The three removals: getting off Google is now three jobs

A client sent me a link last week. It was a five-year-old local news piece that mentioned them in passing, in a way they'd rather not be mentioned. They wanted it "off Google."

I've had some version of this conversation about twice a month for eighteen years. What's changed in the last twelve is what my answer has to be, because the client isn't really asking about Google anymore. They're asking about the answer ChatGPT gives when a prospect types their name into it. They just don't know that's what they're asking, and most of the people they've hired to help them don't seem to have noticed either.

The reputation industry is running a playbook designed for a distribution model that no longer exists on its own. The reactive-content-and-suppression game worked when attention flowed through a ranked list of ten blue links. It works considerably less well when a large language model has already read the corpus, formed a summary, and is delivering that summary to the user before any click happens. And it fails almost entirely when the negative source has been baked into a model's training data — because at that point the page could vanish from Google tomorrow and the model would still know what it said.

The industry needs to stop calling all of this "removal" and start being honest about the three genuinely different things that word can mean. Because right now, most reputation budgets are being spent on the weakest of the three against the surface that matters most.

Three things called removal

There's a piece in Search Engine Journal this week from Erase that draws a distinction I've been trying to articulate to clients for months. It splits "removal" into three outcomes and it's the clearest framing I've seen anywhere in the industry.

Removal means the source page is gone. Deleted by the publisher, taken down by the platform, or removed by court order. The URL 404s or is unpublished. Nobody can find the content because there is no content to find.

Deindexing means the page still exists but Google no longer returns it. The URL is live. Anyone with the direct link can still read it. Crawlers and scrapers can still fetch it. Google just doesn't surface it in its results.

Suppression means the page exists, is indexed, and you've pushed it below the fold with other content that outranks it for the queries that matter. Nothing has been removed. Nothing has been deindexed. You've simply changed the ordering of what a user sees when they search a specific term.

These have always been different outcomes with different costs and different reliability profiles. What's new is that they now interact with AI search surfaces in genuinely different ways — and the ordering of "best" and "worst" has changed. In the old model, suppression was often the most cost-effective play because pushing something to page two was functionally equivalent to it not existing. Nobody clicks page two. In the AI model, that stops being true.

Suppression against a model doesn't work the way it worked against a SERP

This is the point most reputation agencies haven't reckoned with yet, and it's the point I keep having to explain to clients who've already spent budget with someone else.

Suppression is a bet on ordering. Ordering is what AI search stopped rewarding.

When a language model retrieves or has trained on a corpus, it doesn't care what position a source occupied. A negative article that lived at position eight in the classic ten-blue-links world was, for practical purposes, invisible — the CTR curve dropped off a cliff after position three or four, and everyone knew it. That same article, if it's in the retrieval corpus a model uses to answer a question about your brand, is now weighted by relevance and authority, not by SERP position. A page nobody would have clicked can absolutely be paraphrased into a summary that millions of people will read.

Pew's browsing analysis, cited in that Erase piece, found that when an AI summary appears in Google results, users click through to a traditional result only 8% of the time. That's roughly half the click rate of searches without an AI summary. Which means when the summary is doing the work, the ranking of the underlying sources matters far less than whether they're in the source set at all.

Erase reports that their inbound requests specifically to address negative AI Overview and assistant-answer content are up 215% year over year. That number is a lagging indicator of a shift most agencies still haven't priced into their proposals. Clients are noticing what their reputation service isn't touching.

Suppression is a bet on ordering. Ordering is what AI search stopped rewarding.

Removal is the only outcome that touches the training corpus

If a page is genuinely removed — deleted from the source, unpublished, taken down after a legal request — then over time it stops being fetchable, stops being cited, and eventually falls out of retrieval corpora and out of the training data future models are built on. The removal isn't instant in AI terms. The cached content, the archived copies, the model that was trained six months ago all still know about it. But the trajectory is clean. Every subsequent index, every subsequent retrieval, every subsequent model has less of that content available to reference.

