Your company changed. AI hasn't caught up. How to fix it.

Welcome to the first edition. I’m Samanyou, founder and CEO of Writesonic. Each issue, I’ll unpack one real AI experiment from inside the company, including the workflow, the evidence, and what went wrong.

The core idea: When your company changes, AI doesn’t update from your homepage alone. It rebuilds the story from every source that describes you, including the stale ones you forgot existed.

A few weeks ago, I opened a Temporary Chat on ChatGPT and asked a simple question:

I ran the same question more than once. A few minutes apart, the model described two different companies.

One answer led with the company we used to be, a content creation and marketing platform. Another led with the company we’re becoming, an AI search platform, and cited our current site.

That contradiction became the starting point for this experiment.

Writesonic started in 2021 as an AI writing tool. Today, we help brands understand and improve how they appear across ChatGPT, Claude, Gemini, and other AI search experiences.

The product, website, and category all changed. The web’s picture of us changed much more slowly.

The screenshot showed that the old positioning could surface. To understand how often it happened, I needed a repeatable test.

AI answers vary with the model, region, prompt wording, account context, time, and sources retrieved for that specific run.

A Temporary Chat removes memory and chat-history personalization, which makes the test cleaner. It still captures one answer, from one model, at one moment.

That changed the question I was trying to answer:

I used our own product to track the answer across engines, regions, and time. Dogfooding was convenient, but it also forced us to confront the same issue our customers bring to us.

Two metrics matter here, and they answer different questions:

Visibility tells you whether the buyer hears your name. Citation share helps explain why the model said what it said.

Razvan Surdu ran one product through 47 prompts. A competitor was named 31 times. His product appeared three times.

His result sharpened the question for me: when buyers ask for a recommendation, which company gets named instead of yours?

AI search can combine older learned associations with information retrieved from the live web. When those signals disagree, the answer can drift between who you were and who you are.

When the model lands on our current homepage, it usually gets the new positioning.

When it retrieves an old directory, help article, review profile, or comparison page, it can pull us back into the old category.

That surviving description is what I call the inherited answer.

It’s the version of your company the web learned before you changed. You didn’t write it in one place, so you can’t fix it in one place.

And that was the uncomfortable realization: our positioning problem wasn’t contained inside our website.

Over several years, we repeatedly changed the hero, subhead, product story, and category.

Today, our homepage says, “Win customers from AI search.” A few years ago, it sold a different product. A few years before that, it told another story again.

And Writesonic was never only a writing tool.

Along the way, we launched an image generator, an audio product, a chatbot builder, and a broader SEO suite. Each product added another plausible description of the company.

Our 2021 About page captured the earliest version clearly. We said we were building “state-of-the-art AI-powered apps,” starting with a product that helped me write landing-page copy.

Every version was true when we published it, and together they created a messy identity trail.

This is the mental model that finally made the problem click for me.

Your homepage is one vote in an election happening across the web.

Wikipedia votes. Review platforms vote. Product directories vote. Your documentation votes. Reddit threads, listicles, app stores, GitHub repos, and years of backlink anchor text vote too.

A homepage rewrite updates one strong signal while hundreds of weaker signals keep reinforcing the old category.

That’s why the work extends beyond copywriting. You need to align the source layer that search engines and AI systems can retrieve.

We even pruned parts of our blog that kept reinforcing the wrong audience and topic. Some pages ranked, but attracted students and freelancers instead of the marketers we wanted to serve.

One example was a Hindi page ranking for “what should I write in my bio.” It brought traffic. It also taught the web very little about the company we wanted to become.

HubSpot has written about removing a large volume of outdated content. Animalz documented a similar content-pruning program at QuickBooks.

Those cases made me look at the content library as a system. Page-level traffic can hide the audience and category the full collection teaches the web.

A collection of individually successful pages can still point to the wrong audience and category.

We started by listing every high-signal page that still placed Writesonic in the old category. Then we separated pages we could edit from pages we had to influence.

The homepage and About page were current. Our Wikipedia article now opens by describing Writesonic as an AI visibility and generative engine optimization platform.

One layer below, several profiles were still carrying the old company story.

At the time of the audit:

Those descriptions were accurate when they were written, but they no longer matched the company. We’ve recently updated the profiles we can control, including Trustpilot, so they now reflect the current product. The audit still mattered because it showed how easily an old category can survive in places nobody checks after launch.

