Find Websites Feeding Wrong Financial Data to AI Chatbots

Source tracing for investor relations is a method that helps IR teams find websites feeding wrong financial data to AI chatbots through visible citations. The team captures answers from ChatGPT, Google AI Overviews, Google AI Mode, and Microsoft Copilot. It then checks each cited claim against the company’s filing in SEC EDGAR.

Visible citations show which links appeared with a captured answer. They do not reveal every source or process behind that answer. A clear record lets the Head of Investor Relations, CFO, or legal team review the answer, cited page, and filing evidence together.

How can an IR team identify the source behind a wrong AI answer?

An IR team can identify a visible source by saving the complete answer and opening every attached citation. The team should preserve this evidence before changing a company page or requesting an outside correction.

Use this capture order:

1. Save the exact prompt.
2. Save the complete answer.
3. Record the engine, interface, date, and time.
4. Copy every visible citation URL.
5. Save each page title and relevant passage.
6. Record whether the session was signed in.
7. Record the exact disputed statement.

A screenshot may hide a destination or shorten a URL. Save the answer as text and preserve each full URL. A PDF or browser printout can provide a second copy.

What does a visible citation prove?

A visible citation proves that the captured interface displayed a link beside or near the answer. It does not prove that the linked page was the only source used.

Connect each disputed claim to its displayed citation. Note whether the page contains the figure, period, executive name, or business description at issue.

Do not label an uncited page as the source because it ranks highly in search. Search position does not prove that an engine used that page. The audit method should keep direct evidence separate from possible source candidates.

What belongs in a source-tracing record?

A source-tracing record should let another person review the answer, cited page, and filing evidence. The record should keep captured facts separate from the team’s conclusion.

This structure prevents a summary label from replacing the underlying quotes. The audit methodology describes how those records support a filing comparison.

Which source links should the team capture from each AI engine?

The IR team should capture every visible source link from ChatGPT, Google AI Overviews, Google AI Mode, and Microsoft Copilot. Record the full URL, page title, relevant passage, and citation position.

Citation displays can differ by product and interface. Preserve the answer and its source panel together when possible. Do not assume that similar titles lead to the same page. Treat each captured interface as its own evidence record.

What should the team capture from ChatGPT Search?

A ChatGPT Search capture should preserve inline links, the Sources control, the full answer, and the destination of each displayed citation. Open each source control before ending the session.

OpenAI’s ChatGPT Search guidance says search responses can show inline citations and a Sources button below the response. Last checked: 2026-10-16.

Copy every destination URL rather than relying on the visible title. If the capture shows no citation, label it unsourced. Do not assign the answer to model memory or a website without direct evidence. The guide to correcting wrong financial figures in ChatGPT covers the separate correction process.

What should the team capture from Google AI Overviews and Google AI Mode?

A Google AI Overviews or Google AI Mode capture should contain every displayed supporting link and the claim shown near it. Save the link title, domain, full URL, answer text, and visible source controls.

Google’s guidance for AI features and websites says AI features in Search show links to supporting web pages. Last checked: 2026-10-16.

Do not treat the first link as support for the full answer. Match each disputed statement to the page passage that may support it. A focused guide explains how to review financial source links in Google AI Overviews without assuming that one citation explains every claim.

What should the team capture from Microsoft Copilot?

A Microsoft Copilot capture should preserve the complete answer, exact product interface, visible citations, and displayed web sources. Product details matter because the cited guidance covers specific Copilot products.

Microsoft’s web search guidance says Microsoft 365 Copilot Chat and agents can use web search and cite web information. Last checked: 2026-10-16.

Save every displayed destination URL and the statement beside it. Do not apply guidance for one Copilot product to every Copilot interface. The guide to Microsoft Copilot citations for investor relations provides a more focused capture process.

How should the team compare a cited page with an SEC filing?

An SEC EDGAR filing comparison should match the disputed statement with a filing for the same company, period, metric, unit, and scope. Place the filing quote beside the cited page quote.

MS House uses the company’s SEC filing as the source of truth for financial facts. SEC EDGAR search provides access to company filings. Last checked: 2026-10-16.

