Building out the LLM

When we talk about "building out the LLM" here, we don't mean training a new model from scratch or fine-tuning one on your own data. We mean something more practical: creating a reusable set of prompts, tone instructions and structural rules for an existing large language model (like ChatGPT or similar tools) so it can consistently generate on-brand copy for a specific part of your website. Using an LLM is a one-off conversation to get a piece of text.

Building one out is setting up a repeatable system so that every home page, service page or blog post is generated in the same consistent style, rather than starting from a blank prompt each time.

Patience is a Virtue

You need patience and normally you will have to refine the prompts and words that you have used. From my own personal experience, it seems as if the biggest stumbling block is a lack of patience and a genuine understanding of what prompts use them. How often can you read a piece of copy having to refine it? That’s either never or very rarely.

People are proud of their copy and often unwilling to change or learn. I’m not trying to offend anyone, but you need copy that reflects what are people are looking for. This has been an issue since the late nineties, and we are still seeing the same issues today.

What has changed is we are being bullied into a style that LLMs understand. It’s not just content; it’s code as well. You are expected to write your website copy in a style that LLMs understand. Beyond Google, all screen readers and crawlers have always needed this. To be specific: website owners, not LLMs or search engines, are the ones being pushed to adopt this new conversational writing style. Search engines and AI tools set the standard that businesses must now write to, whether that business is a small local firm or a large enterprise, if they want their pages to be understood and surfaced correctly. It’s why the link building industry based on PBNs (private blog networks) have been successful.

And so, we are where we are today. A compromise of being forced into a style of writing that is based on answering questions in a conversational style because LLMs are the new bosses and it’s what they want to see. Even worse is the demand from LLMs to ask and answer these questions in the LLM way. It’s not natural, I’m sure it will mature in time, but we are seeing a demand to answer as many questions as possible in the LLM stylee. The sad fact is we see websites with poorer offerings steal that space from those with authority. It may not be a good thing, but it’s a reality. As a result of this we all need to change if you want any visibility in this AI overview space.

Creating your own LLM is a way of generating the bones of website copy. I’m not asking you to play the game, but I am asking to consider using a few different large language models for different parts of your website.

Creating LLM Templates

I found writing home page copy is entirely different to most other pages of the site. I see home pages like album covers, (an album is a piece of vinyl plastic that has music etched into its grooves). The album cover show cases a 1^st impression. It may not necessarily reflect the songs inside the album but it is there to grab attention. Generating copy from an LLM for a home page is often impossible but it’s trying to do the same thing.

What is easier is creating different style of content for each part of the website. You can create templates for lead generation, blog content, service pages, a FAQ section for these types of pages, a CTA on some pages, timeline and evidence pages, local pages. They’ll all need their own JSON schema mark up for them to be understood, but it’s all doable. A blog post will normally be written in a style that is not the same as a service page, an industry page or a page with a transactional action or a page written with a call to action for lead generation. For those that have been writing copy for decades, or those that written a book or two, this shouldn’t be too much of a problem. You should have a large enough bank of “words” to start building out your LLM for at least some parts of the website, but rarely all of it.

A few terms are worth pinning down here. "LLM styles" refers to the distinct prompt and tone settings you create for each page type, so a blog post and a service page don't come out sounding identical. A "bank of words" is simply the growing collection of past copy, phrases and examples you feed back into your prompts so the LLM has real material to draw on rather than generic filler. And "nuances" means the small, individual quirks of phrasing, rhythm or word choice that each past author leaves behind, the kind of differences that make a page sound like it was written by several different people rather than one consistent voice.

In practice, this means working through each page type one at a time. Start by picking a single page type, say, service pages, and pull together three or four examples of your best existing copy for that type. Write a short brief that sets the heading structure (H1, H2, H3), the word count range, the tone, and the specific questions the page must answer. Feed that brief and your example copy into the LLM as a standing prompt, generate a draft, then compare it line by line against your examples and tweak the prompt until the gap closes. Repeat this process separately for lead generation pages, blog posts, FAQs and local pages, since each needs its own brief, its own examples and its own JSON schema markup.

Too Many Authors Spoils the Broth

There are some caveats. If lots of people have blogged or written on your website, and you build your LLM on it, you will several types of nuances or writing styles in your auto generated copy. If you briefed someone else to do it for you, let them know that you are looking and what to avoid from the start; put it into the brief, if not, you are almost guaranteed a disappointment. The quality of work is heavily based by the prompts you use. If you want the copy to look like AI generated with Chat GPT – (dashes), that’s what you’re going to get. You reap what you sow.