Ask an AI assistant who runs a company that was acquired two years ago, and there's a decent chance you'll get the name of someone who left in 2022. The answer isn't invented. It came from pages the company published itself.
TL;DR: When AI gets facts about your brand wrong, the cause is usually too many old versions of the truth sitting on your own site, not missing content. The fix is correcting the evidence chain behind the answer instead of publishing another page on top of it.
🔎 Too Much Truth Is the Problem, Not Too Little
Most brands treat AI visibility as a production problem. More authoritative pages, more FAQs, more comparison content. Carolyn Shelby's argument at Search Engine Journal is that this rarely fixes anything.
In most cases where AI says something wrong about a company, the problem isn't a lack of data. It's too much of it. The brand already has several versions of the truth on record, and the one that best matches the question isn't the current one.
The website says one thing. An old PDF says another. Product documentation still uses names the marketing team dropped two years ago, and executive bios preserve titles that no longer exist.
Some of that material was simply wrong. Most of it was accurate the day it went live and never got revisited.
Traditional search could rank several of those pages at once and let you decide which was current. AI search retrieves sources and builds a single answer out of them. That answer looks settled even when the evidence behind it isn't.
The Question Itself Picks the Winner
When someone asks an AI assistant a question, the prompt supplies most of the vocabulary used to go find supporting sources. The system may rewrite it or run a few related searches, but it's still trying to answer the question it was handed.
That matters when the question contains an assumption the person doesn't know is outdated. Someone asking who a company's CEO is assumes the company still has one. They don't know to ask who leads the brand now, or whether leadership moved to a parent company.
A search built around the company name and the word CEO will favor pages containing exactly that pairing. Old bios, press releases, and acquisition announcements all have it. A current leadership page that says SVP and general manager, and never mentions the old title, may never enter the retrieval set at all.
An AI system can't cite a source it didn't retrieve. The old answer wins because it matches the language of the question.
Shelby describes a real company she leaves unnamed. Ask an AI assistant who its CEO is, and you might get any of four former executives. None of those answers were made up, because all four held the title at some point and accurate history pages still say so.
The current structure uses different words entirely. The company's team page names its senior leader as SVP and GM, and the parent organization's pages describe executives over a much larger business group. None of it is phrased in a way that answers the question as asked.
It gets messier. The current leader's profile carries the SVP and GM title at the top, but introductory copy on some pages still calls that person CEO. The conflicting signal is sitting on the same profile.
Sometimes the strongest content strategy is fixing the old content you forgot you published.
🛠️ Three Fixes That Change What Gets Retrieved
Write the Bridge Sentence Nobody Published
Publishing a new page isn't the same as correcting the record. Accurate information written in vocabulary nobody searches with stays invisible to the question people actually ask.
Bridge content connects the dead term to the current one inside the same sentence. Shelby's example reads roughly like this: following the acquisition, the company no longer has a standalone CEO, and Jane Smith now leads it as SVP and general manager within the parent group.
That sentence names the old role, explains why it ended, and gives the current equivalent. It hands retrieval systems a source that matches CEO without falsely assigning the title to anyone.
The same move works when products get renamed, plans get retired, certifications expire, or service areas change. Don't assume people know the new name well enough to search for it.
Trace the Evidence Chain Before Publishing Anything New
When you find a wrong answer, follow the citations and searches behind it. The goal is finding what's producing the answer, not burying it under something freshly published.
Pages you own can be updated, consolidated, redirected, or annotated. An old announcement shouldn't be rewritten to pretend history went differently, but it can carry a clear date, a status note, or a link to current information.
Current bios shouldn't keep obsolete titles in their intro copy. PDFs and sales materials that no longer represent the business can be retired or labeled as archival.
Third-party coverage needs more judgment. You can't ask a publication to rewrite an accurate five-year-old story because the company changed later. You can make the current explanation easy to find and ask partners and directories that are supposed to be current to fix their records.
Audit Your Claims, Not Just Your URLs
A content inventory records URLs, titles, traffic, and maybe conversions. A brand claim audit records the factual assertions those pages make, plus the words people are likely to use when they ask about them.
What a content inventory records
URLs and page titles
Traffic and rankings
Conversion data
What a brand claim audit records
The question or prompt a person is likely to use
Any outdated assumption sitting inside that question
The old term and its current equivalent
The approved current fact and its canonical public source
Other owned pages, PDFs, feeds, bios, or profiles carrying older versions
The sources currently cited in inaccurate AI answers
The action required: update, annotate, consolidate, redirect, retire, or bridge
The team responsible for maintaining that claim
Don't stop at HTML pages. Media kits, downloadable sales materials, help-center content, schema values, product feeds, app-store listings, speaker bios, job listings, and forgotten subdomains all make claims about your business.
The goal isn't identical wording everywhere. Different audiences need different levels of detail. The underlying facts just need to resolve to the same answer.
📊 A Mention Isn't a Win If the Answer Is Wrong
Most AI visibility reports track whether a brand was mentioned, cited, or included for a set of prompts. Those are useful observations. On their own, they can also create false confidence.
For the prompts that matter, read the answer itself. Is it accurate, and is it current? Does it answer what the person actually wanted, or does it just accept a false premise buried in their question?
When an answer is wrong, inspect the citations and reproduce the likely searches before assuming the model made something up.
A brand mention isn't a win if the answer names a former executive, quotes an old price, or credits the current product with a feature you discontinued.
Hit reply
What's one page on your own site that still says something about your business that stopped being true?
💭 One Thing I'm Thinking About
The uncomfortable part of this is how much of it is self-inflicted. Nobody outside your company published the bio with the wrong title. You did, and then everyone moved on.
Our industry has a strong bias toward making new things. Auditing old PDFs and emailing a partner directory doesn't produce a chart anyone wants to present in a quarterly review.
But the brands AI systems describe correctly will be the ones whose own record agrees with itself. That's maintenance work, and it may be the highest-leverage thing sitting untouched in most content plans.
Did You Know? In the company Shelby describes, asking an AI assistant who the CEO is could surface any of four different former executives, and every one of them really held the job.
Till next time,
