AI Gets Your Business Wrong? More Content Won't Fix It

A customer forwards you a screenshot. They asked ChatGPT about your clinic, and the answer named the doctor who sold the practice years ago, sitting right next to a service you no longer offer. You run the same prompt yourself and get a slightly different wrong answer. Your first instinct is to write a page that sets the record straight.

That page will not fix it. Neither will deleting every old page you can find, which is the other half of the standard advice. The wrong answer survives because your current truth is written in vocabulary that no longer matches the question your customer actually types, so a machine looking for the best match to their words keeps landing on the version that still uses them. The fix is a claim-by-claim audit that maps the old terminology to the current reality, one line at a time, with a bridge sentence that connects them in language a model can follow.

The wrong answer is not a glitch, it is an older page winning a match

Think about what your business has published over the years. There is a pricing page you replaced. Two directory listings you filled in during a slow week. Somewhere there is a press release announcing the service you have since renamed, and a Facebook About section nobody has touched since the last rebrand. You have three review-site profiles, and one of them you can no longer log into. Every one of those is a statement about your business, and none of them retired when you changed your mind.

When an AI system builds an answer, it is not asking which of your pages is newest. It is assembling the version that best fits the words in the question. Your newest page is one candidate among many, and if it uses different language than the question, it loses to something you wrote years ago.

I know this shape of problem because I keep running into it on my own site. I built a local command center that audits my archive for exactly this class of failure: pages competing for the same intent, meaning two of my own pages chasing the same question; schema that disagrees with the visible page, meaning the hidden data a machine reads says one thing while the words a reader sees say another; and link graphs that point at things I no longer say, meaning my own links still send people to statements I have retired. It runs against 100 plus pages, and the recurring finding is never a missing page. It is two pages that both sound true and only one of which still is. The work that fixed my own archive was subtraction and reconciliation, not publishing.

Carry one example through the rest of this. It is a composite, built from patterns I see repeatedly rather than one real client. A weight clinic used to sell a "slimming program." Two years ago they brought in a physician, changed the protocol, and renamed it "medical weight management." Everything about the new version is better: better outcomes, higher price, a different kind of patient. The old name is gone from the website.

It is not gone from the internet. It sits in an old blog post on a local lifestyle site, in the clinic's own Instagram bio, in an old price list a directory scraped and cached, and in the reviews where patients wrote the words themselves. When someone asks an assistant about slimming programs in that city, the clinic gets described with a name it retired, a price it no longer charges, and a promise the physician would not make. The cost is quiet. Nobody calls to complain about an answer they never saw was wrong. They simply book somewhere else, or they arrive expecting the old price and the front desk spends the consultation renegotiating instead of converting.

Publishing the correction adds a competitor to your own truth

The reflex is understandable. Something is wrong out there, you own a website, so you write. Now there are two current pages, four stale ones, and a model choosing between six versions instead of five.

Volume was the lever for a decade. More pages, more surface area, more chances to be found. In a retrieval system that picks by fit rather than by recency, every additional version you publish is one more thing that can be picked, and you do not get to vote.

The advice I see on this problem falls into three camps, and each one is partly right. Concentrate more structured ground truth on your own site. Syndicate the identical correction across twenty or more third-party sources. Diagnose the cause and make every surface match. None of them tell an owner what to do on the Tuesday morning after the screenshot arrives, and none of them address why the correction failed to take in the first place.

This is a different failure from the one most content audits are built to catch. Pages that slowly lose rankings need a refresh-or-prune decision, and I have written the process for deciding whether to refresh or prune decaying content separately. That piece is about traffic decay. This one is about a machine confidently repeating a fact that stopped being true, which can happen to a page that is performing beautifully.

The question's vocabulary picks the winner, not your publish date

Shelby said it in one line: "publication is not the same as correction, and accurate information written in the wrong vocabulary may remain invisible to the question people actually ask."

Apply it to the clinic. Their new page says "medical weight management," "physician-supervised," "metabolic assessment." Correct, current, well written. The patient types "slimming program near me" or "how much is the slimming package at [clinic]," because that is the phrase they heard from a friend three years ago and phrases stick to people longer than they stick to businesses. The new page contains none of those words. That old lifestyle article contains all of them, plus an old price.

Nothing malfunctioned. The system matched language to language and served the closest fit. Your rebrand happened inside your organization on a specific date. In the world, the old name keeps circulating in reviews, forums, screenshots, and other people's writing, and it will keep circulating for years after you stop using it.

Deleting the old page removes a source. It does not remove the question. Now the words your customer uses appear nowhere you control, and the model reconstructs an answer from whatever fragments remain, which is worse than a stale page you edited.

