What 6 Months of Testing AI Content for SEO Taught Me

Key Highlights

• Lightly edited AI drafts can rank for low-competition keywords, but they are almost entirely ignored by AI Overviews, which want original substance that templates cannot supply.
• The content that performed best used the AI draft as scaffolding and added human expertise, verified data and real examples on top.
• Unedited AI output does not just underperform, it quietly introduces fabricated and unverifiable claims, which is why a human verification layer is not optional.
• The tactics that failed all shared one flaw: they tried to game the system. Schema gaming, publishing at scale, keyword-stuffing answers and neglecting off-page signals each underdelivered.

AI writing tools promised to change content marketing overnight: publish at scale, rank everywhere, cut costs to nearly zero. Six months of actually working with them, set against what the research now shows, tells a more grounded story. The results are neither the utopia the vendors sell nor the collapse the sceptics predict. They are specific, they are genuinely useful, and they point clearly to a workflow rather than a shortcut.

The stakes for getting this wrong are higher than they look, and the evidence is starting to quantify them. According to a 2026 audit of four major computing-conference proceedings, fabricated or unverifiable "mysterious" citations went from appearing in none of the 2021 papers to affecting 2-6% of the 2025 papers, once AI writing tools became common and their output went out unchecked. That is an academic-publishing finding, not an SEO one, but the mechanism transfers exactly: unverified AI text smuggles in errors that look perfectly plausible, and someone has to catch them before they ship. On a content site, those errors become the trust problem that keeps you out of AI answers.

This guide shares what that stretch of testing, and the research around it, actually taught me. It covers whether AI content ranks on Google, whether it appears in AI Overviews, what the best-performing pieces had in common, why unedited drafts go wrong even when they read fine, which gaming tactics failed, how to use these tools well, and how to measure whether the editorial layer is paying off. If you want the flip side of the ranking debate, my piece on whether AI-generated content can rank without backlinks pairs well with these findings.

Does AI-Generated Content Actually Rank on Google?

Yes, AI-generated content can rank on Google, especially lightly edited drafts targeting low-competition keywords. But raw or generic AI output struggles against competitive terms and rarely earns the trust and depth that strong, durable rankings need without human enrichment. Ranking is possible, never automatic, and the gap widens as competition rises.

The nuance lives in competition and quality. For low-competition, informational queries, a decent AI draft can rank and hold for a while. For competitive terms, where depth and trust decide the winner, unedited output falls short quickly, because there is nothing on the page a competitor with real expertise cannot beat. The easy wins are real but shallow, and they rarely survive contact with a serious rival.

Google's own position is that quality matters, not origin. It rewards helpful, people-first content however it was produced, and filters generic material that adds nothing, applying the same bar to AI and human writing alike. So AI content ranks when it is genuinely good and struggles when it is generic, which means the tool does not move the bar, it just lets you hit or miss it faster. The pattern I kept seeing matches that: minimally edited drafts caught easy rankings, then plateaued or slipped as richer competitors moved in, while drafts treated as raw material and heavily improved held far better.

Does AI-Generated Content Show Up in AI Overviews?

Largely no. Lightly edited AI content is almost entirely absent from AI Overviews. Overviews favour sources with clear trust signals, original insight and genuine helpfulness, which templated output usually lacks. Earning a citation there takes the human-added value that generic AI content, by its nature, cannot provide on its own.

This was one of the clearest lessons, and the most important one. Ranking and being cited in an Overview are two different bars, and content that scraped a ranking still routinely failed to earn a citation. The Overview layer does not just want relevance. It wants a source worth synthesising into an answer, and a generic draft gives it nothing distinctive to lift.

The reason is structural. Overviews build from sources they trust, and generic content carries weak trust and no original substance, so there is simply nothing there worth quoting over a stronger page. Without first-hand experience, verified data or distinct expertise, an AI draft is invisible at exactly the layer that increasingly matters most. If you want the mechanics of how that selection happens, my explainer on where Google AI Overviews pull answers from breaks the signals down.

What Did the Best-Performing AI-Assisted Content Have in Common?

The best-performing content treated the AI output as a first draft and added original data, real examples and genuine expertise on top. AI handled structure and speed; a human added the insight, verification and first-hand experience that earn both rankings and, crucially, AI Overview citations. That combination beat pure AI and slower all-human workflows alike.

The editing here is not cosmetic, and that is the part most teams underestimate. It is where the value gets added, turning a generic draft into something worth ranking and citing. In practice the strongest pieces kept only the useful skeleton of the draft, with the rest rewritten, restructured or enriched by someone who actually knew the topic. The AI bought speed on the scaffolding; the human bought everything that made the page worth reading.

What mattered most for AI visibility specifically was original substance. A genuine statistic you gathered, a real case you worked on, a first-hand observation nobody else had: these are exactly the differentiators a template cannot fabricate, and they are what Overviews reward. One piece of original evidence did more for citation odds than any amount of polishing a generic draft, which mirrors what I found in what content ranks in AI search, where originality and structure beat volume every time.