Start with the filing tied to the disclosure. This may be a Form 8-K Item 2.02 exhibit, Form 10-Q, or Form 10-K. Record the form, filing date, section, direct URL, and exact passage.

Which filing details should match?

The cited page and SEC filing should match on the period, unit, metric definition, and business scope. A matching number alone does not establish agreement.

Check these fields:

1. Period: Does the statement cover the same reporting period?
2. Unit: Does the figure use dollars, thousands, millions, or another unit?
3. Scope: Does it cover the company, a segment, a region, or a product?
4. Metric: Does the page use the same definition as the filing?
5. Status: Is the statement a reported result, estimate, target, or guidance?
6. Currency: Does the source use the same currency as the filing?

Copy the filing sentence exactly. Include nearby text when it defines the metric, period, or scope.

How should the filing comparison be labeled?

The filing comparison should use a narrow label that describes the captured evidence. Avoid a broad label such as “false” when the problem is an old period or wrong unit.

These labels describe evidence. They do not assign intent to the publisher or engine.

What if an AI answer about the company has no source link?

An answer with no visible source should be recorded as unsourced for that capture. The IR team should preserve the answer instead of guessing where it came from.

Save the prompt, complete answer, engine, interface, date, time, and signed-in state. The team may test the same prompt in other engines. A citation found elsewhere can guide later research, but it does not prove that the first engine used that page.

A narrower prompt can ask for the source supporting the disputed claim. Save that response as a separate capture. Do not attach a later link to the original record.

Can the team search for matching wording?

The team can search for an exact sentence as an investigative step. A wording match does not prove that the engine retrieved or used the matching page.

Search for a distinctive phrase from the answer. Check whether a result contains the same figure, period, and surrounding language. A close match can guide the next review.

Label the matching page as a candidate source unless the captured answer links to it directly. This distinction keeps visible evidence separate from an inference made during research.

When does OAI-SearchBot access belong in the source investigation?

OAI-SearchBot access belongs in the investigation when a company IR page is missing from captured ChatGPT Search citations. Treat access as a separate technical check, not as source evidence.

Check robots.txt after preserving the answer, citations, prompt, and session details. A check may help when a corrected IR page remains absent from later captures or a third-party page appears instead.

Record the tested URL, user agent, robots.txt rule, and test date. A rule that blocks OAI-SearchBot does not prove that ChatGPT Search used another page. A rule that allows access does not prove retrieval or citation.

The OAI-SearchBot guide explains how to run and document this check.

How can the team distinguish a stale source from a parsing error?

A stale source contains information from an older period and presents it as current. A parsing error occurs when the cited page is current, but the answer changes its meaning.

Open the cited page and locate the disputed statement. Then compare that statement with the relevant SEC filing.

Do not rely only on a page’s displayed update date. Compare the page text with both the filing and the captured answer.

What if the cited page contains old and new figures?

A page containing old and new figures creates ambiguous evidence. Record which passage is closest to the answer’s wording and which passage supports the current filing.

Check headings, tables, labels, and footnotes. One section may contain a current figure while another still shows an older figure.

Use a label such as “mixed page evidence” when no single passage explains the answer. Do not classify the result as stale or a parsing error unless the captured text supports that conclusion.

Which source should the IR team update first?

The IR team should first review a source that contains the disputed claim, appears in a captured answer, and conflicts with an exact filing quote. A company-controlled page may come first when it meets those conditions.

Rank the captured sources with these questions:

1. Does the page contain the disputed claim?
2. Does the page appear in more than one documented capture?
3. Does the same claim appear across several engines?
4. Can the company edit the page or request a correction?
5. Does the claim concern an important investor question?
6. Does the filing provide an exact quote for comparison?

This ranking applies only to the team’s captured answer set. It does not show which websites AI engines cite most across public companies.

Can the team rank which websites AI engines cite most?

The team cannot make a general citation ranking without a defined sample, prompt set, date range, and capture method. The evidence supplied for this article does not support such a ranking.

An IR team can count cited URLs within its own documented captures. Any report of that count should state the engines, prompts, dates, capture method, and sample size.

