Find the pages behind the mix-up, then state the same identifiers (full name, company, city, field) on your site and every profile, with the same photo and bio, and link those pages to each other so a search can separate you from the other person. You cannot edit the model or the other person's pages, so correct what you control and measure the change with a fixed list of questions, each asked three times.
Why does ChatGPT merge two people with the same name?
To a language model, your name is a string of characters, and the same string can belong to many people. Nothing in the text it reads says "this is a different Jane Doe" unless the page makes that clear with an employer, a city or a field.
The mix-up can enter by two routes. The first is training: the model absorbed facts attached to a name, and facts about two people can blur together. The second is live search: the assistant looks up the name, gets pages about different people and writes one paragraph from them.
The second route has been tested in the lab. In the AmbigDocs study, published at the COLM 2024 conference, researchers gave language models sets of documents about different subjects sharing a name, such as several people called Michael Jordan, and asked questions using only that name. They report that state-of-the-art models "often yield ambiguous answers or incorrectly merge information belonging to different entities". Each document answered the question correctly for its own subject; the confusion came from combining them. With examples added to the prompt, GPT-4 gave a complete answer, pairing each fact with the right subject, on 58% of questions. Those were 2024 models, so read that figure as a record of the failure and not as a current score.
It happens outside the lab too. In July 2023 the law blog The Volokh Conspiracy documented a Bing summary that combined facts about two men with similar names, a veteran and adjunct professor and a man convicted of trying to join the Taliban, and connected them with a "However" as if they were one person. The author wrote that the two "have nothing in common other than their similar names". A merged answer reads smoothly, and nothing in it warns the reader that two lives have been stitched together.
How do you find out where the mix-up comes from?
Test before you change anything, in a clean session: signed out or in a temporary chat, with memory off, because your own history colours the answer. This guide to what ChatGPT knows about you covers the set-up. Ask three questions in ChatGPT, then repeat them in Claude, Gemini and Perplexity:
- The bare name: "Who is Jane Doe?"
- The name with context: "Who is Jane Doe, the pricing consultant in Hamburg?"
- The separation test: "Is Jane Doe the pricing consultant in Hamburg the same person as Jane Doe the novelist?"
Copy out every wrong fact and every source shown with the answer. Then read the result against this table:
| What you see | What it probably means | What to do first |
|---|---|---|
| A cited page is about the other person | The search returned both of you and the assistant wrote one biography | Make your own pages easy to separate (next section) |
| A cited page about you carries the other person's facts | That page is wrong: a directory, an aggregator, an event page | Ask the site's owner to correct it |
| Wrong facts and no sources | Probably what the model learned in training, with no page to point at | Fix your pages anyway; training changes only with a later model |
| Wrong on the bare name, right with city or company | The assistant can separate you given context, and the other person dominates the name | Work on pages that pair your name with your field; expect this to move last |
What helps assistants tell you apart?
Each page carrying your name should also carry enough to show which person it means.
- Pick one name form and keep it. If your name is common, a middle initial or second name is the cheapest distinguisher you have. Use exactly the same form on your site, LinkedIn, X, talks and articles.
- Write one identifying sentence and repeat it. For example: "Jane Doe is a pricing consultant in Hamburg and the founder of Doe Pricing." Open your about page, LinkedIn About section and speaker bio with it. A sentence with full name, company, city and field still makes sense when lifted out of its page, and a slogan does not.
- Give yourself a page that is only about you. A personal site works, and so does a page on your company's site. Google's ProfilePage documentation says the primary focus of such a page must be a single person or organization affiliated with the website, and it lists an employee page on a company website as an example. A name under a photo in a team grid does not do this job.
- Link your pages in both directions. Your page links to your LinkedIn and X profiles, and each profile links back to your page. For the page's markup, the same Google documentation recommends a
sameAsproperty with the URLs of your other external profiles or home pages. Google documents this for its own search, so do the visible links first: anyone who can open the page, person or crawler, can follow them. - Use the same photo and the same bio. Words do most of the work in a written answer, so treat the photo as the smaller signal: it lets an editor, a journalist or a client see at a glance that two profiles are one person. Bios that contradict each other do more harm, because they give a model two versions of you to reconcile.
- Check what LinkedIn actually shows. LinkedIn lets you set which sections of your profile may be displayed to people who are not signed in, through the edit icon next to "Public profile & URL". Leave on the sections that carry your identifying sentence. Your custom URL is edited there too; if your plain name is taken, add your field instead of a number.
