Is AI Content Bad for SEO? the Controls That Matter Before You Publish

Is AI content bad for SEO? No — Google’s spam policy that governs this question does not name AI as the problem. It names scaled, unoriginal pages published to manipulate rankings, and then says the abuse counts no matter how the pages were created. The tool is not the violation. The volume, the thinness and the intent are.

That distinction is the whole answer — and it is also why most advice on this question cannot help you. You are not deciding whether to use a language model. You are deciding which controls sit between a model’s output and your live site.

Take four things from this page before anything else. Google’s spam policy acts on scaled unoriginal content however it was produced. Google’s own guidance calls generative AI useful for research and for adding structure to original work. Neither document promises you a ranking outcome for disclosing or concealing how a page was written. The five controls at the foot of this page are what we actually run before anything of ours goes live.

What the Myth Claims and What the Policy Says

Three things separate a page the policy targets from one it does not. The tool, meaning which model drafted the page, is not the test on its own. The purpose, meaning pages made to rank rather than to help, is the named abuse. The value, meaning what the page adds for the reader, is the stated standard.

Generative AI does appear in the policy’s examples. It appears there as a way of producing many low-value pages at once, which is a use rather than an offence in itself. Read the three in order and the usual worry inverts: you stop asking whether a machine helped write the page and start asking whether you can stand behind what it says.

Can Google Detect AI Content?

Google’s published guidance on generative-AI content does not tell you to worry about detection. It points you at the spam policies and the Search Essentials, and it asks whether a page helps the person reading it.

You will find plenty of tools that score a page’s probability of being machine-written, and you will find people treating those scores as though they were Google’s. They are not. A third-party detector’s verdict tells you how that detector scored the text. It does not establish whether the page is accurate, whether anyone competent reviewed it, or how a search engine will treat it.

There is a more useful question underneath the one you asked. If a reader cannot tell whether a person thought about the page, the page has a problem — and that problem is real whether or not any machine flags it.

What a Detector Can’t Tell You

A detector cannot see whether the claim in your second paragraph is true, whether the number came from a source you actually opened, or whether anyone who knows the subject reviewed it. Those are among the things that decide whether a page earns its place, and they are the ones sitting squarely inside your control.

Does Google Penalize AI Content?

Not for being AI-written. Google’s spam policy penalizes scaled content abuse, which it defines as many pages generated for the primary purpose of manipulating search rankings and not helping users. The first example listed under that definition is using generative AI tools to produce many pages without adding value.

Read that example closely, because the weight sits on many pages without adding value rather than on the tool. Google’s separate guidance on generative-AI content is, if anything, encouraging. It calls generative AI particularly useful when researching a topic and for adding structure to original content, then points you back at the spam policies and the Search Essentials for the standard your work has to meet.

So the risk is not the model. The risk is the pattern the model makes cheap — publishing at a volume you cannot review, on subjects you have not checked, for a reason that has nothing to do with the reader.

What the Spam Policy Actually Says

Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it’s created.

Google Search Central, spam policies, read 13 September 2026

That phrase does the work: no matter how it’s created. It closes the loophole in both directions. Hand-writing a thin page does not exempt it, and using a model to draft a genuinely useful one does not condemn it.

From Brief to Published Page

Google’s generative-AI guidance adds one practical suggestion rather than a rule. If you generate content automatically, it says, consider telling readers how the content was created, with background on how automation was used and image metadata where that applies. That is a transparency suggestion aimed at your audience. It is not a ranking promise, and you should not treat it as one.

Why Humanizing AI Content Isn’t the Fix

If you are searching for how to humanize AI content, the need behind the search is real. The draft reads flat and you want it to read as though a person wrote it. The fix is editing at the draft stage rather than a rewriting pass bolted on at the end.

Here is the distinction that matters in practice. A humanizing pass takes finished generic text and varies its surface: swapped synonyms, broken-up sentences, a contraction dropped in. The text stops sounding machine-made and carries on being generic, because nothing was added that a person knew. Editing at the draft stage does something different. It cuts the paragraph that could belong to any vendor, replaces the vague attribution with a source you opened, and puts in the thing you learned last month that nobody else can write.

