Using AI for SEO without producing content nobody wants
What Google's spam policy actually prohibits, which parts of the job AI is genuinely good at, and a division of labour that survives a core update.
Entrovix AIAug 5, 2026 11 min read
Using AI for SEO without producing content nobody wants
The question is usually asked as whether Google penalises AI content. That is the wrong question, and answering it wrongly has cost a lot of sites a lot of traffic.
What the policy actually says
Google's spam policies name a violation called scaled content abuse, defined as when many pages are generated for the primary purpose of manipulating search rankings and not helping users. The examples explicitly include using generative AI tools to generate many pages without adding value for users.
Read that carefully, because the emphasis is not where people assume. The violation is producing many low-value pages to manipulate rankings. It is method-agnostic: the same policy covers spun articles, scraped content and outsourced filler written by humans. The spam policy documentation (opens in a new tab) is the primary source and it is readable.
Google does not have a rule against AI writing. It has a rule against publishing a lot of pages that are not worth reading, and AI made that considerably easier to do by accident.
Where AI is genuinely good at this work
The useful division is not between writing and not writing. It is between tasks where being wrong is cheap and visible, and tasks where being wrong is expensive and invisible.
| Task | Use AI? | Why |
|---|---|---|
| Grouping a keyword list into topics | Yes | Tedious, mechanical, and errors are obvious on review |
| Drafting outlines and structures | Yes | Cheap to reject, and it breaks the blank page |
| Rewriting a passage that will not flow | Yes | You already know what it should say |
| Writing meta titles and descriptions at scale | Yes, with review | Formulaic work with a clear length constraint |
| Summarising research you have gathered | Yes | You can check it against the source |
| Generating statistics or citations | No | The failure mode is a confident fabrication you will not notice |
| Writing about your own service or process | No | It does not know, so it will produce plausible generalities |
| Producing opinions or recommendations | No | The value was that a person with experience said it |
| Publishing anything unread | No | This is the actual policy risk, whatever wrote it |
The failure that actually gets sites hurt
It is rarely one AI-assisted article. It is the pattern: two hundred pages published in a month, each targeting a keyword variation, each containing the same generic advice reworded. Every page is individually defensible. Collectively they are exactly what the policy describes.
The tell is that nobody would link to any of them, nobody would bookmark any of them, and if you removed half at random nobody would notice. If that is true of a batch of content, the volume is the problem regardless of how it was produced.
A workflow that holds up
The principle: AI accelerates the parts of the job that do not require having done the work, and a person supplies the parts that do.
- 1Choose the topic yourself, from questions customers actually ask. This is where most AI-led content fails, because the model does not know your customers.
- 2Research from primary sources yourself. Official documentation, your own data, your own experience. Do not let a model supply facts.
- 3Use AI for the outline, then rearrange it. The first structure it offers is usually the most generic one.
- 4Write the sections that need judgement yourself, especially anything containing a recommendation, a trade-off or a number.
- 5Use AI to draft connective passages and to tighten prose you have already written.
- 6Add what only you have: a specific example, the thing that surprised you, the case where the usual advice does not apply.
- 7Verify every factual claim against its source. If you cannot find a source, cut the sentence.
- 8Read the whole piece aloud. Anything you would not say to a client goes.
Step seven is not optional. Fabricated statistics are the single most common way AI-assisted content damages a business, because they are specific, plausible and quotable, and the damage lands when a customer checks.
How to tell if what you have is thin
Four honest questions, applied to a finished draft.
- Does it contain anything not already in the top five results? If not, there is no reason for it to exist.
- Could a competitor publish it verbatim, changing only the company name? If yes, it says nothing about you.
- Would you send it to a client who asked this question? If you would rather write a fresh email, the page is not finished.
- Does it commit to anything? Content that recommends nothing risks nothing and is worth nothing.
Tools, and what they are for
We build AI tools and we are not neutral about this, so it is worth being precise. They are useful for drafting, restructuring and summarising work you are already doing. They are not a content strategy, and using one to produce a hundred pages is the failure mode described above.
One thing worth saying plainly: humanising tools that exist to defeat AI detectors are solving the wrong problem. Google's policy is about value, not about authorship, and rewording thin content does not make it valuable. If a page needs disguising, the honest fix is to make it worth reading.
Questions people actually ask
Will Google penalise me for using AI?
Not for using it. The policy targets mass-produced low-value pages regardless of how they were made. A well-researched article that a model helped draft is not what it describes.
Do I need to disclose that AI was involved?
Google does not require it. Whether you disclose is an editorial and trust decision for your own audience, and reasonable businesses land in different places on it.
Can AI detectors tell?
Unreliably, in both directions. They flag human writing and miss edited AI writing. More to the point, Google has not said it uses them, and building a content process around defeating a third-party classifier is optimising for the wrong judge.
How much content should I publish?
As much as you can make genuinely worth reading, and not one page more. For most small businesses that is two to four pieces a month, not forty. Content that gets traffic and generates leads covers how to choose which ones.