Revisi started from a fairly mundane problem. When a website has a handful of pages, keeping the writing consistent is mostly a matter of paying attention. You remember how you described the product, which phrases you tend to use, what you deliberately avoid and roughly what the brand is supposed to sound like.
That gets harder once the site grows. If you have 58 landing pages, 17 products and 104 blog articles, you are not realistically going to remember everything you wrote six months ago, whether one page uses completely different terminology to another, or whether the tone has gradually shifted as new content was added. None of those things are catastrophic individually, but they accumulate.
That was the problem I was trying to solve when I started building Revisi. A website usually does not suddenly wake up with a completely different voice; it drifts there gradually as pages are added, rewritten and revisited at different times. Some content might have been written when the business was new, some after the positioning changed, and some because somebody needed a landing page live by Friday afternoon. Once there is enough content, manually checking all of it against everything else stops being particularly realistic.
“Does this sound like us?” is a surprisingly difficult question
Brand voice guides are supposed to help with this, but they are often more subjective than they first appear. A brand might want to sound confident, clear, conversational and helpful, which makes perfect sense when a person reads it but leaves quite a lot of room for interpretation when you actually try to apply it consistently.
What does confident mean at sentence level? Does conversational mean contractions, shorter sentences, fewer formal words or more first-person language? At what point does helpful become over-explaining? Most people can make a reasonable judgement when they are looking at one paragraph, but making that same judgement consistently across a few hundred pieces of content is much harder.
I ran into exactly the same problem when I started using AI to help review the content. The obvious approach was to give an LLM a brand voice guide, give it some copy and ask whether the copy was “on brand”. It worked in the sense that the model always had an answer, and the answer often sounded convincing, but that turned out to be part of the problem.
One review might focus heavily on sentence structure while another concentrated on word choice or formality, even when the underlying content was similar. The feedback was not necessarily wrong, but it was loose enough that I could not be sure the same piece of writing would receive the same type of criticism the next time it was reviewed.
The biggest improvement came when I stopped treating “brand voice” as one large judgement and started breaking it into smaller things that could actually be evaluated. Instead of asking whether a page simply sounded right, Revisi could look at more explicit characteristics such as formality, sentence structure, clarity and word choice.
The exact metrics are less interesting than the principle behind them. If you want reasonably consistent judgement, you have to be much clearer about what you are asking someone, or something, to judge.
That changed how I thought about Revisi as a product. I originally thought of it mainly as an AI that could review a website and tell you whether the writing matched the brand, but over time it became much more useful to think of it as a review system.
I do not particularly want AI rewriting every page on a website for me, because that can create a different consistency problem where everything starts sounding like it came from the same model. What I actually want is something that can point me towards the pages that deserve attention, explain why they stand out and give me enough context to decide whether anything needs changing.
Reviewing a page and reviewing a website are different problems
That distinction became more obvious as Revisi grew. Looking at a single page is mostly a question of whether the copy works on that page, while looking at an entire website means understanding how that page relates to everything around it.
A product page and a long-form article should not necessarily have identical sentence structures or levels of detail, but they should still feel like they belong to the same brand. You also need to know which pages have already been reviewed, which ones have changed since the last scan and whether an apparent inconsistency is actually a problem or simply the result of two pages serving different purposes.
A surprising amount of the work therefore ended up having very little to do with the LLM itself. There is the website crawl, storing and comparing content, identifying what changed, avoiding unnecessary re-analysis of pages that have not changed, grouping inconsistencies and deciding which recommendations are important enough to show first.
A model can generate hundreds of observations if you ask it to, but that is not particularly useful if the person using the product still has to work through hundreds of observations to figure out which five actually matter. The workflow around the model became at least as important as the model's ability to evaluate the writing.
That has probably been the broader lesson for me from building AI products. The interesting part is rarely just getting a model to produce an answer, because producing an answer is now relatively easy. The harder part is deciding what context it needs, how narrowly the judgement should be defined, when the model should run and what the user is supposed to do next.
Revisi became more useful as I spent less time thinking about how to make the AI sound clever and more time thinking about how to make the review process manageable.
I still do not think brand voice can, or should, be reduced to a perfect score. Writing is subjective, context matters, and sometimes the right sentence will break several of the rules you have defined elsewhere. The purpose is not to automate that judgement away.
For me, the useful part is reducing the amount of content that needs human attention in the first place. If I have more than a hundred articles and dozens of other pages, I do not need an AI to decide what my brand should sound like. I need it to help me find the handful of pages where something feels noticeably different, give me a sensible reason why, and let me make the final call.
Revisi came from trying to solve that problem for myself. The amount of content had grown beyond what I could reasonably keep in my head, while simply asking an AI whether something was “on brand” was too vague to be dependable. The interesting part has been turning that subjective judgement into something structured enough to apply repeatedly across a large website, while still accepting that the final decision belongs to a person.