The audit kept expanding. Product docs, help-center articles, browser extensions, app-store pages, WordPress plugins, GitHub repos, and old integration pages all contained language written years ago.

The oldest description of your company is often sitting somewhere your marketing team never opens.

I expected old listicles. Finding the same problem in our own developer docs and help center was more embarrassing.

Our API docs still introduced Writesonic as “an AI-powered content automation platform.”

And an old help-center article still answered “How is Writesonic different?” by calling us an AI writing tool.

Here’s what made the problem feel finite: the AI answer cited that help-center article.

The model had effectively pointed us to the stale source behind the stale story.

The source audit gave us a concrete order of operations:

Nothing about this work is glamorous, but it gives us something concrete to change. We update the source itself instead of betting on a temporary quirk in one model.

As Lily Ray says about fragile search tactics, “It works until it doesn’t.”

A corrected source should remain useful through model changes because the underlying information is now accurate.

The first changes helped, but branded and category prompts moved at very different speeds. That difference showed us where the remaining work lived.

For branded prompts such as “What is Writesonic?”, the new AI-search positioning now appears more consistently. Those answers often cite pages we control.

Category prompts tell a different story.

Ask for “alternatives to Writesonic,” and many answers still place us beside AI-writing products. The sources for those answers tend to be third-party comparison pages and old listicles.

The split was consistent:

I now use the distance between those two groups as a rough measure of how far the new positioning has travelled beyond our own domain. The gap is smaller than it was, but it is still visible.

Parts of this work resemble link building, but the scoreboard is different. I care less about the number of links and more about whether the sources shaping buyer questions describe the right company.

Here is the process I would use after any meaningful repositioning.

Choose the branded and category prompts that matter commercially. We run ours through Writesonic’s AI Visibility Tracker, so the same queries are collected across the engines, regions, and markets we care about.

The method does not depend on one vendor. What matters is keeping the prompt set stable and storing the answers, citations, and competitor mentions over time. That gives you a bird’s-eye view of the pattern instead of a folder full of isolated screenshots.

Track three things separately:

Review the bird’s-eye view weekly while you are actively changing sources. A biweekly or monthly cadence can work once the positioning is more stable. The trend view shows what manual checks hide: the dominant description, recurring sources, regional differences, and the competitors that keep replacing you.

Keep the prompts consistent long enough to see whether the distribution is actually moving.

Start with pages you can change today:

Look for old category labels, outdated feature lists, and examples that attract the previous customer.

For some profiles, an edit will not be enough. If your reviews sit inside an old category on G2 or another marketplace, you may need a new category listing and new reviews there.

Use the sources appearing in actual AI answers as your queue.

For each one, ask:

We started with the handful of pages that appeared repeatedly in commercially important answers. Updating those matters more than emailing every listicle author who has ever mentioned the company.

Some third-party pages will never change. Fresh, credible evidence can gradually outweigh them and give AI systems better sources to retrieve.

That includes:

You don’t directly write your Wikipedia positioning. You earn the independent sources its editors can use.

After each batch of changes, run the original prompt set again.

Compare the full distribution with your baseline.

Which descriptions are becoming more common? Which stale pages still appear? Where are competitors still recommended instead?

Then repeat the source audit.

By this point, the strategy was clear. The bottleneck was the volume of execution.

Correct one profile. Update one forgotten page. Contact one publication. Earn one new citation. Repeat across the web.

This won’t scale through a heroic weekend.

So I’m turning the repetitive parts into experiments:

Some work will stay manual because automation without judgment would quickly turn into spam. The research, prioritization, drafting, follow-up, and measurement loops are better candidates for systems.

I’ll break down those systems in future issues, including the prompts, workflows, numbers, and whatever fails.

Publishing a new hero section does not erase the older descriptions already spread across the web. Those descriptions accumulated over years, and the new one needs time and deliberate distribution to catch up.

I now treat repositioning as an identity migration with five steps:

That is the experiment I’m running now.

This newsletter will be the operating log of running Writesonic with AI, with product, growth, and marketing at the center. I’ll also cover sales, support, customer success, finance, and the unglamorous internal workflows where AI either creates leverage or creates more work.

I’ll share the implementation, numbers, and failures instead of polishing every experiment into a success story.

If that’s useful to you, subscribe. And send this issue to the founder or marketer whose company has changed faster than the internet remembers.

PS. Have you fought an inherited answer after a repositioning? Reply and tell me what moved it. I’m collecting patterns across companies, engines, and categories.