It helps to be concrete about what "original substance" means, because it is easy to nod at and hard to do. It is not a rephrased version of what the top-ranking pages already say, which is most of what an AI draft produces by default, since the model was trained on those very pages. It is the thing that was not already on the internet before you wrote it: a number from your own data, a lesson from a project you actually ran, a comparison you tested, a nuance an expert on your team knows that the generic consensus misses. When an engine assembles an answer, that kind of material is what it has to quote you for, because no one else has it. Everything an AI can generate unaided is, almost by definition, already available from a source with more authority than you.

There is a second, quieter property the winning pieces shared: a clear point of view. Generic AI drafts hedge, because hedging is the safest average of everything they were trained on. The pieces that earned attention took a position, said plainly what worked and what did not, and were willing to be specific enough to be wrong. That decisiveness is both more useful to a reader and more quotable to an engine, and it is precisely what a model will not do for you on its own. Adding it back is human work.

Why Does Unedited AI Content Go Wrong Even When It Reads Fine?

Because fluent text is not the same as accurate text. Unedited AI output reads smoothly while quietly introducing claims, figures and references that are wrong or unverifiable, and a reader cannot tell by looking. That is the real risk of publishing drafts as-is: the errors are invisible until something breaks trust.

The academic audit makes the danger concrete. When AI tools spread through those conference proceedings, fabricated citations appeared where there had been none, affecting a small but real share of published papers, and tellingly no author disclosed using AI even where policy required it. The failure was not that the text looked bad. It looked fine. It was that plausible-sounding fabrications slipped through because nobody verified them, and confident prose is exactly what makes such errors hard to catch.

Carry that onto a content site and the cost is trust, which is the currency AI Overviews actually spend. A page with a fabricated statistic or a misattributed claim is not just wrong, it is a liability the moment anyone checks, and generative engines increasingly weigh source reliability. This is why the verification pass is not optional polish but core production work. The human layer is not there to make the draft prettier; it is there to make it true, and truth is what the citation layer is screening for.

The failure mode is particularly dangerous because of how confident the output sounds. A human writer who is unsure tends to hedge, flag the gap, or leave a note to check later. A model fills the gap smoothly with something that reads exactly like the verified parts around it, so there is no visible seam where the fabrication begins. That is what makes "it reads fine" such a poor test. Fluency is the one thing these tools reliably deliver, which means fluency can no longer be your signal that a claim is sound. The only reliable check is to trace each specific fact, figure and reference back to a real source before it ships.

In practice this reshapes the editing job. The expensive part of reviewing an AI draft is not smoothing the prose, which is already smooth, but interrogating it: pulling every statistic, every named study, every confident assertion and confirming it exists and says what the draft claims. That is slower and less satisfying than rewriting a clunky human draft, and it is the step teams under time pressure quietly skip. Skipping it is exactly how the fabrications in that audit reached publication, and on a content site it is how a single unchecked page can cost you the credibility that took months to build.

"Mysterious citations are routinely appearing in peer-reviewed publications throughout the scientific community."

Amanda Bienz and colleagues, The Case of the Mysterious Citations (arXiv, 2026)

What Gaming Tactics Failed, and Why?

The tactics that failed all tried to game the system rather than serve the reader: over-relying on schema, publishing AI content at scale, keyword-stuffing direct answers and neglecting off-page brand signals. Each assumed AI visibility could be manufactured, and each consistently lost to genuinely helpful, well-structured, trusted content.

These are the confessions worth sharing, because the failures teach more than the wins. Here is what did not work, and the lesson underneath each.

Over-Relying on Schema Markup

The most seductive trap is treating schema as a lever that forces citation. Schema helps a machine understand your content, but it cannot rescue a thin or unhelpful page, and no amount of markup manufactured visibility the content had not earned. Added to genuinely strong content it helps; bolted onto weak content it does nothing. Use it to describe real substance, not to fake it.

Publishing AI Content at Scale

Spinning up volumes of AI content to blanket a topic backfired: generic, templated pages were largely filtered out of Overviews, and publishing at scale with minimal editing dragged down overall site quality signals. Scale is not a strategy when the underlying pages add nothing, and the systems are increasingly good at ignoring mass-produced text, sometimes at the expense of the whole domain.

Keyword-Stuffing the Answers

Cramming keywords into answer passages reduced AI visibility rather than raising it. Engines extract clean, direct passages, and keyword-dense text reads as noise to both machines and people. Directness and clarity beat frequency every time, and stuffing actively suppressed citation on pages that might otherwise have been lifted.

Neglecting Off-Page and Brand Signals

The quietest failure was pouring months into on-page work while ignoring brand mentions and entity building. Off-page signals carry more weight for AI citation than many admit, because recognised, trusted entities get cited more readily than anonymous ones. Credible mentions and earned recognition often outweigh on-page tweaks, a theme I dig into in why branded websites rank better in AI search.

