Why Good AI Content Costs More, Not Less

“You use AI. So is this slop that will get my site penalized?”

It’s the question a lot of prospects hold and don’t voice. We answer the half that needs a direct answer: yes, we use AI across our pipeline. This post is about how we ensure quality while using it.

Generating text became nearly free. That moved the cost — it didn’t remove it. What you pay for now is everything around the writing itself: the research before a draft exists, verifying each claim against an opened source, human sign-off, and checking the live page after it goes up. The failures you’ve seen — invented experts, fabricated references, a consulting report that got refunded — all share the same root. Not AI. Unverified AI.

What You’re Paying For When Writing Costs Nothing

In October 2025, arXiv’s computer science category stopped accepting review articles and position papers unless they had already been accepted at a peer-reviewed venue.¹ The reason, in arXiv’s own words: “due to the unmanageable influx” of submissions. Large language models had made papers — especially ones that didn’t introduce new research — “fast and easy to write,” and its volunteer moderators “do not have the time or bandwidth” to handle the volume. arXiv’s fix was to make someone else pay the checking cost.

Springer Nature’s editorial AI policy², read this month, draws the same line from the publisher’s side. It states plainly that “Human accountability is non-transferable.” AI use that “generates unverifiable outputs” sits in the not-permitted tier. Nobody’s rule is “don’t use AI.” The rule is that the output must be verifiable.

One measure of what “nearly free” means: generating text at a fixed benchmark level of capability cost $37.50 per million tokens in March 2023 and $0.18 per million tokens by February 2025, per Epoch AI’s tracked prices³. The cost of writing fell by orders of magnitude. The cost of checking did not.

And yet the measured problem is not that AI writes worse prose. A 2026 study co-authored by Internet Archive staff⁴, built on a representative sample of websites, found no statistically significant evidence that a rising share of AI-generated text reduces factual accuracy or stylistic diversity — while a majority of surveyed US adults believe it does. The honest picture: the damage isn’t in the sentences. It’s in the claims nobody verified.

Google Doesn’t Care That AI Wrote It. It Cares If It’s True.

That fear about an AI penalty is aimed at the wrong thing — and Google has been saying so for years.

In its February 2023 guidance⁵, still linked as current from the live helpful-content doc, Google answers the direct question: “Appropriate use of AI or automation is not against our guidelines.” It adds the symmetric half: “Using AI doesn’t give content any special gains. It’s just content.” No penalty, no bonus.

What Google’s spam policies⁶ actually prohibit is scaled content abuse — generating many pages primarily to manipulate rankings, “no matter how it’s created.” When Google rewrote the policy⁷ in March 2024, it explicitly removed automation as the trigger, naming instead “whether automation, humans or a combination are involved.” The spam policies page lists 16 named practices. None of them is a rule against AI-generated content.

So what does Google grade? Its own self-assessment questions⁸ ask: “Does the content have any easily-verified factual errors?” and “Does the content present information in a way that makes you want to trust it, such as clear sourcing …?”

The document used to train the humans who grade its results — the Search Quality Rater Guidelines⁹, September 2025 edition — says in §4.6.6 that “the use of Generative AI tools alone does not determine the level of effort or Page Quality rating.” The guidelines define Trust as “the extent to which the page is accurate, honest, safe, and reliable.” Accuracy is the first word.

Hallucinated facts are not graded as “AI content.” They are graded as factual inaccuracies that make a page Untrustworthy. (And no, E-E-A-T isn’t a specific ranking factor; Google’s own page says so.)

What Unchecked AI Content Actually Costs

The failures that make the news aren’t about prose style. They’re about claims that crumble the moment someone reads the source.

In October 2025, The Guardian¹⁰ reported that Deloitte agreed to partially refund the Australian federal government for a AU$440,000 report after errors were found — including nonexistent references and citations. The updated report disclosed for the first time that part of it was produced with a generative AI toolchain. It’s the cleanest example available: unverified AI content had an invoice attached.

Newsrooms have their own version. In May 2025, NPR¹¹ reported that a summer reading list published by the Chicago Sun-Times — syndicated through King Features — invented 10 of its 15 books, and The Verge¹² recorded that the same section quoted experts who don’t appear to exist. The writer admitted to 404 Media that he used AI and didn’t check the output: “On me 100 percent and I’m completely embarrassed.”

