ChatGPT shows incorrect or outdated information about your company for four ordinary reasons: it is answering from memory of an older version of you, it is reading live pages that are themselves out of date, it has you mixed up with a similarly named business, or your own records disagree and it picked one. None of those is a judgment on your work, and each one needs a different repair, which is why the diagnosis has to come before the fix.
You probably arrived here the way most owners do. You typed your own company name into an assistant on a quiet afternoon, half curious, and the answer came back almost right. The old suite number, a service you stopped leading with two years ago, a partner who left in 2022. Almost right is its own kind of wrong, because the buyer reading it has no way to know which half to doubt.
Nothing in that answer is the machine forming an opinion about you. It's a machine repeating what it was able to check, from sources that stopped keeping up. The uncomfortable part is the delivery: it reads the stale version in exactly the same confident voice it uses for the current one, and your buyer cannot hear the difference.
So here is the plain version. The four causes in the order they actually occur, a fifteen minute triage that tells you which one you are looking at, and what each one genuinely needs. Find the cause before you buy the cure.
The Four Ordinary Reasons an Assistant Gets You Wrong
Almost every wrong answer about a real company traces to one of four causes. They are not equally likely for every business, and they leave different fingerprints, which is what makes them sortable. A fifth case exists and is worth naming: sometimes an assistant simply invents a detail with nothing behind it, which is why the triage below asks it to show you its sources.
One. It is remembering an older you. A model is trained on data collected up to a fixed date. OpenAI publishes that date for every model it ships as a knowledge cutoff, and its GPT-5.6 models list February 16, 2026. Cutoffs differ by model and move each time a new one ships, so the honest read is not a single date but a lag: whatever an assistant knows from memory is months behind the day you asked. Ask a question the assistant answers from memory alone and you get the version of your company that the internet described before that date. Nothing in the answer will tell you it is out of date. It just answers. That is the pattern we called the training data echo, and it is the cause with the longest tail.
Two. It is reading live pages that are stale. Assistants that search do check the web, which sounds like the fix and frequently is not, because no page announces that nobody has touched it since 2019. Google's own description of the Knowledge Graph, which governs Google's records rather than every assistant's, says its facts come from a variety of sources that compile factual information, and that a panel gets created automatically when there is enough information available on the open web. Enough is not the same as current. Every system doing this kind of retrieval favors what is well indexed, and old pages have had years to get that way.
Three. It has you mixed up with someone else. This is the cause owners least expect, and it produces the strangest answers: a service you do not offer, a founding year that is not yours, a city you have never worked in. In research published in July 2026 by the AI search platform Searchable, which put more than 13,000 prompts about London companies to ChatGPT, Perplexity and Gemini and checked the answers against Companies House records and official LinkedIn biographies, small and midsize firms had their brand names confused or misattributed in 4 percent of answers, against 0.7 percent for large companies. That is a vendor's study of one city, reported in the trade press rather than published with a full methodology, so hold the numbers loosely. The direction is the part that travels: the less distinctive your public footprint, the easier you are to mistake for somebody else.
Four. Your own records disagree, so it picked one. Machines settle disagreement by weight of evidence, not by asking you. Google keeps one Business Profile per business, and says a profile it considers a duplicate will not show on Google Search or Maps. Multiply that across listings, professional directories, social profiles and your own pages, and an assistant choosing between two versions of your address is not malfunctioning. It is doing exactly what it was built to do with the evidence you left behind. When your own sources disagree, nobody stops to ask you which one is right.
A Fifteen Minute Triage to Find Which One You Have
You can sort your own case in about fifteen minutes without buying anything. The move is not asking an assistant whether it recommends you. Ask it to describe you, then change one variable at a time and watch what moves.
Step one, five minutes. Ask with search turned off. Open a fresh chat with memory and custom instructions off, turn web search off where the assistant lets you, and ask it plainly: tell me about [your legal business name] in [your city], what do they do and who do they serve? Write the answer down word for word, with the date, the engine, and the fact that search was off. That is your memory reading.
Step two, five minutes. Ask again with search on. Same question, same wording, fresh chat, search enabled. Then ask it which pages it used, and save that list. That is your retrieval reading. The two readings together are the actual diagnostic, and neither one is much use alone.
Step three, five minutes. Open the pages it named. Read them the way a stranger would. Are they yours? Are they current? Do they describe the company you run today, or one you used to run?
Now read the pattern. Four fingerprints, four causes:
- Wrong with search off, materially better with search on. You are hearing memory. The live record is doing its job and the model has not caught up.
- Wrong both ways, and the pages it cited are real but old. The sources are the problem, and your website being correct is not enough on its own.
- The description belongs to a business that is not yours, or blends two companies into one. That is mistaken identity, and it usually shows up as a fact that is not merely stale but foreign.
