When a client asks for AI-written pages

We fulfil search and content work for other agencies, which means the same client request reaches us from several directions at once and we can tell when something has gone universal. Since the spring it has been this one: can we just have ChatGPT write the service pages.

Sometimes it comes from the client, who stayed up until midnight producing forty city pages and wants to know why you are charging for what took them an evening. Sometimes it comes from inside the agency, from a project manager looking at a content backlog that has not moved in two months.

Either way, “AI content is bad for SEO” is the wrong answer. It is not what Google said, it is not what we do, and it will lose you the account to whoever says yes. Here is the position we use instead.

What Google has actually said

Worth being precise, because most of the advice going around is not.

In February Google published guidance on AI generated content, and it was not a ban. The line was that Google rewards high quality content however it is produced, and that using automation to generate content whose main purpose is manipulating search rankings is a spam policy violation. Same standard as always, applied to a new production method.

Two systems are doing the enforcing on local sites:

• The helpful content system, running since last August and updated in December, which judges whether a site as a whole reads like it was made for people or assembled for search engines. It is site wide, and that is the part clients miss. A pile of thin pages can weigh on the pages you care about.
• Core updates, the most recent one in March, which reweight quality signals across the board.

So the risk was never that a machine wrote it. The risk is that what got written is the same thing forty other sites in the metro published, about a business the writer has never visited.

Where AI drafting genuinely works

We use it daily, and we are specific about where.

• Turning an owner’s voice memo into prose. Record the owner or the lead technician answering questions for twenty minutes, transcribe it, have a model shape it into sections. Everything in the output came from the business.
• Expanding an outline a strategist wrote into a first draft that a person then cuts by a third.
• FAQ sections built from real questions, pulled from call recordings and the front desk inbox, not from a model’s idea of what people ask.
• Meta descriptions, alt text and internal link suggestions at volume, reviewed in a batch by someone who knows the site.
• Reformatting what exists: breaking a wall of text into a structured page, drafting the summary paragraph, writing the schema.

What all of those have in common is that the source material belongs to the client and the judgement belongs to a person. The model is doing transcription and shaping, which it is good at.

Where we refuse

• Service pages with no interview behind them. If nobody spoke to the business, there is nothing in the page that belongs to the business.
• Location pages produced by find and replace on a city name. Fastest way there is to make a local site look spun, and that was true long before anyone typed a prompt.
• Anything carrying a clinical or safety claim. Dental, med spa and anything medically adjacent gets written from the practice’s own material and reviewed by the practice. We do not let a model invent a treatment claim on a provider’s letterhead.
• Review responses sent unread. Drafted, fine. Sent without a human reading them, no.
• Volume for its own sake. When a client asks for thirty posts a month, the useful question is what any of them are for.

Change the unit you sell

Here is the commercial problem underneath all of this. If you price content by the page or the word, AI has made your price look absurd and no amount of quality argument will fix it. The client can see that the marginal cost of a word is now close to zero.

So move the unit. What we bill for, and what we suggest partners bill for:

1. The interview. Thirty minutes of the owner’s time per service line, recorded, plus the work of getting them to sit down for it at all. That is the scarce input now.
2. Assets from the job. Photographs, before and after, names and roles of real people, the detail only someone who was there would know. Clients underrate this and it is the one thing competitors cannot copy.
3. Structure. Which pages exist, what each one is for, how they link to the pages that make money.
4. Editing and accountability. Someone’s name is on it being accurate.

Content produced this way does cost less to make than it did two years ago. Say so. Pass some of that back, and put the saved hours into interviews and photography. A client who feels the process got cheaper does not go shopping for a cheaper vendor.

What we are telling agency partners

• Say yes to AI in the process and no to AI as the source, and put it in the scope in writing so nobody relitigates it every month.
• Audit each client site for pages that exist only to hold a keyword, and decide before the next core update whether to improve, merge or remove them.
• Replace word count line items in your next proposal with interview and asset line items, and watch how the conversation changes.
• Give every service page one thing that could only have come from that business: a named person, a photo taken on site, a job they did, a policy they actually follow.
• Build location pages from real evidence of work in the area, or do not build them.
• Fix the client’s Google Business Profile before the content plan. Social profile links are rolling out on profiles now, and the listing still moves the phone more than any page will.

None of this is a defensive position. The agencies we fulfil for are not losing work to models, they are losing it to whoever gets the owner talking for half an hour. That is the part we built our white label fulfilment around, and it is the same reason local SEO work still starts with the listing and the business rather than the word count.