In August 2025, The Guardian¹³ reported a Press Gazette investigation that led Wired and Business Insider to remove articles bylined by a freelancer who appeared not to exist, the stories AI-generated.

The rates are similarly blunt. In a March 2025 report, the Tow Center at Columbia¹⁴ tested eight AI search engines on 1,600 queries built from real article excerpts and found they collectively answered more than 60% of queries incorrectly; more than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs. The BBC and EBU’s October 2025 study¹⁵ — 2,709 responses across 18 countries and 14 languages — found 45% of responses contained at least one significant issue; the errors were “systemic, spanning all languages, assistants and organizations involved.”

None of these are prose failures. Every one is a checking failure. And when AI assistants misquote the web — fabricating citations or crediting syndicated copies instead of original sources — that’s also why owning your own rankings still matters.

How We Ensure Content Quality While Using AI

Google now asks publishers to explain how automation was used — listing “AI-assisted” as its own category and inviting publishers to share the details. This section is that explanation.

The angle is a human decision

Nothing gets written until a person picks the angle. The system proposes; a human decides, and the job stops dead until they do.

A search result is not a source

Every fact in the draft comes from a page that was actually opened, with its URL, title and publication year recorded. A year is recorded only when it’s visible on the page — if the page doesn’t say, we don’t say. Time-sensitive data works to a defined freshness window; anything outside it is either replaced or dated in the sentence itself. For listicles and comparison pieces, every named company is checked to be a real, trading business — the right industry, real products, a real executive in the stated role.

The check that doesn’t trust the writer

A separate pass reads the finished article — and only the finished article — and verifies every checkable claim independently against the live web. A verifier must quote the exact line on the page that confirms or contradicts the claim; “verified” with no evidence is not a permitted output. Corrected claims are cleared and must earn their verification again from sources. The tool that writes the prose is not the tool that checks it, and the part that gathers the evidence is not the part that rules on it.

That separation produces a concrete deliverable: every delivered article ships with its own fact-check record — a claim-by-claim table of type, claim, verified, source URL and evidence excerpt, headed “N of M verified,” plus a sources tab. An unverifiable fact never reaches the draft. On the finished draft, an unsupported claim is either cut or corrected to what the source actually says; an unsupported superlative gets downgraded, not left standing.

A person signs off — and usually the client does too

Automated consistency, spelling and brand checks run before an editor opens it. Then an editor decides — propose a precise line, make one change without touching the formatting, or rewrite properly — and nothing moves on to the publisher until a person signs it off. On most accounts that means the client signs it off too.

From day one the workflow was flawless. For every opportunity, I could approve the site, see the target URL/anchor, and review the article before it went live. It felt like an extension of our team, with quality control baked in.

— Span Chen, CMO, VidMage.ai

The article going live is the start of QA, not the end

The published page is checked directly: it loads, it can be indexed, the link is present with the right anchor and no value-stripping attributes, nothing has been appended after handover, and the page isn’t labeled sponsored or user-generated or syndicated somewhere else. Findings are graded — breach of what you ordered, advisory, or an honest “we couldn’t determine this” — and a link that fails is held. We hear about it, not you. The delivery notification only fires when QA clears it.

After that it’s re-checked every night for as long as it’s live — the same machinery behind outreach link building and the accountability promise in our satisfaction guarantee. That ongoing watch is also what our link journey post traces across a placement’s life.

Content produced to this standard is included in the placement price, with revisions included — never a separate line item.

Judge It Yourself

The judges we answer to are our clients — including editorial teams whose job is catching exactly the failures listed above.

The content quality has been consistently excellent. It reads naturally and meets our editorial standards without us having to push back.

— Mario, Kape (ExpressVPN, CyberGhost, Private Internet Access)

That’s a client’s verdict from his own editorial review. He clears AI-assisted content on quality, which is the distinction this whole post exists to draw. Portfolio-level results sit in our case studies.

Ask to see the work. Request samples of published content at presshero.io/contact/ — and ask for the fact-check record that ships with every article. The claim-by-claim table is the difference between saying content is checked and showing it.

This article was itself produced with the assistance of generative AI, using a process similar to the one it describes. The numbered markers above and the Sources block below are that process’s visible output.