A buyer sends your team a screenshot: an AI assistant says that single sign-on is included in every plan. Your company offers it only on the Enterprise plan. The mistake could send a buyer into the wrong sales conversation or make your pricing page look inconsistent.
The tempting response is to publish more content immediately. A better first move is to identify the exact claim, establish the correct fact and find out what information the assistant could have used. You may uncover a page worth fixing. You may also find that your pages were clear and the model still got it wrong.
A screenshot is a starting point, not a diagnosis
Save the full question, the answer and the sentence that is wrong. Record the model, date, language and market, plus any links shown alongside the answer. Note whether the assistant searched the web. A cropped screenshot that says only “SSO is included” loses the context needed to investigate the error.
Then ask the person who owns the fact to verify it. For an SSO claim, that might be the product owner and the current plan documentation. Write the verified statement beside the AI claim:
| What you observed | What needs verification |
|---|---|
| “SSO is available on all plans.” | Which plans include SSO today? Is there a separate add-on or contract condition? |
Until that second column is settled, the team has an apparent discrepancy, not a safe replacement sentence for the website.
Follow the claim to its sources
Open every cited link and read the passage that could support the statement. Check its publication or update date. A citation is useful evidence of what appeared near an answer; it is not proof that the source says what the assistant claims. OpenAI advises readers to check cited sources because search results and citations can be incomplete, outdated or incorrect.
There are several different findings, and each calls for a different response:
| Finding | Useful next step |
|---|---|
| Your own pricing and product pages disagree | Resolve the contradiction with the fact owner, then update both pages. |
| Your page says “SSO available” without naming the plan | State the plan and any conditions next to the feature claim. |
| An outside directory repeats an old plan description | Keep your primary page current and request a correction from that publisher. |
| The answer has no usable source, or its citations do not support the claim | Record the model error. Do not invent a website problem to explain it. |
If the page you want to improve is hard to access, check that separately. For ChatGPT search, OpenAI says a site must allow OAI-SearchBot to crawl it to be eligible for inclusion. Google has its own Search requirements for AI features. Passing a technical check does not guarantee that either system will use your page.
Sparse or ambiguous information can make a specific company fact harder to establish. Yet that is only one possible explanation. OpenAI's research on hallucinations describes how models can produce plausible false statements even when the question appears simple. Do not claim to have found the cause of one answer merely because a page could be clearer.
When a comparison invents a limitation
Imagine a procurement manager comparing two contract platforms. Their shortlist requires a Salesforce integration and EU data residency. Product X supports both, according to facts its team can verify internally. Its public page, however, offers only “easy integrations” and a “secure cloud.” Product Y names Salesforce and EU data residency on its website.
An assistant might answer: “Y meets both requirements. X does not integrate with Salesforce.” The second sentence is false. It could also remove X from the shortlist before anyone from Sales has a chance to explain the product.
For the buyer, the distinction is consequential: the capability exists, but the public description does not establish it. Before editing, inspect any source behind the answer. It might be an old directory listing, a misunderstood product page or no usable source at all. Then have the integration owner verify the exact Salesforce support and the data residency conditions. Publish those facts on the relevant product or documentation page, with their plan, region and any limits. A buyer should be able to check the same criteria used in the comparison.
This approach does not assume X is the best product overall. It corrects a false statement about one buying criterion. Whether X belongs on the shortlist depends on the buyer's other requirements and the evidence available for them.
Correct the page a buyer would trust
Consider a B2B software company whose feature page says, “Enterprise-grade security, including SSO.” Its pricing page lists SSO under “Advanced features” without saying which plan includes it. The team has confirmed that SSO is part of Enterprise only.
The useful edit is not another article repeating “we support SSO.” It is an unambiguous statement on the page where a buyer checks availability:
Single sign-on (SSO) is included in the Enterprise plan. See the plan comparison for current availability and implementation requirements.
The team should use that wording only after checking the actual offer. If availability varies by contract, region or date, the public text needs that boundary too. Link the plan comparison to the relevant documentation and remove conflicting copy elsewhere. A person should be able to answer the pricing question without consulting an AI assistant.
The same method applies when an assistant confuses a founder with another person or attributes a product to the wrong company. Name the person, company and relationship clearly on the relevant profile or product page; use a date where the relationship can change. A single well-maintained source is more useful than several vague repetitions.
Use TextRank to turn the observation into an edit
TextRank starts with a page and a buyer question. It keeps the answers from supported models and their available sources separate, then compares those materials with passages on your page. The result is a prioritized content task tied to the question, source and passage. The audit method explains what it compares.
For the SSO example, the task might be: “On the pricing page, name the plan that includes SSO, link to the current feature documentation and have Product approve the conditions.” TextRank can help locate and describe that gap. It cannot confirm a private contract term on its own or force an AI assistant to change its answer. Your team remains responsible for the fact and for publication.
If an answer provides no usable sources, the trail is shorter. Record that limitation. You can still check whether your page gives buyers a clear answer, but you cannot reliably attribute the model's error to a particular source.
Check what changed after publication
Keep the original prompt, answer, sources and page wording. After the edit is live, repeat the question with the same language and market, and compare results for each model separately. Log the new date and any change in search behavior or cited sources.
Judge two outcomes separately: Does the page now state the verified fact clearly? And what did the model answer this time? One retest cannot prove that your edit caused a change in the model's response. It can show whether the website problem you identified was fixed and whether further investigation is worthwhile.
If you have a specific page and question to investigate, start a TextRank audit. Begin with the claim that matters to a buyer, then assign the resulting content task to the person who can verify it.
Frequently asked questions
Can changing our website make an AI system correct its answer?
A clearer, accessible source can address a gap you control, but it cannot guarantee that a model will use it or stop making mistakes. Compare later answers with the original observation.
What if the false claim comes from a third-party site?
Save the cited page and the incorrect passage. Keep your own primary source accurate, and ask the third-party publisher to correct its page when appropriate.
Should we publish a new article for every false claim?
No. First check whether the fact belongs on an existing product, pricing or documentation page. Add a new article only when it answers a distinct reader question.