Your Old Content Is Talking Behind Your Back

AI doesn't just read what you published last week. It remembers everything you never cleaned up.

Somewhere on your website there is a blog post from four years ago that describes a product you no longer sell, quotes a price you no longer charge, and recommends a strategy you now tell clients to avoid. You forgot it exists. The models didn’t.

This is the uncomfortable asymmetry of the AI search era: you experience your content as a stream, but machines experience it as a pile. You think in terms of what you published this quarter. A language model retrieving an answer about your category doesn’t care about your content calendar. It cares about what’s retrievable, and your 2022 pricing page is exactly as retrievable as the thing you shipped on Tuesday.

In the Google era, this mostly didn’t matter. Old posts sank. Rankings decayed, traffic dwindled, and neglected content quietly composted at the bottom of page seven, where it could embarrass no one. Obscurity was a garbage collection service, and it was free.

AI answers don’t work that way. When a model synthesizes a response about your product, your category, or your point of view, it pulls from whatever it can find and treats it all as roughly simultaneous. The result is a collage: a sentence from your current homepage, a claim from a 2023 case study, a feature description from a help doc you deprecated but never deleted. The user sees one confident paragraph. You see — well, you don’t see it at all. That’s the problem.

The pile has no timestamps

Here’s a thought experiment. Imagine every page you’ve ever published printed out and handed to a smart, fast-reading stranger, with the dates torn off. Then imagine a customer asks that stranger to describe your company. What would they say?

That’s uncomfortably close to how your brand gets represented in AI answers. Retrieval systems do weigh recency signals, but weakly and inconsistently — and once a claim is in the answer, the user has no idea which decade it came from. A model that says “the platform starts at $49 a month” sounds equally confident whether that fact is from this morning or from three pricing changes ago.

The old advice was to keep publishing because fresh content signaled a living site. The new reality is harsher: everything you’ve ever published is your current messaging, whether you meant it to be or not.

Publishing is easy. Custody is the job.

Marketing teams are structured around production. Briefs, calendars, launches, the satisfying green checkmark of “published.” Almost nobody is structured around custody — the ongoing responsibility for everything already out there.

Ask a content team how many posts they shipped last quarter and they’ll know instantly. Ask how many of their live pages contain claims that are no longer true, and you’ll get silence. Not because they’re careless, but because no one has ever owned that question. The archive is nobody’s job.

In the AI search era, the archive is the job. Your corpus — the full pile — is what gets read, chunked, embedded, and quoted. A corpus with contradictions produces contradictory answers. A corpus where your old positioning outweighs your new positioning ten pages to one will describe you as the company you used to be. You can’t out-publish a pile that disagrees with you. Twenty fresh posts sit next to two hundred stale ones, and the model has no reason to prefer yours.

What custody actually looks like

The fix is not glamorous, which is probably why nobody does it.

Start with an inventory. Not a spreadsheet of URLs — an honest read of what each page claims. Prices, features, names, numbers, opinions. Most teams doing this for the first time find whole strata of fossilized messaging: the rebrand that never propagated past the top ten pages, the killed feature still starring in three tutorials, the founder take that the founder no longer takes.

Then make a call on every stale page: update it, redirect it, or delete it. Deletion feels like burning equity, and sometimes it is — but a page that actively misinforms models about your business has negative equity. It’s not an asset that stopped working. It’s an employee giving wrong answers to customers, around the clock, in your name.

Finally, change the ratio of effort. If your team spends ninety-five percent of its content energy on net-new and five percent on maintenance, the pile wins. Flip enough of that budget to keep the corpus coherent. One accurate, current, quotable page beats five new posts stacked on a foundation of contradictions — because when a model goes looking, coherence is what it finds, and coherence is what it repeats.

The bottom of the pile

There’s a reason this discipline doesn’t exist yet: for twenty years, neglect was safe. The search engine forgot, so you could too.

That deal is over. The systems that now describe your company to the world have long memories, no sense of chronology, and no ability to know that you’ve moved on unless you tell them — by cleaning up the record they read from.

You’ve spent years worrying about what to say next. Meanwhile, everything you’ve already said is still talking. Go find out what it’s telling people.