Reproduce the wrong claim before deciding what caused it or what to fix.
When an AI answer gives old pricing, misses a new feature, uses a previous product name, assigns the wrong customer segment, or describes an integration incorrectly, preserve the exact question, claim, answer, displayed sources, run conditions, and date. Then write the approved current truth and its official source. Repeat the observation before treating it as stable. An outdated third-party page beside the answer is worth investigating, but its appearance does not prove it caused the error. Prioritize the discrepancy by buyer risk, correct current owned information and obsolete evidence you control, request important external corrections where appropriate, and retest. Do not promise that a page update will immediately change a model response.
The discrepancy itself can be documented.
You can show that a recorded answer conflicts with current approved pricing, capabilities, positioning, plans, integrations, or naming. That is enough to create an evidence-backed accuracy issue.
A displayed source rarely explains the full causal path.
A cited old page, missing owned detail, or contradictory description is evidence to investigate. Do not state that it caused the answer unless the causal relationship is independently established.
Turn an alarming screenshot into a reviewable discrepancy record.
What is wrong?
Write the exact inaccurate claim and the current expected truth without paraphrasing away the discrepancy.
Can it be reproduced?
Repeat the exact question under recorded model, provider, time, language, location, and session conditions.
Is it branded or unbranded?
Separate a direct company-description failure from an error inside a category or recommendation answer.
What evidence accompanies it?
Retain the answer text, displayed URLs, and which claim each source appears to support.
Is the current truth published?
Check the canonical owned product, pricing, integration, security, and positioning pages.
Is an old third-party page present?
Record the page and its date, but do not assume its appearance proves causal influence.
What can we control?
Identify obsolete owned pages, unclear current copy, genuinely appropriate structured information, and correction requests you can make.
How will we retest?
Freeze the original question and acceptance criteria before publishing a correction.
Keep the wrong claim and the current source of truth in the same review packet.
| Evidence | What to retain |
|---|---|
| Exact question | Preserve the wording that produced the inaccurate or outdated description. |
| Exact inaccurate claim | Quote only the bounded claim under review, not an invented summary. |
| Current expected truth | State the approved product, pricing, availability, integration, positioning, or plan information. |
| Answer text | Retain enough surrounding text to understand how the claim was presented. |
| Sources and citations | Save displayed URLs, visible dates, and the claim each source appears to accompany. |
| Affected run | Record provider, model, language, location where relevant, session conditions, and timestamp. |
| Repeated observation | Keep the results needed to tell a recurring risk from one answer. |
| Official source of truth | Link the current approved owned page or internal product record used to verify the discrepancy. |
Prioritize by buyer and business risk, not by how embarrassing the screenshot looks.
P0 — Immediate business risk
Incorrect pricing, major product availability, security, or core-capability claims that could materially mislead a buyer.
P1 — Material fit risk
Wrong target customer, category positioning, integration support, or important feature claims that could exclude or misqualify the product.
P2 — Descriptive cleanup
Minor terminology, old naming, or secondary descriptive differences with limited effect on a buying decision.
Fix the current truth you control, then measure what actually changed.
- 01
Confirm the current truth.
Assign an accountable product, pricing, security, or marketing owner to approve the expected claim.
- 02
Correct the canonical owned page.
Make the current product information explicit where buyers and reviewers should verify it.
- 03
Retire contradictory owned evidence.
Update, redirect, noindex, or clearly date obsolete owned pages according to the real publishing need.
- 04
Address important third-party discrepancies.
Request corrections where appropriate; RecoProof cannot edit or compel an external publisher.
- 05
Use structured information only where it is truthful and appropriate.
Do not add schema as a guarantee or substitute for clear current product copy.
- 06
Retest the original question.
Repeat under recorded conditions and preserve whether the wrong claim changed, persisted, or varied.
An old integration page appears beside an outdated answer.
Imagine a buyer asks whether Vendor A integrates with a current CRM. The answer says no and displays a two-year-old comparison page, while Vendor A's approved integration directory now says yes. The team should preserve the answer and source, verify the integration with the accountable product owner, make the canonical owned evidence unmistakable, request a correction from the comparison publisher where appropriate, and retest. The old page is a useful lead, not automatic proof of causality. This is illustrative, not a RecoProof customer result.
Wrong representation is not the same problem as missing commercial inclusion.
If the answer describes your SaaS accurately but omits it from unbranded category questions, use the separate diagnosis for why the company can rank on Google without entering ChatGPT recommendations. This page is for factual or temporal discrepancy, not shortlist position.
Establish what was said, where it appeared, whether it repeats, and what deserves action.
RecoProof retains the approved question, answer, displayed source evidence, model and run conditions, repetition, and discrepancy context. Confidence and human review help prioritize the issue before a team changes product copy or raises an external correction request. RecoProof cannot edit third-party sites, force a model update, or prove that a displayed source caused the wrong claim.
Review the measurement methodology and the fictional sample report before deciding whether the full Recommendation Audit fits the risk.
Current OpenAI guidance used for accuracy and source-review claims.
Last checked August 27, 2026. OpenAI advises that ChatGPT can produce incorrect or misleading output and that search results and citations can be incomplete, outdated, or wrong. That supports verification and source review; it does not establish which page caused a specific response.