The through-line across all four is simple: AI search closes the gap between looking good and being good. Tactics that exploited that gap in old SEO now decay into quiet invisibility, because the systems assess substance and trust rather than surface signals. The reassuring flip side is that the honest strategy is also the durable one, which strengthens the trust signals AI actually recognises instead of gaming them.

What makes these failures especially costly is how they fail. Old-school manipulation, when it stopped working, usually produced a visible drop you could diagnose and reverse. These tactics fail silently. You keep publishing, the schema is valid, the pages are indexed, and nothing obviously breaks, yet the citations never come and you cannot point to the moment it went wrong. Months disappear into work that was never going to pay, and the absence of a dramatic penalty is exactly what lets the wasted effort continue. Recognising a silent failure early is worth more than recovering from a loud one, because there is no alarm to tell you to stop.

There is also a compounding cost that the mass-publishing trap in particular carries. When a wave of thin AI pages drags down a site's overall quality signals, it does not only fail to rank itself, it can dampen the pages that would otherwise have done well. So the downside is not merely wasted effort on the weak content; it is the drag that weak content places on your genuinely strong pages. That asymmetry is why "publish everything and see what sticks" is a worse bet in AI search than it ever was in classic SEO, and why restraint has become a real tactic rather than a lack of ambition.

How Should You Actually Use AI Content Tools?

Use AI for first drafts, structure and speed, then add human expertise, verified data and real examples before anything gets published. Treat the tool as an accelerator for editorial work, not a replacement for it, and reserve publishing for content that genuinely adds something a generic draft cannot.

A workflow that consistently held up looks like this. Let the AI draft the structure and cover the obvious ground quickly, so you skip the blank page. Then add original insight, first-hand experience and a clear point of view that no template could produce. Layer in human-verified data and real examples, checking every figure and reference rather than trusting the draft. Edit hard for clarity and easy extraction. And publish only the pieces that clear a simple test: does this add something a generic draft could not?

The winning mindset is AI plus human, never AI instead of human. That combination earned both rankings and Overview citations in practice, while pure AI content reliably earned neither, and unedited drafts carried the added risk of shipping fabrications. For a closer look at which approaches survive contact with reality, my guide to AI writing tools that produce content that ranks covers what holds up and what does not.

The economics of this are worth sitting with, because they invert the pitch the tools are sold on. AI does not remove the expensive part of content, it relocates it. The drafting that used to take the most time is now cheap, and the value has migrated to the parts a model cannot do: sourcing original material, verifying claims, and adding a genuine point of view. So the honest way to budget an AI-assisted workflow is not to plan for less human time, but to redirect that time from typing toward thinking, checking and contributing what only a person can. Teams that treat the saved drafting time as a licence to publish more, rather than to enrich each piece more, tend to get the volume trap and none of the upside.

Restraint deserves to be an explicit policy, not an afterthought. A useful rule is to publish fewer pieces than the tools make possible and to hold each to the standard of adding something the open web does not already contain. That single gate filters out most of what would have failed anyway, protects your quality signals, and concentrates your editorial effort where it actually earns citations. It feels counterintuitive to use a speed tool to publish less, but that is precisely the discipline the results reward.

How Do You Measure Whether the Editorial Layer Is Paying Off?

Measure it by tracking both traditional rankings and AI citations for the pieces you publish, comparing lightly edited drafts against human-enriched versions on the same topics. If the enriched pieces rank better and earn citations while the raw drafts do not, the editorial layer has proven its worth in numbers rather than opinion.

Set a baseline for each piece first: the queries it targets, and whether it ranks or is cited today. Then compare an enriched version against a lightly edited one on the same topic and watch the citation gap. That gap is usually the most persuasive argument for the editorial investment you will ever put in front of a budget holder, because it converts a philosophical debate about AI content into a measured difference.

Give the comparison a fair window, too, and check the same queries more than once, because AI answers vary from run to run and a single reading can flatter or punish a page by chance. A pattern that holds across several checks is signal; a one-off is noise. Do not lean on rankings alone, because a page can rank while staying invisible in AI answers, and the two move independently. Tracking citation share separately is what reveals the full picture, and it is the metric that tells you whether the human substance is landing where it counts. To see how your AI-assisted content performs in AI answers specifically, a free AI-visibility audit is a quick way to get that baseline before you scale a workflow.

Conclusion

Six months of working with these tools, checked against the research, settled the debate with nuance rather than slogans. AI content can rank, mostly for easy terms, but it stays largely invisible in AI Overviews unless a human adds real substance, and unedited it carries a quiet risk of shipping fabrications that erode the very trust citations depend on. The origin of the content matters far less than whether it is genuinely helpful, verified and differentiated.

So treat AI as a first-draft engine, not a publish button. The editorial layer, the original data, the real examples, the genuine expertise and the verification, is where rankings and citations are actually earned, and it is also what keeps fabrications off your pages. Combine AI speed with human substance, applied with a little restraint about what actually deserves to be published, and you get close to the performance the tools promised but cannot deliver alone.

Want to know whether your AI-assisted content is genuinely earning visibility? Run a free AI-visibility audit with Rank in AI Overview and see exactly where you stand across the AI engines.