Use it to draft and research, not to publish unread. AI is good at structure, summaries, and first passes. It is bad at knowing what is true about your business. The honest rule is that a human has to own every claim on the page, because Google''s guidance and your customers both hold you to that standard.
This is not a purity argument. We use AI tools in our own process every week. The question is not whether a machine touched the text. It is whether anyone checked it.
Most bad AI copy is not bad because a model wrote it. It is bad because nobody read it afterwards, nobody verified a single number in it, and nobody had anything specific to say in the first place. Those were problems before ChatGPT existed. AI just made them cheaper to scale.
Google judges the purpose, not the tool. Its guidance on creating helpful content asks you to be transparent about automation and to explain why you used it. Content made to help people can use AI. Content generated mainly to game rankings breaks the spam policies regardless of how it was made.
Google frames this through what it calls the Who, How, and Why of your content. Who made it should be clear through bylines and author pages. How it was made should be disclosed, including any automation. Why it exists is the one Google treats as decisive: the content must exist to serve a real audience rather than to attract search traffic.
That third question is the whole test. If you can answer "why does this page exist" with something other than "to rank for a keyword", you are probably fine. If you cannot, no amount of human editing will save it.
Google''s self-assessment questions are worth reading directly. They ask whether the content offers original analysis, whether it would make someone bookmark or share it, and whether the author is demonstrably knowledgeable. Those are hard questions to pass with a generic draft, whoever wrote it.
Most of it, in some form. Ahrefs analysed 900,000 newly created English-language pages crawled in April 2025 using its own detector, bot_or_not, and found that 74.2% contained AI-generated content. The split is more interesting than the headline. Only 2.5% were purely AI, 25.8% were purely human, and 71.7% were a blend.
That blend number is the real story. The dominant pattern is not machines replacing writers. It is writers using machines, then editing. Ahrefs also surveyed 879 content marketers and found 87% reported using AI to create or help create content, with blog posts the most common use.
So the question "should you use AI" is already settled in practice. Almost everyone is. The competitive question has moved on to how much human judgement sits on top of the draft.
The study, published by Ryan Law with data work from Xibeijia Guan, is worth reading if you are still debating this internally. It reframes the conversation from a moral one to a practical one.
In the parts of writing that are structural rather than substantive. Outlining a page, drafting alternative headlines, tightening a paragraph you have already written, summarising research you gathered yourself, and rewriting jargon into plain language. All of these are transformations of material you already own.
The pattern is worth naming. AI is reliable when it is working on input you supplied and unreliable when it is producing input you did not. Rewriting your own case study is safe. Inventing a statistic to support your case study is not.
It is also excellent at the unglamorous middle of a project. Meta descriptions for forty CMS entries. Alt text drafts. Consistent product descriptions across a Webflow CMS collection where the facts come from a spreadsheet you already maintain. This is where the time actually goes on a build, and where automation earns its place.
We talked about the wider workflow question, including design and build tasks, in our piece on using AI in a web design workflow.
Anywhere a claim carries risk. Pricing pages, service descriptions, case studies, compliance language, and anything with a number in it. A model will produce a confident, well-formed sentence that is simply not true about your company, and it will not flag that it guessed.
The failure mode is specificity without grounding. Ask a model to write about your onboarding process and it will describe a plausible onboarding process. It reads well. It may not be yours. On a marketing site, that becomes a promise you did not intend to make.
Homepage copy is the other weak spot. The job of a homepage is to say the one thing that is true about you and not true about your competitors. A model trained on everyone''s homepage will produce the average of all of them, which is the exact opposite of what you need.
The tell is usually vagueness that sounds confident. Phrases like "cutting-edge solutions" and "tailored to your needs" survive because nobody can prove them wrong. They also do not persuade anyone.
Because it is the one thing a model cannot fake for you. Google''s documentation asks whether content clearly demonstrates first-hand expertise and depth of knowledge, giving the example of expertise that comes from having actually used a product or service, or visiting a place. A model has done neither.
This is the first E in E-E-A-T, and Google''s framing of the whole set is useful. Google states that of experience, expertise, authoritativeness, and trustworthiness, trust is the most important, and that the others contribute to trust. Content does not have to demonstrate all of them.
Practically, that means your unfair advantage is the thing that happened to you. The migration that went wrong and what you learned. The client question you get every week. The tradeoff you made and regret. None of that is in a model''s training data about your company, because it only exists in your head until you write it down.
Our own rule is that every article has to contain at least one thing only we could say. If a piece could have been written by anyone in our industry, we do not publish it. That constraint is what stops AI assistance from turning into AI averaging.
Scaled content abuse is Google''s term for mass-producing pages to manipulate rankings. Google''s spam policies define it as many pages generated for the primary purpose of manipulating search rankings and not helping users, and explicitly name using generative AI tools to generate many pages without adding value for users.
Read that definition carefully, because the trigger is volume plus intent, not authorship. Ten thoughtful pages written with AI assistance are not scaled content abuse. Four hundred near-identical location pages spun from a template are, whether a human or a model produced them.
Google''s policy also covers the older versions of the same trick, including scraping feeds or search results and running automated transformations like synonymising or translating to obfuscate the source. The tooling changed. The policy did not need to.
The practical guardrail is simple. If you could not defend a page to a customer as worth their time, do not publish it. Volume targets are what push teams over this line, which is why we would rather publish fewer pages that each say something.
Google''s guidance says yes when it matters to the reader. It asks creators to disclose automation and explain why it was used, as part of the How in its Who, How, Why framework. That does not mean a disclaimer on every page. It means not pretending a machine-assisted process was purely human when a reader would care.
In practice the standard we apply is whether disclosure would change how someone reads the piece. A meta description drafted by a model needs no notice. An article claiming first-hand testing does, because the claim itself is the value.
The stronger move is showing your work rather than confessing your tools. Name your sources inline. Say which year a figure comes from. Link to the original. A reader who can check you does not need to be told how the sentences were assembled.
Give it the research, not the authority. Gather your sources first, hand them over, and ask for structure and drafting. Then verify every number against the original before it goes live. The order matters: research, then draft, then verify. Drafting first is how invented statistics get published.
That is close to how we work. We use AI to summarise material we have already retrieved, to propose outlines, and to tighten prose. We do not let it supply facts, and any figure that cannot be traced back to a named source that we actually opened gets deleted rather than softened.
Deleting is the discipline people skip. The tempting move when a stat will not verify is to reword it as "studies show". That is worse, because it keeps the persuasive weight while removing the accountability. If you cannot name the source, the sentence should not exist.
If your goal is to be visible in AI answers rather than just to produce copy faster, the requirements are different again, and we set them out in our guide to getting cited by AI search engines.
Use AI for the draft, own the facts yourself, and never publish a page you would not defend in a meeting. That combination gets you the speed without the liability. The businesses getting hurt right now are the ones that removed the human review step, not the ones that added the tool.
If you take one thing from this, make it the verification habit. Everything else about AI writing is a preference. Publishing an unverified claim about a real company is a risk with a real cost attached.
The same question comes up about building the site itself, and we answered that separately in our piece on whether AI can build your website. If you would rather talk it through than read another article, let''s talk. We are happy to look at your current copy and tell you which parts we would keep. You can reach us at phoenix.studio.
Tell us where you want to go. We'll tell you how we'd get you there.