Run the audit claim by claim, then write the bridge sentence

A content inventory lists pages. A claim audit lists statements about your business that can be true or false, which is what an AI answer is actually made of. Seven columns, one row per claim. Build it in a spreadsheet, on a phone if that is what you have.

Start with the claims that cost money when they are wrong: who owns or leads the business, what you actually sell, where you are located, what it costs, who you serve, what you no longer do. A short sheet you finish beats a long one you abandon.

The bridge sentence is the piece people skip, and it is the piece that does the work. It keeps the retired vocabulary present on a page you control, attached to the correction, so the words your customer types still lead somewhere you wrote. Delete the old term everywhere and you hand that query to strangers.

If this is going to your agency rather than your own afternoon, the question to ask is narrow enough that you will know immediately whether they have done it: "Show me the claim-level sheet with the prompt, the retired term, the canonical page, and every surviving copy you found off our site." A list of pages updated is not an answer to that question.

Count correct answers, not appearances

Most AI visibility reporting counts mentions. Your brand appeared in more answers this month than last, the line goes up, everyone nods. That number goes up when the model repeats the name of the doctor who left, because the tool counting mentions cannot read. An answer that names your former owner, quotes a retired price, or sends someone to a location you closed is a mention and a loss at the same time.

Score accuracy instead. Take the prompts from column one, run them, and mark each answer correct, partly correct, or wrong on the specific claim you were auditing. That is a scoreboard that moves for real reasons. The mechanics of running the checks are laid out in my walkthrough on how to check whether AI actually recommends your business, and the distinction that matters most is covered in the piece on being mentioned versus being believed.

Here is what one row looks like on the clinic's sheet. Column one holds the prompt: "how much is the slimming program at [clinic]." Column two holds the claim being audited, which is the price. Column three onward is one column per month.

Month one, the answer comes back wrong. The assistant quotes the retired program name and the old price, and when you ask it where that came from, it points at that old lifestyle article. You log the grade, you log the source it cited, and you open the row.

Then you do the work. You add a bridge sentence on the clinic's own service page saying the slimming program is now medical weight management, in the customer's words, not the clinic's. You correct the directory listing you still control.

Month two, you run the same prompt and grade it again. Partly correct. The assistant now uses the right service name, which tells you the bridge sentence landed. It still quotes the old price, and this time it cites a different directory, one you have not corrected yet. So you write that directory into the row as the next action and leave the row open.

That is the whole discipline. One prompt, one claim, one grade a month, and a row that stays open until the grade says correct and stays there. You are not chasing a score. You are closing sources one at a time and watching the answer move.

On timing, be patient in a specific way. Sources get recrawled on their own schedules, third-party sites move when they feel like it, and no vendor publishes the lag. Monthly re-runs of a fixed prompt list will show you movement without turning this into a daily anxiety. A claim that has not budged after several cycles means you missed a copy, not that the fix failed.

Run the audit once and every wrong answer traces back to a decision your business made and never finished communicating: the rename nobody told the directories about, the location change that stopped at the website, the price that moved in the booking system and nowhere else. The machine is not misrepresenting you. It is quoting you accurately, from an earlier draft of your company that you forgot to retire.

If you want a second pair of eyes on which claims about your business are still circulating wrong, and which ones are costing you bookings rather than just pride, book a call with me and bring the screenshot that started this. We will build the first six rows of your claim sheet together and you will know by the end of the call whether this is an afternoon of work or a quarter of it.

Frequently Asked Questions

How do I get ChatGPT to stop using my old business name?

You cannot edit the answer directly, so you work on what it reads. Find every place the old name still appears, starting with sources you control like your own site, your social bios, and your directory listings, then move to third-party pages and request updates where you can. On your own site, keep the old name present in one bridge sentence that states the change and the current name, so the query still lands somewhere accurate instead of somewhere random. Then re-run the same prompt monthly and log whether the answer changed.

Should I just delete the old pages with the wrong information?

Deleting removes a source but not the question people ask. If customers still search using the old term and no page you own contains it, the answer gets rebuilt from fragments you have no control over. The better move on your own site is usually to correct the page and add the bridge sentence, or redirect it to the page that owns the current truth. Save outright deletion for thin pages that have nothing worth keeping.

How long until AI answers reflect the corrections I make?

There is no published schedule, and it varies by how quickly each source you fixed gets re-read. Your own site tends to move before third-party listings, and pages you had to ask someone else to change move last. Check monthly using the exact same prompts so you are comparing like with like. If a claim has not moved after several checks, assume there is still a stale copy you have not found rather than that the correction failed.