AI written content and what it costs you

The cost nobody puts in the spreadsheet

Everyone talks about AI content in terms of what it saves you. A brief that used to take a writer three days now takes an hour of prompting and editing. That part is real. What doesn’t show up on the savings side of the ledger is the cost that shows up later: the editing tax, the crawl budget you burn on pages that never earn a click, and the trust you lose when a reader or a link partner notices the pattern.

I run a portfolio of content sites and I’ve published a lot of AI assisted pages across them. Some of that content ranks fine. Some of it is dead weight I’m still cleaning up. This is what I’ve actually seen, not a theory about what Google “wants.”

What the model is actually doing

A language model generates text by predicting the next likely word given everything before it, trained on a huge pile of existing text. It doesn’t know facts the way a person who did the work knows facts. It doesn’t have a client it talked to, a test it ran, or a mistake it made last quarter. It produces the statistically likely sentence, which is why AI drafts tend to converge on the same structure, the same hedging phrases, and the same safe, general claims no matter what tool you use.

That mechanic is the root of every cost below. It’s not that the output is “bad” in some vague sense. It’s that the output is, by construction, the average of what’s already been written on a topic. If your competitors are running the same prompts, you’re all converging toward the same page.

The editing tax

The first real cost is time, and it doesn’t disappear, it just moves. Raw model output almost always needs:

• Fact checking, because the model will state something plausible and wrong with the same confidence as something true.
• Removing filler transitions and hedge phrases that add length without adding information.
• Adding the specific detail that only someone who’s done the thing would know, a number, a screenshot, a decision you made and why.
• Rewriting anything that reads like five other articles on page one already.

If you skip that pass to save time, you’ve just moved the cost from your hours to your rankings, because thin, generic pages are the ones that struggle to hold position once a query has any competition. If you don’t skip it, you’ve spent close to the time a competent writer would have spent anyway, just reallocated from drafting to editing. Either way, “AI made this free” is rarely true once you account for the pass a page needs before it’s worth publishing.

Crawl budget and indexing are not infinite

This is the part site owners with a few hundred URLs don’t think about but anyone running a larger site learns fast. A crawler allocates a limited amount of attention to your domain based on signals like site size, server response, and how often past crawls turned up anything worth indexing. Flood a site with hundreds of thin, AI generated pages that all say roughly the same thing in different words, and you’re not adding hundreds of ranking opportunities. You’re diluting the crawler’s attention across pages that individually don’t deserve much of it, and you can end up with slower indexing on the pages that actually matter.

I’ve watched this happen on a site I scaled too fast: pushing volume before the pages had anything distinct to say. The fix wasn’t writing better prompts, it was cutting the page count and merging thin pages into fewer, denser ones. Fewer URLs, more per URL, faster indexing on what was left.

Duplication is easier to spot than people think

Because the model is predicting likely text, AI content across a site (or across sites, if you’re running a portfolio and reusing prompts) tends to fall into the same sentence rhythms and the same paragraph shapes. You notice it as an operator scrolling your own site at 2am. Readers notice it too, even if they can’t say why a page feels off. And if you’re doing outreach or link building, the people on the other end notice it fastest of all, because they read pitches and articles all day and pattern matching is their job.

That’s a real cost that’s easy to underweight: a link partner, a guest post editor, or a journalist who reads one AI shaped page from you is less likely to trust the next one, and trust is the actual currency in link building, not the anchor text. Nothing about a tactic being efficient makes it safe if it burns the relationship you needed the content for in the first place.

Where it genuinely doesn’t cost you much

I don’t think the honest answer is “never use it.” Some jobs are closer to the model’s strength: turning a set of data points you already gathered into a first-pass table, drafting an FAQ section from questions you already answered a dozen times in support tickets, restructuring an outline, or handling the boilerplate parts of a page (definitions, format explanations) where being the tenth site to say the same true thing correctly isn’t a liability. The cost there is close to zero because the editing pass is short and the page was never trying to be your differentiated asset anyway.

The costly move is using it for the part of the page that was supposed to be the reason someone linked to you or trusted you over the next result: the case study, the actual test, the specific number, the thing that required you to have done the work.

No tactic here is a shortcut around the work

I want to be direct about a few things I won’t tell you, because plenty of content on this exact topic will.

I’m not going to tell you AI content ranks on a guaranteed timeline, or at all, because that depends on competition, intent, and execution, and anyone who quotes you a number is guessing. I don’t have inside knowledge of how a search engine’s ranking systems weigh AI text specifically, and I’d be lying if I implied otherwise; what I have is years of watching my own pages win or lose position and drawing conclusions from that, which is a different thing from knowing the algorithm. And no, running AI generated content through a private blog network or an automated link scheme doesn’t make either tactic safer. It just adds a second layer of thin, patterned content sitting on top of the first one, on infrastructure that’s easier to connect back to you than most people assume.

The actual checklist I use

Before a page with AI assisted drafting goes live on any site I own, it has to pass three questions:

1. Does this page contain at least one thing a reader can’t get from the first three results on the same query? A number, a decision, a specific outcome, something.
2. If I strip the topic-boilerplate paragraphs, is there still a page here, or does it collapse to nothing?
3. Would I be comfortable sending this page to someone I want a link or a mention from, knowing they read this kind of thing daily?

If the answer to any of those is no, the page either gets more work or it doesn’t get published. That’s the whole system. It’s not exciting, and it’s slower than “generate and publish,” but it’s the version that doesn’t cost me crawl budget, trust, or a cleanup project six months later.

The real cost of AI written content isn’t the content itself. It’s what happens when you skip the part of the process that made your content worth ranking in the first place.

If you want help figuring out where that line sits for your own site, from keyword research through the technical side, The SEO Desk is where I write up what’s actually worked and what hasn’t.