Plan the site first, generate second: why AI content buildouts fail

The short version: when an AI content buildout goes wrong, the problem is almost never the writing. It is that nobody decided what pages should exist before the writing started.

The pattern is always the same. You open ChatGPT, ask it for a page about one of your services, and what comes back is fine. Not brilliant, but fine. So you do another one. Then a page for the suburb next door, then one for the suburb after that, then a few blog posts because someone said you should be blogging. Two weekends later the site has forty new pages on it. Three months after that, the rankings are flat or worse, and nobody can tell you why.

The order was backwards from the very first prompt. Generating pages is now the cheap part. Deciding which pages should exist is the expensive part, and it is the part that got skipped, because it is the only bit the AI cannot quietly do for you while you get on with running the business.

What actually goes wrong when you generate first

Two things break, and they compound each other. The first is cannibalisation (where two or more of your own pages target the same search, so Google has to pick between them and often ranks neither of them well). If you generate a services page, and then a separate page for the same service in your own city, and then a blog post answering the same question, you have not built three chances to rank. You have built three weaker versions of one page. The tell-tale sign is a keyword that bounces around: this week Google shows your services page, next week the blog post, and it never settles anywhere near the top. I have written the mechanics up in full in more pages, worse rankings.

The second is thin content. When you generate without a plan, every page is written from roughly the same prompt with a word swapped, so every page says roughly the same thing. Google is good at spotting that, drops the weakest ones, and a site carrying a pile of near-duplicate pages can end up with its genuinely good pages crawled and trusted less as well. The weak pages drag the strong ones down with them.

What makes AI buildouts specifically dangerous rather than just unhelpful is speed. Doing this by hand, you would write six pages, get sick of it, and stop. The friction limits the damage. With AI you can do sixty in a weekend, so you build the problem faster than you ever could manually, and you will not see it for months.

Volume without a map does not get you more rankings. It gets you more of your own pages competing with each other for the same one.

A keyword map is a list of pages, not a list of keywords

This is what most people get wrong when they finally do sit down to plan. They export a list of keywords, look at four hundred rows, and think each row needs a home. It does not. A keyword map says: these are the pages this site should have, and this is the pool of searches each one is responsible for.

It is a spreadsheet, and it does not need to be fancy. One row per page. Columns for the URL, the one job that page does, the group of search terms it owns, and which existing page it might overlap with. For most local service businesses the finished map runs to a few dozen rows, not four hundred, and building it is an afternoon of thinking rather than a project.

The map is also the thing you hand to the AI. Instead of "write me a page about drain unblocking", the brief becomes "write the drain unblocking page, it owns these eight searches, it must not cover general plumbing or hot water because those have their own pages, and it links to these two pages". Same tool, completely different result, because you have done the deciding and left the AI the typing. That is the whole philosophy behind using AI for your website content without wrecking it and this is the part of it that does the most work.

How to actually build the map

Step one: get the raw list of terms

You need to know what people actually type, which is rarely what you call things internally. Search Console tells you what you already get impressions for, Google's own autocomplete and the related searches at the bottom of a results page cost nothing, and your enquiry inbox is genuinely underrated - people describe their problem in their own words when they email you. AI is fine here too, as long as you treat its suggestions as a list to verify rather than a list to build from. If you want the longer version of this bit, I have covered it in keyword research for small business. Do not stop to judge anything yet - you just want the raw pile.

Step two: group by what the searcher wants, not by how the words look

This is the step that decides whether the whole exercise works, and it is the step software gets wrong, because software groups by string similarity - which words the phrases have in common. That is not the question. The question is what the person wants, and whether one page would satisfy all of them.

Some examples of how badly those two come apart. "Emergency plumber", "plumber open now" and "24 hour plumber Ponsonby" share almost no words with each other, but every one of those people wants the same thing: someone in a van, today. One page. Meanwhile "braces" and "braces cost" are near-identical strings, and they belong together too, because the price question is a section of the braces page rather than a page of its own. But "braces" and "Invisalign" also look related, and they are two separate pools of demand: somebody searching Invisalign has already ruled out metal braces and will bounce straight off a general braces page. Two pages.

