What surprised me after interviewing 20 content managers

When I joined Sighted, the first thing I wanted to do was talk to content managers. Not about us. About their week. I asked each of them to walk me through one real piece of content, from the idea to the published page, and to show me where AI sits in that chain and where it doesn't.

After 20 conversations across the US, France and Canada, my product manager's notebook was full and most of my assumptions were wrong. Here are the five things that stuck.

They write less now. Most of their time goes into building AI content workflows.

Of the 20 content managers, 14 have built their own agents or AI content workflows. Not a prompt saved in a doc. A system: a research agent, a drafting agent, an editing agent, a Notion database that feeds them, a connection into the company's own data.

I didn't expect that. I've worked with excellent content managers for years and I filed them, unfairly, under "literature": people who care about words first and tooling last. I went in expecting pushback, maybe some nostalgia for hand-made quality. I heard none. They moved faster than many engineers I know, and towards the same goal they always had: tell the company's story to drive traffic and revenue. Only the audience changed. They used to write for humans and Google. Now they write for LLMs, Google and humans, roughly in that order.

One pattern held on almost every call: they buy tools to measure and build tools to produce. Nobody built their own visibility tracker. Nobody bought generic content production off the shelf either.

And the team shrank around them. Where there used to be a content team, there is now one person and a stack of agents. Hence the line at the top.

The only content worth writing is what only you can say. It's called information gain.

The most interesting thing I heard was about information gain, what Google also calls non-commodity content: the thing that gets a page cited by an LLM. Once writing costs an hour, writing stops being the hard part. The hard part moves upstream: finding something worth saying.

Avoiding AI slop came up on every call, and so did making AI sound human. But what mattered more was what the content says, not how it sounds: a page gets picked when it adds something the model hasn't already read a thousand times. Sameness is exactly what an AI answer filters out, so the generic article that used to rank on volume now gets skipped.

The best content managers I spoke with have turned this into a sourcing discipline rather than a writing one. They sit on sales calls and note the exact words prospects use for their pain. They mine customer stories. They publish their own data, even small sets. They get the founder to take a position and put a name on it.

Measuring GEO impact is hard

Asked how they measure the impact of AI citations, 10 of the 20 content managers called it a black box. Their words, not mine.

GEO is less predictable than SEO, and the rules aren't published. With SEO you had a keyword, a ranking and a click. With AI search you have a mention you can't see, inside an answer that changes every time someone asks. The tools that track it disagree with each other.

The harder problem is tying any of it to pipeline. Nobody clicks in an AI answer. Pew Research measured it in July 2025: people clicked a link inside an AI summary on about 1% of visits. What they do instead is read the brand name ChatGPT gave them, remember it, and google it a week later. Google Analytics files that visit under direct traffic, and the channel that earned the lead stays invisible.

So several teams have gone back to the oldest tool in marketing: a "How did you hear about us?" field at sign-up. Low-tech, slightly embarrassing, and the only honest attribution most of them have.

Content refresh: a big cost, with no system

Of the 20 content managers, 12 raised content refresh as a worry, and none had an automated way to handle it.

This is the one I was waiting for. What I've learnt helping brands with their AI visibility is that LLMs reward consistency: the same facts, stated the same way, on every page, all the time. The research points the same way. A model tends to side with the majority of its sources, and a claim repeated across many pages beats a claim made once on an authoritative one. A price that changed in March and still reads old on four comparison pages gives the model a reason to doubt everything else you say.

Here's how it usually plays out. A SaaS company raises its price and ships two features in the same quarter. The pricing page is updated the same day. The twelve blog posts that quote the old price, the three comparison pages that list the old feature set and the help centre article from 2024 are not. Six months later, ChatGPT describes the product with the 2024 price and without its best feature. Nobody did anything wrong. It just wasn't anyone's job.

Everyone I asked handles this by hand: a spreadsheet of several hundred articles, a clean-up twice a month, a spring review of every country page. AI solved production. It never touched maintenance.

Reddit is this year's obsession. Distribution is the lesson.

Reddit came up on call after call. For a while, AI answers gave Reddit threads a far bigger voice than their share of the web, because that's where real people compare tools in their own words. So teams went looking for threads, seeded answers, and watched their mentions.

I don't think the Reddit wave lasts. One content lead already described it in the past tense.

What stays is the lesson under it. The pages that feed an AI answer are mostly pages you don't own: LinkedIn, YouTube, Reddit, review sites, other people's comparisons. Publishing on your own site is the entry ticket. Distribution is the game, and it is a bigger part of the content manager's job than it was two years ago.

What I'd do with this

If I ran content at a B2B SaaS company today, these five lessons would change my week like this.

• Spend the hour I saved on writing finding what only we can say: one sales call a week, one customer's words, one number from our own data.
• Add the "How did you hear about us?" field at sign-up tomorrow.
• Treat distribution as half the job, and pick the two outside places where our buyers actually compare tools.
• Give someone the job of keeping the facts true across every page, and give them a system, because a spreadsheet of 700 articles is a job nobody finishes.

That last one is what we build at Sighted: a system that reads every page you have, finds where they disagree about your product, and keeps them saying the same true thing. If you have a few hundred pages and nobody owns keeping them true, write to me and I'll show you where your site contradicts itself.

Thank you to everyone who was kind enough to show me how they work today.