A Google deindex is a promise Google makes about its own results. It's not a promise anyone else honours.

Deindexing is weaker. The page is still live. It can still be scraped by any crawler that isn't Googlebot, which is most of them at this point. It can still be cited by anyone linking directly. And critically, it can still be pulled into RAG systems that don't rely on Google's index. Perplexity, ChatGPT search, Claude's browsing tool, the various vertical AI answer engines — none of these ask Google's permission before fetching a URL. A Google deindex is a promise Google makes about its own results. It's not a promise anyone else honours.

Suppression is weakest of all. The page is indexed. It's crawlable. It's fetchable. It's in every corpus that's ever been assembled that touched that URL. You've simply arranged for it to appear on page two of a specific results surface for a specific set of queries. In AI terms this is nearly a null operation.

---

The order of operations most agencies have wrong

Here's what I see when I look at reputation proposals coming across my desk, from agencies that were solid at this ten years ago and haven't updated their thinking.

The proposal starts with content creation. Build a portfolio site, launch a personal brand blog, seed a few LinkedIn posts, publish some earned media placements. The theory is that this material will outrank the negative content and push it below the fold.

Sometimes this is the right play. Often it's the third-best play being sold as the only play. Because the two conversations that should have happened first — is this content legally removable, and is the source willing to take it down for other reasons — never got had.

Mistake 1

Skipping the removability check. There are content types that are actually removable through defined processes. Right to be forgotten in the EU and UK, defamation with a court order, expunged criminal records with the appropriate legal documentation, factually incorrect statements where the publisher can be shown documentation, non-consensual imagery under platform policies. These aren't guaranteed successes, but they're processes with actual clear pathways. Every negative URL should be run through them before a single suppression asset is commissioned. The Erase piece frames this as three questions: is the page still live, is there a specific factual error, and is there a court order or legal instrument attached. Anyone selling reputation services who isn't asking those three questions on intake is selling the wrong service.

Mistake 2

Treating outreach as beneath the strategy. A quiet email to a small publisher explaining a situation, offering context, and requesting an update or removal succeeds more often than the industry admits. Not always. Not with major news outlets on legitimate stories. But with the long tail of blogs, forums, local news sites, and small directories — the tail that makes up most of the surface area of a typical negative footprint — direct outreach works considerably more often than the twelve-month content build that gets sold in its place. The problem is that outreach doesn't scale into a retainer as neatly. So it doesn't get sold.

Mistake 3

Selling suppression as the default. If the majority of the queries you're worried about are now returning AI-generated summaries rather than ordered lists of links, ranking assets above the negative content produces a smaller effect than it used to. That doesn't mean suppression is worthless — for direct-navigation searches, for queries that still return classic SERPs, for the users who scroll past the summary — it still moves the needle. But it's now part of the toolkit rather than the whole toolkit, and priced accordingly.

What the reputation stack actually looks like now

If I were designing a reputation engagement from scratch today, without the constraints of what's already been sold to the client, it would look roughly like this.

The first thirty days are diagnostic. Every negative URL gets categorised: is it legally removable, is it eligible for platform policy removal, is it eligible for outreach-based removal, or is it a legitimate news or government source where none of those apply? For each category, the corresponding action is initiated. Legal removals go to counsel or to specialists in the relevant jurisdiction. Platform removals go through the platform's formal process. Outreach cases get personalised emails, not templated ones. This phase is where the largest gains happen and where the least effort is currently being spent.

The next sixty days are structural. For content that can't be removed, the question becomes: is this content being cited by AI systems, and if so, on which prompts? This requires actual monitoring — not rank tracking, because rank isn't the right concept, but prompt-based visibility monitoring across ChatGPT, Perplexity, Claude, Gemini, and Google's AI Overviews. Kevin Indig's H1 2026 data, which I've written about before, found that 91% of citations appear in only one of these systems. So this isn't a single dashboard. It's more like polling. You sample the prompts that matter for the client, you see what shows up, and you build the picture from repeated observation. Any agency claiming to give you a single "AI visibility score" is selling you a scoreboard that doesn't reflect the game.

