An AI Overview is not a new copy of the article. It is a paragraph Google generated from pages it can still see, printed above the results, with citations that look like proof. Australians who have just finished a Search removal meet it as a fresh injury: the blue link is gone, and the first thing on the page is a fluent summary of the same allegation, sometimes with the de-indexed URL still hanging off the bottom as a source.
If you remember one distinction from this piece, make it this one: the source still wins. Models do not adjudicate your reputation. They compress whatever the retrieval layer puts in front of them. While the article is live, or while a scrape of it is live, the overview has a job it knows how to do. Take the source down, and de-index the URLs, and the overview usually starves. Ask it to forget you while the Herald still has the piece, and you are writing to the wrong desk.
The rest of the search and archives hub is about leftovers that were stored on purpose. This leftover is computed on the day you search.
What an Overview is actually doing
Google’s AI Overview — the successor to the experiment that was called Search Generative Experience — sits on top of the same index the ten blue links use. It retrieves. It summarises. It cites. It is not a separate library of harmful pages that you can petition as if it were the Wayback Machine. It is a reading of the current search.
That architecture has three consequences people dislike, and they are all logical.
First, a URL that still ranks can be used as a source even if you have already written to the publisher. The overview is not waiting for your concerns notice to expire. It is reading what the crawler last fetched.
Second, a URL de-indexed for your name can still be read if it ranks for a different query, or if another page quotes it. Overviews are greedy. They will take a secondary write-up, a blog that copied the brief, a Wikipedia paragraph. The news-archive problem does not go away because the generator is new. It gets louder, because the generator will flatten three sloppy secondaries into one confident sentence.
Third, the safety filters that do exist are not a reputation product. Google has said Overviews should refuse or hedge on some classes of harm — child sexual abuse material, suicide methods, some medical and violent queries. It has not said that a ten-year-old court report or a spent charge in a regional paper will be withheld from a name search. Those remain “information”.
The product also moves. Whether you see an Overview at all depends on the query, the device, and whether you are in Australia without a VPN. Capture the page you actually have, on the day you have it. Do not argue from a screenshot a cousin took in London in 2024.
Chatbots are a second, messier layer
Gemini, ChatGPT, Copilot, Claude and Perplexity are not Google Search and they are not each other. For a harmful result they split into two behaviours that both get called “the AI said it”.
Live retrieval. The bot ran a search, or a browse, and summarised what it found today. This is the Overview problem in a different wrapper. If the source is down and the indexes have dropped the URL, this behaviour improves. If the source is up, the bot will find it, and it will find it on Bing as easily as on Google.
Training memory. The bot is producing a paragraph from weights frozen on a snapshot of the public web. This is the behaviour that survives a perfect takedown, and the one that is easiest to overstate. Most of the name-search answers we have seen in the last year were retrieval: the model cited a URL, or it went quiet when the URL died. A smaller set — usually people written about in a high-circulation masthead — still attract a bare, uncited summary after the live pages are gone. That is a real residue. It is not a reason to skip the source work.
Each vendor has a feedback control and a legal address. Use them for a specific false statement, a copyrighted work, or an intimate image, and attach the output. Do not send a general request to “delete me from the model”. There is no such product.
Australia does not yet have a statutory right to have a training set rebuilt around your name. A court order that names a URL still matters, because it can change what the retrieval layer is allowed to see. Do not assume a defamation judgment against a newspaper is a judgment against every chatbot that once read the newspaper.
Thin-profile risk
The overview is worst for people who barely exist online. One local-court brief, one university disciplinary note, one “former employee of” paragraph in a trade blog — and nothing else. The retrieval layer has a single cluster. The generator writes it up as a biography.
This is why the old reputation advice about “flooding the SERP” keeps being sold as a package of micro-sites that fool nobody. The honest version is smaller. A current, ordinary public page — an employer bio, a professional register, a LinkedIn that is actually maintained — gives the retrieval layer a second cluster. It does not erase the court report. It stops the court report being the only sentence a model can form.
The dishonest version is worse than nothing. A grid of doorway pages with your name in the title and no publisher anyone has heard of will not be cited. The overview will skip them and go back to the Age. Thin spam does not cure a thin profile. It advertises that someone is trying.
Two Australian peculiarities make the problem sharper. Court lists and older news CMS templates still throw up a full name, a suburb and a charge on a page that has no other content. If the damaging page is a news archive, do that job at the masthead. Do not build a fake life on the internet in order to dilute a real paragraph. People with dense public lives have a different problem: the overview will mix the scandal with the CV. That mix is painful, and it is evidence that the system is reading more than one source. It is not a takedown failure.
What to do, in order
- Capture the Overview or the chatbot answer, with the date, the citations, and an incognito window from Australia. Save the cited URLs. Those URLs are the work.
- Take the sources down, or start that process. A live 200 will be summarised again tomorrow.
- File Google personal-information or legal removals for every cited URL that qualifies, as the Search removal guide describes. File Bing if the chatbot is clearly reading Bing.
- Use the Overview feedback control for a specific factual error, not for “please stop mentioning me.”
- Use each chatbot’s feedback or legal path the same way — one output, one claim, one source list.
- If you have no other public page and the remaining source is a single old article, decide whether a genuine, current listing is available to you. Do not buy a dozen.
- Recheck in a fortnight, not that afternoon. Then recheck in ninety days. Overviews refresh when the underlying results refresh. They also refresh when a scrape appears.
What not to do: paste the overview into a prompt that asks another model how to “wipe” it; send the intimate image to a general inbox; pay for “AI reputation monitoring” that screenshots the same paragraph every week. Monitoring is a calendar reminder. It is not a lever.
What success looks like
Success is an Overview that no longer runs, or that runs on the current sources and does not lead with the old allegation. Success is a chatbot that declines to invent a biography, or that cites only the pages you have decided can stay. It is not a letter from Google that says the model has been retrained. That letter is not a product they offer to ordinary Australian name searches.
A generated paragraph can be wrong in a new way: it can merge two people, invent a conviction, or promote a comment-section claim into a fact. Those are good feedback tickets. They are also a reason to keep the source file. If the only public text is a sloppy secondary, the generator will be sloppy. If the only public text is gone, the generator often goes quiet. That is the entire method, dressed up as a new product.
If you want the desk to look at a general question about Overviews and chatbots, use contact. Do not send the intimate image or the court file. We publish method. We do not run your case.