Do not turn an internal count into a market-wide claim. The Investor Answer Audit sample for fictional Acme Inc. shows a source record tied to filing evidence rather than a general ranking.

Who can check AI answers about a company against SEC filings?

MS House can check what ChatGPT, Google AI Overviews, Google AI Mode, and Microsoft Copilot tell investors after an earnings release. The Investor Answer Audit compares each captured answer with the company’s SEC filing.

The audit records the answer, visible source, disputed claim, filing passage, and suggested owner for a fix. This gives the Head of Investor Relations, CFO, or legal team a reviewable evidence trail.

The service delivers the audit within 48 hours of an earnings release. The Investor Answer Audit sample shows the format using fictional Acme Inc.

How can an IR team handle repeated investor questions about wrong figures?

An IR team can use the same source-tracing record for figures that investors report from AI chatbots. The team should first capture the engine’s answer directly rather than relying only on an email summary.

Compare the captured claim with the relevant Form 8-K exhibit, Form 10-Q, or Form 10-K. Then identify whether the visible source is stale, unrelated, or correctly cited but misread.

The record should show the exact answer and filing quote. It should not claim that the engine caused the investor’s question.

How can source tracing support a public campaign review?

Source tracing can document what the selected AI engines display during a public campaign. The review should separate claims about company results from claims about the campaign.

Capture each answer, visible citation, date, interface, and disputed statement. Compare financial claims with the relevant SEC filing. Preserve other cited passages without treating them as filing evidence.

The record can show what shareholders could see in the captured interfaces. It cannot reveal every hidden source or prove why an engine produced the answer.

How should the team document a later re-check?

A later re-check should repeat the original prompt under conditions that are as similar as practical. Store the new answer beside the original capture.

Record:

• The original and new capture dates
• The engine and product interface
• The exact prompt
• The signed-in state
• The complete answer text
• Every visible citation
• The source page change between captures
• The new evidence label

Do not delete the original answer after a source changes. The two captures show whether the answer, citation, or both changed. The cited guidance does not set a waiting period, so re-check after the cited page changes.

What counts as a complete re-check?

A complete re-check gives the team enough evidence to compare the new capture with the original. The answer does not need to improve for the re-check to be useful.

Use one of these result labels:

1. Corrected: The answer now matches the filing.
2. Citation changed: The visible source changed.
3. Answer changed: The wording or figure changed, but the citation did not.
4. No change: The answer and citations remain the same.
5. Unsourced: The new answer has no visible citation.
6. Inconclusive: The test conditions changed too much for a reliable comparison.

These labels separate observed changes from assumptions about cause.

Does a changed source prove that the edit fixed the answer?

A changed source does not prove that an edit caused the answer to change. Source tracing records the sequence of events but does not establish cause.

Keep the original capture, changed page, later answer, and later citations together. Record the result without claiming that the page edit produced it.

A complete record ends with the captured answer, visible citation, cited passage, filing quote, date, and evidence label. The methodology page explains how these records fit into the wider audit process.

FAQ

How do I find which websites feed wrong financial data to AI chatbots?

Capture the answer and every visible citation from ChatGPT, Google AI Overviews, Google AI Mode, and Microsoft Copilot. Compare each cited statement with the relevant SEC filing. Treat the links as visible evidence, not proof of every source used to create the answer.

What if an AI answer about my company has no source link?

Record the answer as unsourced for that capture. Save the prompt, answer, engine, date, and session details. Check the other engines, but do not assign a source without a visible link or other direct evidence.

Which engine cites the most financial sites?

The evidence reviewed for this article does not support a general ranking. An IR team should count citations only within its own documented prompt set, engines, dates, and captured answers.

How long should I wait before re-checking a source?

The cited guidance does not set a waiting period. Re-check after the cited page changes. Keep the prompt, engine, account state, and other test conditions as consistent as possible.

Does source tracing fix a wrong AI answer?

No. Source tracing identifies the visible citation, disputed statement, SEC filing evidence, and capture date. Changing a source page and testing the answer again are separate steps.