The test I use is embarrassingly low-tech. Imagine two of these people ring you. Do you give them the same answer, or a different one? Same answer, one page. Genuinely different answer, and you have found a second page. It is more reliable than any clustering tool I have used, because you are the one who knows your customers.

Step three: name the page, not the keyword

For each group, write one sentence describing the job of that page in plain English: "the page for someone with a blocked drain who wants it fixed today". If you cannot write that sentence without it sounding like the sentence you wrote for another group, the two groups are one page. If the sentence comes out easily and sounds like a different person, it is a real page.

When a new page is justified, and when the term belongs on a page you already have

This is the judgement call the whole map turns on. My default is that a term belongs on an existing page, and it has to earn its way out. Four questions get me there:

• Would the ideal answer be a genuinely different page? If someone searching this term would be perfectly happy landing on a page you already have, it is a section, a heading, or a paragraph on that page. Not a new URL.
• Is there enough that is unique to say? If you cannot write several hundred words that would look out of place on the existing page, there is nothing to build. Padding it out with AI to hit a length is exactly how you manufacture thin content.
• Does an existing page already rank for it, even badly? If Google is already showing one of your pages for that search, building a second one is picking a fight with yourself. Strengthen the page that is already in the race.
• Is the demand real, and worth the work? A page producing one enquiry a year is not free - it needs writing, linking, and keeping accurate. If you have three candidates and time for one, build the one with real demand behind it and leave the others on the map for later.

Two situations trip people up more than any others. The first is the homepage, which everyone tries to make carry every service at once. It cannot, and the reasons are worth understanding properly - I have set them out in why your homepage cannot rank for everything.

The second is the service-versus-location grid, which is where AI buildouts do the most damage. Someone realises they offer six services in eight suburbs, sees forty-eight potential pages, and lets ChatGPT write all forty-eight. Almost none of them will be justified, and the ones that are follow a rule rather than a multiplication - I have gone through where that line actually sits in service pages versus location pages, and how the local picture fits together in the local SEO guide.

What a sane buildout sequence looks like

Once the map exists, the order you build in matters nearly as much as what you build.

• Map the whole site first, including pages you already have. You are looking for overlaps between new pages and existing ones, and that only shows up when everything is on the one sheet.
• Fix what exists before you add anything. Often the map shows that the page you needed is already live and just underdone. Improving it is faster and safer than launching a competitor to it.
• Build the money pages first. Core services, in demand order. These are the pages that produce enquiries, and they deserve your attention while you still have some.
• Locations next, and only the ones the map justified. Real service areas, written with something genuinely different in each. Not a suburb name swapped through a template.
• Supporting and informational pages last, and only where they earn a place. Most local businesses need far fewer of these than they think.
• Link each page in as you publish it, not in a batch at the end. A page nothing points to is a page Google treats as unimportant, and linking is the thing AI handles worst if you leave it to decide.
• Publish in small batches and watch what happens. Five or ten pages, then a few weeks of Search Console. If something you already had starts sliding, you want to catch it while five pages are in play, not fifty.

That last one is the discipline nobody wants. The whole appeal of AI is doing it all at once, and doing it all at once is exactly what removes your ability to tell which change caused which result.

If you have already generated the pages

Plenty of people reading this are past the planning stage and looking at a site that already has the problem. It is fixable, and it is not a rebuild. Build the map anyway, from what is already published. Put every existing URL on the sheet with the job it is doing and the searches it targets, and the duplicates will be obvious inside half an hour. Where two pages chase the same thing, pick the stronger one, fold anything useful from the other into it, and redirect the loser to the winner. That consolidation usually produces a faster improvement than any of the new pages did, because you are stopping your own pages fighting each other.

None of this makes AI the villain. I use it constantly, and this blog is written with it - the mechanical work is automated and every judgement call is made by a person, which is exactly the split I am arguing for. The tool is not the risk. The risk is that it removed the friction that used to force you to think about structure.

So plan the site first. One spreadsheet, one row per page, one sentence saying what each page is for. Then generate, one page at a time, against that plan. It is an afternoon of work, and it is the difference between a buildout that compounds and one you pay someone to untangle in a year.