Google used to be the problem. Now the answer is.

There is a small ritual at the start of most of my calls. Someone tells me what has happened, and while they are still talking I open a browser and search their name. It is the least sophisticated thing I do all week, and it is still the most useful. What sits on page one is, for most people, the entire story.

For about fifteen years, that was the job. Ten links, in an order you could see, published by organisations you could contact. Reputation work meant understanding that list and slowly changing it. You could ask a site to correct something. You could ask Google to delist a page under data protection law. You could build enough credible material that the bad result slipped to page two, where almost nobody goes. It was slow and often boring, but it was legible. You knew where you stood because you could see where you stood.

That is quietly ending, and I do not think most people have noticed yet.

The question has changed shape

More and more, people do not search a name. They ask about a person. They open an assistant and type something like "is this person legitimate" or "what was the controversy with them" and they get back a paragraph. Confident, tidy, written in full sentences. No list. Often no links, or a couple of citations most people will never click.

This is a different thing entirely, and it is worth being precise about why.

A search result is a pointer. It says, here is a page, someone published it, go and look. You are aware you are being handed a source, and you retain some instinct to evaluate it. You notice if it is a tabloid. You notice if it is eight years old. You notice if the headline has a question mark in it, which is usually a sign that nobody could prove the thing.

A generated answer is not a pointer. It is a conclusion. It has already read the sources, weighted them in a way you cannot inspect, and handed you the verdict. All the friction that used to sit between a person and a claim has been removed, and with it, most of the scepticism.

I am not making an argument here about whether that is good or bad. It is simply what is happening, and it changes what my job is.

Why the old tools do not quite reach

Here is the practical problem. If a newspaper publishes something inaccurate, I know what to do. There is a named publication, an editor, a complaints process, a correction policy, and if it comes to it, a legal route with a defined timetable. It is often frustrating and it takes longer than anyone wants, but there is a door to knock on.

If a model tells someone you were investigated for something you were never investigated for, who exactly am I writing to?

There is no page. There is no publication date. The answer was assembled at the moment it was asked, which means it may not appear the same way twice. You cannot screenshot a ranking and prove the ranking, because the next person to ask gets a slightly different sentence. You cannot ask for a correction to a sentence that does not persist anywhere.

You cannot even reliably reproduce the harm, which matters enormously if you ever want to do anything formal about it. Clients come to me with a screenshot and I have to explain that the same prompt, run again, might give a perfectly reasonable answer, and that this is not evidence that the problem has gone away. It only means the problem is inconsistent, which is arguably worse, because it will resurface at the exact moment someone important is looking.

The oldest pattern in my work, now built into the machine

I have said for years that corrections never travel as far as accusations. It is the single most reliable rule in this field. The allegation is a story. The retraction is a paragraph on a page nobody links to, published weeks later, at a length nobody notices.

That has always been unfair. Now it is structural.

Think about what a model has learned from. The accusation was covered by twenty outlets, syndicated to another thirty, aggregated, discussed on forums, turned into video essays, quoted in follow-up pieces for years afterwards. The correction was one short item in one place. When a system learns from volume, the accusation is simply louder. Not because anyone decided it should be, but because there is more of it.

So you end up with people whose case was dropped, whose claim was dismissed, whose story was corrected in print, and the summary still leads with the allegation. Sometimes with a small note that it was later disputed, which is the modern equivalent of a correction on page 27.

I find that genuinely difficult, and I do not think the answer is to be dramatic about it. It is a design consequence, not a conspiracy. But it does mean the people who were wrongly accused now carry it further than they used to, and I see that in real lives, not in the abstract.

The legal route is thinner than people hope

The first question clients ask is whether they can sue. It is a reasonable instinct and I understand it. The honest answer is that this is not, at the moment, a solid plan.

In Walters v OpenAI, decided in Georgia in May 2025, a radio host sued over output that falsely said he had embezzled funds from an advocacy organisation. The court ruled against him. Part of the reasoning was that the product carries a disclaimer telling users the output may be inaccurate, and that a reasonable reader would therefore not treat it as a statement of fact. The court also took the view that the company had made serious efforts to reduce errors.