Then and only then does content creation start. And it starts with a different goal than the old suppression brief had. You're not trying to rank above the negative source. You're trying to be the more citable source for the prompts the negative content is currently answering. That means depth, authoritativeness, unambiguous factual grounding, structured data where it helps machines parse the material, and — this is the part the industry keeps underselling — earned mentions on high-authority third-party sources. Because brand mentions correlate more closely with AI visibility than citations do, and both matter more than SERP position.

The reputation playbook isn't dying. It's becoming a subset of the AI visibility playbook, and being priced by people who don't realise that yet.

The counterargument, taken seriously

The strongest pushback to what I've just written is this: for the majority of consumer searches, classic Google results still exist, still rank, still get clicked, and suppression against those still works. The AI answer might dominate the top of the page but the ten blue links are still under it, and the same users who read the summary sometimes scroll to check the sources. So suppression isn't dead. It's just diluted.

This is fair. And for a specific class of client — usually B2C, usually high-volume search demand around a specific name or business — classic suppression continues to produce measurable results. If someone's Googling a restaurant to decide whether to book, they will scroll past the AI summary to look at reviews. If someone's checking a professional's name before a meeting, they'll click through to LinkedIn. The AI summary changed the top of the funnel but didn't dismantle the whole funnel.

Where I still hold my position is on the trajectory. Every quarter, the click-through rate on AI-summarised queries drops. Every quarter, the share of queries that trigger a summary grows. Every quarter, more of the initial impression of a brand is formed by a paragraph the user didn't click to read. A reputation engagement designed for the world as it existed in 2022 will still produce partial results in 2026. A reputation engagement designed for 2026 will produce those same partial results and the results that matter for the surface where consumer research is increasingly happening. The question isn't whether suppression works. It's whether suppression alone justifies the retainer.

What this means for anyone commissioning reputation work in 2026

If you're a business owner sitting with a reputation problem, the practical implications are these.

Ask your provider — or your prospective provider — the three questions on intake. Is each negative URL potentially removable through legal or platform processes? Have they attempted outreach to the source? Are they monitoring AI answer surfaces or only SERP positions? If the answers are no, no, and no, you're being sold 2018's playbook at 2026's prices.

Assume that anything currently in a training corpus is functionally permanent for the lifetime of that model. New model releases will re-sample the web and may or may not carry the content forward. But the model your prospects are using right now isn't going to be retrained tomorrow because your reputation situation improved. The best you can do for currently-trained models is starve them of fresh signal on the negative angle and flood them with authoritative alternate signal on the positive one.

Prioritise removal over suppression whenever removal is possible. It's more expensive per URL. It's less scalable as a service. It's often slower. It's also the only outcome that actually addresses the AI surface, and the delta between removal and suppression is now large enough that the maths favours it for any content that qualifies.

Treat brand-building as reputation infrastructure, not as a separate discipline. The strongest defence against a negative source being cited by an AI system is a large, well-known, authoritative body of positive material that the model prefers to cite instead. This isn't reputation management. This is brand as the moat, which happens to be the same moat that protects you against every other adverse discovery event as well.

The close

The reputation industry was built for an ordered world. Ten links, ranked one to ten, and the game was to move things up or down that list. That world hasn't ended — it still exists underneath everything — but it's no longer the surface where most first impressions get formed. The surface where first impressions get formed is a paragraph that summarises whatever the model has decided is authoritative on your name.

You can't push a paragraph below the fold. You can only change what the paragraph says by changing what the model has to work with. And you change what the model has to work with by removing what you can, deindexing what you can't remove, suppressing what you can't deindex, and — most of all — building enough positive signal that the model prefers to paraphrase from that instead.

Three different jobs. Three different price points. Three different success rates. Anyone selling them as one thing is selling the wrong thing.