We do not run a humanizing pass on our own work, and that is our editorial choice rather than a claim about Google’s systems. We write to the standard first — a rewrite cannot add knowledge that was never in the draft. What we will not tell you is that any of this helps you evade a detector, or that evading one would improve your rankings. We have no evidence for either, and anyone offering you that trade is selling you something.

The practical version, if you are editing an AI-assisted draft today: read it for claims first and sources second, cut every sentence that survives having its subject swapped for a competitor’s, and keep the passage only a person in your seat could have written.

What Ranked for This Question on 12 September 2026

We pulled the first page for this exact query on 12 September 2026 and kept the raw response. Ten positions returned eighteen result rows, and exactly seven of those rows are classed as organic results.

That gap is the interesting part. The other eleven rows are an AI Overview plus its three sitelinks, four question rows, and three community discussion results from Reddit and Quora. Two of the three AI Overview sitelinks point at URLs that also appear as organic results further down the same page, so the first page offers fewer distinct destinations than its ten positions suggest.

The authority spread should change what you expect to compete against. The highest-authority organic results on the page are a Reddit thread and a LinkedIn post — a specialist page from a smaller domain sits at position nine. You are not up against ten polished publisher pages here. We are describing one dated measurement of one query, and we are not claiming it generalises.

What the First Page Looks Like at a Glance

The First Page by Position, Type and Authority

Eighteen result rows on ten positions. Seven rows classed as organic. One AI Overview with three sitelinks.

What Our Own Search Console Data Shows, and What It Doesn’t

Our Search Console property returned eight query-and-page rows for the thirty days to 11 September 2026, and the honest summary is that our own pages rank far enough down to prove nothing about this question either way.

The clearest row is the AI-marketing-for-contractors page against the query for AI marketing for contractors: eight impressions, zero clicks, average position 72.5. Four further queries touched the same page at a single impression each, at average positions 82, 86, 90 and 100. The homepage and the blog index picked up a handful of brand-shaped impressions around position 51 to 52. If you are reading that as a benchmark, read it as the low end of one — these are early pages on a young property.

We are showing you this because the alternative is describing our own results in the abstract, and you have no way to audit an abstraction. Two limits belong with the numbers. It is a selected slice rather than every page we have published, and nothing in it establishes that AI-assisted drafting caused, avoided or attracted anything. Position data for a handful of queries cannot carry a causal claim, and we will not make one.

Search Console Results for This Site, 13 August to 11 September 2026

Five Controls Before You Publish AI-Assisted Content

If you take one thing from this page, take the list. Each control maps to something the spam policy or the generative-AI guidance actually says.

The Five Controls at a Glance

• A source you opened. Every number, date and named-entity claim traces to a document someone read this run, because a citation carried over from an earlier draft is not a source.
• A named gap, not a filled one. Where the number is not available, the draft says so and ships with the gap named. Nothing gets rounded, smoothed or inferred into place.
• A person on the publish gate. Not a spot check on a sample. Verification before publish is not a slogan, it is a gate, and a page that cannot clear it does not ship.
• A volume you can review. The policy’s test is purpose and value rather than a page count, so this one is our safeguard and not its rule. If your publishing rate exceeds your review rate, low-value pages ship unread, and that is how a library drifts into the pattern the policy describes.
• A reason the page exists. Written before the draft rather than reverse-engineered after it. If the only answer is to rank, you have your problem.

Our own loop runs in a fixed order: an approved brief, drafted against sources, verified claim by claim, published by a person. It is not mysterious. It is a loop you can audit, which is the only kind worth buying.

Where does AI stop paying on a page like this one? At the point where the work stops being repeatable. A model can draft a section, hold a structure, and save you the hours that were never the valuable part. It cannot open a source on your behalf and mean it, decide what your company is willing to claim, or carry the consequence of a sentence that turns out to be wrong. Those remain yours, and the five controls above exist to keep them yours.

If you want that loop running on your own content rather than described to you, the next step is a conversation about scope — and the honest version of that conversation sometimes ends with us telling you which part to specify before you buy anything at all. Founder-operated. Limited client intake.