Your old crises are still showing up on AI. Here’s what to do about it.
A study of how AI engines describe the Fortune 100 found that serious negatives from years or decades ago still show up. Here’s what in-house teams can do about it.
Sam Michelson is the founder and CEO of Five Blocks.
Anyone who has applied for a mortgage knows how a credit report works. A late payment from six years ago can still be sitting there, even if you’ve paid every bill on time since. But a credit report at least has rules. Most negative marks fall off after seven years.
AI has no such rule. When an investor, a reporter or a job candidate asks ChatGPT or Google to tell them about your company, the answer draws on everything the engine can find. It doesn’t care that the story ran in 2009 and the news cycle moved on years ago. If it’s in the file, it can show up.
We wanted to know how often that happens. So my team used our AIQ platform to prompt eight AI engines, including ChatGPT, Gemini, Claude and Google’s AI Mode, one simple question, “Tell me about [company],” for each Fortune 100 company. We asked once a day on each engine from September 6 to October 6, 2026, which gave us close to 25,000 answers for our study.
The old entries are still on the report
The finding that stayed with me most: of the serious negatives that named a year (lawsuits, scandals, safety problems, regulatory trouble), 43% were about events from 2020 or earlier.
Some go back a long way. The engines still bring up the Exxon Valdez spill, which happened in 1989. They bring up a 1998 Medicare fraud case at HCA Healthcare, General Motors’ 2009 bankruptcy and the 2016 Wells Fargo unauthorized accounts scandal.
Most of those companies paid, settled, reorganized or replaced leadership years ago. The crisis ended. The record of it didn’t.
The study doesn’t tell us why a given story keeps coming back, so here is my read, from years of working with communications teams at large companies. An engine can only tell the rest of the story if the rest of the story is somewhere it reads. If the resolution was never written down in a place the engines trust, the old coverage is the most complete account in the file. So that’s what gets repeated.
Recent events, by contrast, are rare. Only 7.7% of answers mentioned anything from the previous three months. When they did, those answers were about twice as likely to include a serious negative, 11% compared with 5%.
The engines don’t agree on your problems
In our study, we found that the major AI engines mostly agree on the basic facts about a company. They don’t agree on the problems. About a third of the negative points we found were raised by only one of the eight engines. Even two Google products split: AI Mode raised a serious issue in 12% of its answers, about 11 times as often as AI Overview (1.1%).
So if your team checks one engine and finds nothing, that’s good news about one engine.
What I’d tell an in-house team
You can’t delete an entry from this kind of report. What you can do is add to the file, so that newer, accurate information is easy for the engines to find.
- Pull every report, not just one. Start with the engines your stakeholders use most, and ask what they would ask. “Tell me about [company]” is the baseline. Then try the questions that hurt: “Is [company] safe?” “Should I work at [company]?” “[Company] lawsuit.” Write down each negative and which engine said it.
- Find the source behind each old story. Most engines show their citations. For every old issue, see which page the engine is reading. It might be a decades-old news story, a reference site or one of your own pages that nobody has touched in years. If it’s your own page, that’s the easiest fix you have.
- Put the resolution on your own website. This is the source you control, and the engines lean on it heavily: 82% of the answers we collected cited the company’s own site. If the settlement, the reform or the safety record since then isn’t on your site in plain language, the engines have nothing to balance the old story with.
This is the step legal may push back on, and that’s fair. Nobody wants a new page that makes an old crisis easier to find. Keep it to facts already on the public record, and put it where that history belongs, such as a governance, safety or company history page, not the homepage.
- Keep Wikipedia accurate, by its rules. Wikipedia was the most cited source outside company websites, though engines differ a lot: 92% of Google AI Mode answers cited it, against 3% for ChatGPT. Wikipedia strongly discourages companies from editing their own articles. Disclose who you are and suggest changes on the article’s talk page, backed by independent sources. Go in knowing that a request puts the article in front of editors who may add to the criticism section as well as fix it, and that answers can take weeks. Bring sources strong enough to hold up. After Wikipedia, the next most-cited sources were Yahoo Finance and Britannica, so check what they say about you too.
- Check again, on a schedule. AI answers change, and the engines update often, so one check is a snapshot. Run the same questions every month or quarter and keep the answers. That’s how you see an old issue moving up, or a new one arriving, before your CEO asks about it.
No one is going to give your company a seven-year rule. The old entries will stay in the file. What you can change is what sits next to them: accurate, current information in the places the engines already read. It’s slow work, but it’s the part of the record you control.