Whatever you make of that reasoning, and there is a lot to argue with in it, the practical takeaway is stark. The warning label may do quite a lot of work.

Other cases are moving. Starbuck v Google was filed in Delaware in October 2025 over allegations the plaintiff says were fabricated by a chatbot, including criminal accusations. That litigation is ongoing and the law here is genuinely unsettled. I would expect the position to look different in a few years, and I would not be surprised if courts in England and Wales take a different view from courts in the United States, given how differently the two systems treat defamation in the first place.

But if you are sitting in this problem today, "wait for the case law to mature" is not advice. It is a shrug.

What I actually do about it

The shift in approach is roughly this. You stop trying to fight the output, and you start changing what the output is built from.

The old model was suppression. Push the bad thing down. The new model is closer to supply. Make sure that the accurate, verifiable, well-sourced version of a person exists in enough places, clearly enough written, that it is the easiest thing for any system to find and repeat.

In practice that means a few unglamorous things.

Fix the sources, not the summary. If there are three articles carrying the original error, those three articles are the problem, not the paragraph they produced. Corrections, amendments, updated copy and clearly dated notes on the original pages are worth more now than they were five years ago, because they are what gets read next time.

Have a first-party page that answers the obvious question. Most people never write anything factual about themselves anywhere they control. Then they are surprised that the internet's account of them is assembled entirely from other people's writing. A clear, plain, boring, factual page on your own domain is not vanity. It is the only version of the story you own.

Be findable in the places that get treated as reliable. Not every source is weighted the same. Established outlets, professional bodies, court records, official registers and long-standing reference sites carry more than a post that will be gone in a year. Getting the accurate account into one of those is worth more than getting it into ten of the others.

Deal with syndication properly. One story reprinted across a group of titles is not one problem. It is a dozen. If you correct the original and ignore the copies, you have corrected the least-read version.

Use the data protection routes where they apply. Delisting requests still work on search engines, and the criteria are more generous than people assume, particularly where the material is old, inaccurate, excessive, or relates to something that was never proven. It is a form-filling exercise with a lot of judgement in it, and it is worth doing properly rather than firing off a template.

None of this is exciting. When I explain it, I can usually feel the disappointment on the other end of the call. People arrive expecting strategy and drama, and I am talking about correction requests and structured pages. But this is the work that changes what is there in two years, and two years is the timescale that actually matters.

The thing I will not help with

I should be clear about a limit, because this is where the field gets a bad name.

Everything above is about making sure the accurate version of someone is available and findable. It is not about making true things disappear. If somebody did the thing, the answer is not to engineer the record. It is usually to deal with it, properly, and then let time and behaviour do what they do.

I take on work where the backlash was disproportionate, the account was wrong, or the story dropped the context that made it make sense. I have turned down work where the honest answer was that the coverage was fair. That is not moral posturing on my part, it is practical. You cannot build a durable reputation on a corrected record if the record was right.

Where this leaves us

I keep coming back to a thought that I find slightly uncomfortable. For all its faults, the old system had one thing going for it. It showed its workings. You saw the sources and you made up your own mind, badly and with bias, but you did it yourself.

What we have now hands people a settled answer and quietly removes the step where they might have doubted it. And for anyone whose name is attached to something unresolved, contested, or plainly untrue, that missing step is everything.

So the practical advice is duller than it should be. Do not wait for the storm to pass and assume the record fixes itself, because the record is now the thing that lasts. Fix the sources while anyone still cares enough to amend them. Write down the accurate version somewhere you control. Deal with the copies, not just the original.

And if you are in the middle of this and it feels like there is nothing to push against, that is not you being slow. There genuinely is less to push against than there used to be. It just means the work moves earlier, and further back, into the places nobody thinks to look.

If any of this is sitting on your chest at the moment, I do not charge for a first call. I am not going to sell you anything on it either. I tried a sales job once and lasted two weeks, so you are quite safe.