When search results and AI answers keep repeating an outdated or wrong fact, like an old CEO or a discontinued product, the problem is usually the source record, not the model: fix owned pages first, request corrections from third-party publishers, publish the missing current record, then retest search and AI separately since they refresh on different schedules.
Everything-PR's Reputation Management coverage emphasizes that reputation now lives across earned media, search, structured data, social sources and AI-generated summaries. Correcting an outdated or false narrative therefore starts upstream.
Step 1: Define the Exact Error
Do not begin with "AI is wrong." Write the factual correction in one sentence. Identify what is outdated or false, what is currently true, when it changed and which primary evidence proves it.
Step 2: Trace the Sources Repeating It
Run the relevant searches and AI prompts across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, several times each. An error that appears once in one engine is often noise. An error that repeats across engines or runs usually traces back to a source problem.
Then record the pages cited or repeatedly surfaced. When an engine does not cite, ask a follow-up such as "What sources support that claim?" to surface the underlying reference. Look for a common source: an old company bio, outdated directory profile, stale press release, archived executive page or third-party article.
The goal is to identify the evidence chain, not to argue with the final answer.
Step 3: Fix Owned Primary Sources First
Update the pages the organization controls: leadership bios, product pages, location pages, newsroom material, investor-relations pages and structured data. Preserve dates and historical context where they matter.
A correction should make the current state unmistakable without pretending the earlier state never existed.
Step 4: Request Corrections Where the Source Is Wrong
When a third-party publisher contains a factual error, use its correction process. Provide concise evidence and distinguish an error from information that was accurate when published but is now old.
Journalists and publishers are more likely to correct a verifiable factual problem than to rewrite legitimate historical coverage because a company dislikes how it reads today.
How Hard Is It to Fix Each Type of Source?
The difficulty of a correction depends on the authority of the source repeating the error, so score each source before choosing a tactic.
| Source type | Typical fix | Difficulty |
|---|
| Tier-1 outlet with the fact wrong | A correction request to the publisher, plus follow-on coverage that states the correct fact at equal or higher authority | Hardest |
| Mid-tier outlet or Wikipedia | PR outreach for a correction, or a well-sourced Wikipedia edit with any required conflict-of-interest disclosure | Faster |
| Low-authority blog or scraped content | Stop trying to delete it and publish authoritative content that outranks it | Hard to remove, easier to displace |
Step 5: Publish the Missing Current Record
Sometimes the old story persists because the new story is poorly documented. Publish current facts in formats other sources can cite: executive announcements, FAQs, dated updates, regulatory filings, research, case results and clear company reference pages.
The correct fact does not have to delete the wrong one. It has to outweigh it, through sources the engines already weight heavily: tier-1 earned media, a well-cited Wikipedia article, and an owned fact sheet, press kit or leadership bio that states the fact in a format a retrieval system can read cleanly.
Step 6: Retest Search and AI Separately
Search-result composition and AI answers do not update on the same schedule. Retest the original queries on a consistent cadence and record whether the source mix, wording and citations change. Engines that lean on live retrieval, such as Perplexity, tend to reflect new sources sooner than engines that rely more on older training data.
EPR's AI Visibility Audits guide recommends fixed prompts and repeatable measurement precisely because one test is only a snapshot.
What Not to Do
- Do not create dozens of thin pages repeating the correction.
- Do not ask publishers to remove accurate historical reporting merely because it is unfavorable.
- Do not change dates to make old material appear new.
- Do not assume schema alone will override contradictory source material.
- Do not measure success from one favorable AI answer.
- Do not expect individual correction requests to AI companies to scale, because the engines generally do not run manual review queues for company factual disputes outside clear policy violations.
- Do not treat a lawsuit against the model as the primary repair channel. The legal route is unsettled, as covered in When AI Generates Your Crisis.
No communications firm can guarantee specific outputs inside third-party AI systems. The discipline is shaping the inputs the engines retrieve from, not directing the engines themselves.