- The answer changes shape between runs, or holds two versions of the same fact at once. Your records disagree, and the assistant is picking a side each time.
Three cautions before you act on any of it. The off switch is not clean everywhere: some assistants retrieve by default and give you no way to stop them, and some route through licensed listing data even when browsing looks disabled, so treat this as a heuristic rather than a laboratory. One engine is a data point, not a verdict, so run the same two questions on a second assistant before you conclude anything. And answers move by design: same question, same day, slightly different wording, different answer. Run each question twice per engine and treat the pattern as the finding rather than any single sentence.
This triage answers what an assistant knows about you. Whether it will actually name you when a buyer asks who to hire is a separate question with a separate check, run without your name in the prompt at all.
What Each Cause Actually Needs
Here is why the fifteen minutes were worth spending. The four causes take four different repairs, and choosing the wrong one is how owners spend a quarter and watch nothing move.
Stale memory. You can't reach into a model's memory and edit it, and nobody honest will sell you that. What you can do is make the current version of you easy to retrieve and easy to corroborate, so an assistant that searches has a current, corroborated account available to find, and the open web the next model learns from describes the company you actually run. This is the slowest of the four to pay off, and the only one where patience is genuinely part of the plan.
Stale sources. Repair, then retire, in that order. Correct the surfaces you control, close the ones you abandoned, and make the survivors tell one story: same name, same city, same specialty, same people. Nobody enjoys this work, and it's still the fastest of the four to show a difference, because a corrected page can be read the next time an assistant goes looking, rather than waiting for a new model.
Mistaken identity. Your job is to become harder to confuse. Use one legal name and one trading name consistently, keep the city and the specialty next to the name wherever a machine can read them, and publish your identity in structured data so nothing about who you are is left to inference. That is the plain-English case in our guide to how AI decides your business is real, and the work itself is what we build as Entity Identity.
Records that disagree. Choose the canonical version, then make every surface match it. Start where machines lean hardest: your Business Profile, your main directory listings, and your own contact and service pages. A duplicate profile is more than clutter. It is a second candidate for the truth, sitting there with your name on it.
One sentence sits underneath all four, and it is worth keeping. AI can't feel your reputation in a room. It can only work with what it can verify.
Why One Assistant Sounds Current and Another Does Not
Owners often arrive with two answers from two assistants and a reasonable question: which one is telling the truth about my company? Usually both are telling the truth about their own inputs.
The assistants make different choices about when to search, what they are allowed to reach, and how much weight to give what they find. One may answer from training memory unless you push it to look. Another searches by default and shows its sources. A third leans on licensed listing data for anything involving a place. Same question, different machines, different evidence, and neither one is broken. We mapped where each one actually looks in a companion piece on where ChatGPT gets information about your business.
Two practical consequences follow. First, don't try to fix AI. There is no single system to fix, only a record that several systems read differently, so the work is on the record and the checking is per engine. Second, do not average your results across engines into one comfortable number. If one assistant has your current specialty and two do not, you have a two out of three problem, and the average hides exactly the part you needed to see.
One more thing worth knowing before you trust a tidy-looking answer. A citation tells you where an answer came from. It does not tell you the page was right, or current, or even about you.
What to Do First
Start with the reading, not the repair. Run the triage on two engines this week and write down the date, the exact wording, the engine and whether search was on. That record is worth more in a month than any screenshot is worth today, because it gives you a before to compare against.
Then fix in this order. Correct the single most consulted wrong surface first, which for anything involving a location is almost always your Business Profile. Make your own site state the plain facts in one obvious place: legal name, trading name, city, what you do now, who you serve, who works there. Retire or correct the abandoned profiles that still describe a company you no longer run. Then wait thirty days and run the same two questions again, in the same words.
If the picture is worse than you expected, that's information rather than a verdict. Observed sources can change substantially month to month, so a single bad reading is a starting point and a repeated one is a pattern.
When you want the full version rather than the fifteen minute one, the free Recommendation Check is where to get it: your visibility scored across 109 checkpoints in eight zones, plus what five engines actually say when buyers ask who to trust in your category. No discovery call first. The diagnostic does the talking, it takes about an hour to present, and you decide what happens next. See where you stand.
Want to Learn More?
Probably Genius was built by Jacquie Baker and Christopher Shaw, who spent more than 40 combined years translating what makes an expert the best in the room for the systems that decide who gets chosen. We are a done-for-you operator rather than a dashboard: the work is mapping what you know, structuring it so machines can verify it, and building the record around it. If an assistant is describing a company you stopped being three years ago, the problem is not your reputation. It is the paperwork the machines can reach. You're probably a genius at what you do. We make sure AI gets the memo.