- Ask third parties to add context. Conference pages, podcast notes, press pieces and directories often carry your bare name, and a page like that can attach to either of you. Send the owner your identifying sentence and a link to your page, and ask them to remove anything that belongs to the other person.
- Publish in your field under that name. Every article and post pairs your name with your topic, which makes it harder to attach to anyone else. How to get ChatGPT to recommend you as an expert goes further on this. If keeping it up is the hard part, ItGrows's LinkedIn product drafts LinkedIn and X posts in your own voice on topics you choose, and you approve or edit every post.
One optional line helps when the other person is well known: "Not to be confused with Jane Doe, the novelist." It states the separation in words a model can repeat. The trade-off is that your page now points at the other person, so mention only their field, and add the line only if the confusion keeps coming back.
What can't you change?
The other person. You cannot remove their pages or ask them to be less known. If they are far better known, the bare name will tend to go to them. The AmbigDocs authors give a reason for this: when one subject is substantially more likely to be meant, answering about that one can be more useful than declining to answer. With a famous namesake, the realistic target is a correct answer when someone adds your city, company or field, and no merged biography on the bare name.
What the model has already absorbed. The privacy group noyb reported on 20 March 2025 that ChatGPT "now also searches the internet for information about people, when it is asked who they are", which stopped a false criminal story it had told about a Norwegian man. noyb adds that the incorrect data "may still remain part of the LLM's dataset". So your pages can change searched answers, while answers from training wait for a later model.
Whether the assistant searches at all. In a test we ran on 23 September 2026 with 20 German business questions, the model answered 10 without searching at all and cited nobody. Your pages cannot help an answer that never looks for them.
The next person's chat. Correcting ChatGPT in your own conversation does not rewrite what someone else sees.
How do you check whether it is improving?
Turn the three questions into a fixed list and freeze the wording. Five to ten questions are enough: the bare name, name plus city, name plus company, name plus field, the separation test, and one question about something only the other person did. Run them in clean sessions, in the same assistants, with the same settings.
Repeat each question, because one answer is only a sample. In our own measurements, asked the same question three times, an assistant often names different businesses. We count a business as named for a question if any of three asks names it, and compare checks only on the same questions and the same number of asks. Use the same rule here: a question counts as merged if any of its three asks merges you with the other person.
Sort every answer into one of four groups:
- Merged: one biography with facts from both of you.
- The other person only: it describes them and not you.
- You only: it describes you, with correct facts.
- Separated: it says more than one person has this name and gives each their own facts.
Progress means fewer merged answers and more in the last two groups, first on questions with context and later on the bare name. Keep the cited URLs too: when your own pages start appearing among the sources, a search can reach them.
Allow for slow indexing. LinkedIn warns that public profile changes may take several weeks or even months to reach search tools like Google, Yahoo or Bing, and that it does not control when they refresh. A monthly round is a workable rhythm.
If your company name has the same problem, the method carries over with buyer questions, as this guide to checking whether assistants name your business shows. ItGrows makes businesses and experts visible in AI answers and measures every week whether ChatGPT, Claude, Gemini and Perplexity name them, for readers who would rather hand the work over.
Frequently Asked Questions
Will telling ChatGPT it has the wrong person fix the answer?
It can fix that conversation, and ChatGPT's memory feature may carry the correction into your own later chats. It does not change the pages a search finds for the next person or what the model learned in training. Put that effort into your pages instead.
Should I start using a middle initial or a different professional name?
A middle initial is worth it if your name is common and you can apply it everywhere: site, profiles, bylines and speaker bios. Update the old pages you control and ask the owners of the rest. A completely new name cuts you off from every page that already mentions you, so keep it as a last resort.
How long does a correction take to show up?
Answers built from a live search can change once the corrected pages are found. Answers from what the model learned in training wait for a later model. Date every check so you can see which change moved what.
Can I ask OpenAI to correct a false statement about me?
You can send OpenAI a privacy request, and people covered by the GDPR can rely on data protection law. noyb reports that in a case from April 2024 OpenAI argued it could not correct data and could only "block" it on certain prompts, and its March 2025 complaint to Norway's data protection authority relies on the GDPR's accuracy principle, Article 5(1)(d). Its report describes the complaint and no decision, so keep dated screenshots and proof of the correct facts, and take legal advice if the statement is damaging.