summary. At that point the summary has been converted from prose into evidence, and the interesting structure appears: which claims the record supports, which drifted in the retelling, and which have nothing under them.

Why this stays descriptive

Nothing in this procedure evaluates whether the company is good or what anyone's position ought to be. It answers one question: which sentences of this summary can be traced to a dated public source, and which cannot. That question has a checkable answer, which is exactly what makes the audit worth doing and what keeps it inside ThesisCheck's descriptive boundary. It does not produce buy, sell, hold, rating, sizing, or price-target recommendations, and neither should the summary it audits.