How to Verify an AI Stock-Research Answer in 60 Seconds

You can verify most AI-generated stock-research claims in about 60 seconds. Define the exact claim, open the primary source, match the company, period, reporting basis and unit, reproduce the calculation, and check whether newer evidence has superseded it.

This is not a substitute for full due diligence. It is a fast gate between an AI answer and your spreadsheet, note or investment discussion. The objective is to catch the plausible near-miss before it becomes an input.

The 60-second verification protocol

Time	Check	Question
0 to 10 seconds	Define the claim	What exactly is being asserted?
10 to 25 seconds	Open the source	Does the primary document contain it?
25 to 40 seconds	Match the labels	Company, period, basis and unit correct?
40 to 50 seconds	Reproduce the math	Do the cited inputs produce the result?
50 to 60 seconds	Check freshness	Is there a newer or corrected disclosure?

Let us work through each step.

0 to 10 seconds: turn prose into one testable claim

AI answers often combine a fact, a calculation and an interpretation in one smooth sentence:

Revenue grew 18 per cent, margins expanded and the company is gaining market share.

That is not one claim. It is at least three:

1. Reported revenue grew 18 per cent over a specified comparison period.
2. A defined margin increased by a stated amount.
3. The company’s growth exceeded the relevant market or competitor set.

The first two may come from company financials. The third needs external market evidence. Separating them prevents one sourced number from lending false credibility to the rest of the paragraph.

Write the claim in a strict format:

[Company] reported [metric] of [value] for [period], on a [consolidated or standalone] basis, in [unit].

If the answer cannot be made that specific, it is not ready to verify.

10 to 25 seconds: open the primary source

Follow the citation to the actual company or exchange document. Prefer, in order:

1. audited annual or quarterly financial statements;
2. an exchange filing;
3. a company investor presentation;
4. a concall transcript for management commentary;
5. a secondary source only when the primary evidence is unavailable.

Do not stop at a search-result snippet or a generic link to the company’s investor-relations page. The evidence should lead to the relevant document and, ideally, the relevant page or table.

Then ask the simplest question: does the source actually contain the claimed fact?

If the answer is no, the verification stops. The claim remains unsupported even if the document is real and related to the company.

25 to 40 seconds: match company, period, basis and unit

This is where many confident answers fail.

Company and entity

Is the figure for the listed parent, a subsidiary, a business segment or the consolidated group? A subsidiary’s growth can be true without describing the listed company’s overall growth.

Period

Is the number quarterly, year to date, trailing twelve months or annual? Is the growth sequential or year on year? “Latest” is not a period.

Reporting basis

Is it consolidated or standalone? Both sets of financials may be valid. Mixing them inside one calculation is not.

Unit and scale

Is the value in rupees, lakh, crore, million or billion? Is a percentage a ratio, a percentage point change or a share of a category? Unit mistakes can change a conclusion by 100 times while leaving the sentence grammatically perfect.

A useful mental checklist is C-P-B-U: company, period, basis, unit.

40 to 50 seconds: reproduce the arithmetic

If the AI answer contains a calculated number, use the cited inputs.

Suppose it says revenue rose from ₹100 crore to ₹118 crore:

Growth = (₹118 crore / ₹100 crore) - 1 = 18 per cent

That calculation is easy. More complex ratios still follow the same rule: identify the numerator, denominator, averaging convention and reporting basis.

Watch for these shortcuts:

• comparing a full year with nine months;
• calculating a margin from net income when the label says operating margin;
• using closing equity instead of average capital where the method requires an average;
• treating a one-time gain as recurring profit;
• filling a missing input with zero;
• rounding each input before calculating.

An AI explanation may be persuasive while the underlying inputs are incompatible. Reproduction is what turns the number from prose into a calculation.

50 to 60 seconds: check freshness and supersession

Look at the document date, then ask whether a newer relevant filing exists.

The company may have:

• released a more recent quarter;
• revised the result;
• changed guidance in a later concall;
• completed a split, merger or demerger;
• reclassified a business segment;
• corrected an exchange disclosure.

Freshness is not the same as using the newest document in every situation. If you are reconstructing a decision made on a past date, the correct source is the newest document that was available by that date. A later correction belongs to today’s view, not to the historical information set.

That is the role of point-in-time data: it preserves what was knowable then as well as what is known now.

A worked example: a low P/E that is not what it seems

Imagine an AI answer says:

The company trades at 4 times earnings because its profit rose sharply in FY26.

Apply the protocol.

Define the claim. The market capitalisation divided by FY26 earnings equals 4, and FY26 profit represents recurring earnings.

Open the source. The FY26 statement shows a large exceptional gain from an asset sale.

Match the labels. The number is consolidated annual profit, correctly labelled.

Reproduce the math. The reported P/E may indeed be 4.

Check the interpretation. Removing the non-recurring gain produces a very different earnings base.

The AI did not necessarily hallucinate the P/E. It failed to distinguish a mechanically correct ratio from an economically meaningful one. Verification catches both factual errors and interpretation errors.

Why “has citations” is not enough

There are four levels of evidence quality:

1. No source. The answer is unverified.
2. Document-level source. A relevant filing is named, but the claim is hard to locate.
3. Passage-level source. The exact page or passage is linked.
4. Claim-level support. The passage supports the exact company, period, basis, unit and wording of the claim.

Aim for the fourth level when a statement matters. A source list at the end of a report can be useful for further reading, but it should not be mistaken for proof of each sentence.

When 60 seconds is not enough

Slow down when the claim involves:

• a material restatement or accounting-policy change;
• related-party transactions;
• cash-flow quality or working-capital classification;
• guidance reconstructed across several quarters;
• segment reorganisation;
• a forecast, target price or probability;
• portfolio exposure or position sizing.

The 60-second protocol is a triage gate. Complex claims need the full source trail and, often, a second reviewer.

How Altys shortens the check

Altys is designed to place the evidence beside the answer. Financial figures come from source-linked data, derived numbers are calculated by code, and material research claims carry citations. The system retains point-in-time history and returns an unavailable state when the supporting evidence cannot be established.

For calculation-heavy work, verification continues in the export. Strategy screens and scorecards download as formula-native Excel workbooks with inputs, eligibility, weights, calculations and provenance. GenGrid and analytical tables export as Excel-ready CSV. This lets an analyst recalculate, stress a weight, filter the underlying universe and inspect the evidence without depending on the Altys interface.

That does not remove the analyst from the loop. It compresses the mechanical part of verification: finding the document, identifying the period and basis, checking the calculation and seeing what was knowable on a date. The design is explained in How Altys prevents hallucinated financial numbers.

Save this checklist

Before using an AI-generated financial claim, ask:

• What exactly is the claim?
• Where is the primary source?
• Does it match the company and entity?
• Does it match the period and reporting basis?
• Is the unit correct?
• Can I reproduce the calculation?
• Was the information available on the date being studied?
• Has a newer disclosure changed it?

If one answer is unclear, the number is not ready for the model.

This article is educational. Altys Labs is a financial data and analytics platform, not a SEBI-registered Research Analyst or Investment Adviser. Nothing here is investment advice or a recommendation to buy, sell or hold any security.