What Responsible AI Content at Scale Looks Like
TL;DR
- AI search queries average 60 words in ChatGPT versus 4.5 in Google. Most content strategies haven't caught up to that gap.
- The brands showing up in AI answers aren't producing the most content. They're producing content specific enough, accurate enough, and structured enough to cite with confidence.
- A content operation built for AEO starts with listening, not writing: sales calls, buyer reviews, competitor gaps. That ensures the decision-making engine is still human.
When I met Josephine Cahill, she told me something about their content operation that really stuck with me:
"Why should I be bothered to read what you can't be bothered to write?"
That question has driven every decision she’s made in building an AI-driven content operation. Josephine leads web for Oyster HR, a global employer-of-record platform operating across 180 countries, and she's been thinking about AI content longer and more critically than most. When we met in San Francisco during Flow TV week, she walked me through what she'd built. I told her the rest of the world needed to see it.
The place to start is with a mistake.
Oyster's team was using the Claude MCP for Webflow to expand their internal linking strategy. The instructions were reasonable. The scope was not. Before anyone caught it, the tool had quietly rewritten around 80 articles in their 800-piece library. A site snapshot saved them, but fifteen days of cleanup followed.
Josephine describes their content library the way you'd describe a storage unit on Storage Wars: "What are we breaking into from the Oyster archives? Is it going to be good for you? Is it going to be bad for us? We're going to find out." That’s pretty hilarious, but it's also a pretty good description of what AI search does to content you've forgotten you published. The stale pricing page, the outdated compliance guidance, the blog post from 2020 that's still getting cited are still looming out there.
What Josephine built after that experience is one of the most rigorous content operations we've seen at this scale, and one of the most directly applicable for enterprise marketing teams trying to figure out what responsible AI production actually looks like in practice.
Why most enterprise content teams are optimizing for the wrong search behavior
Google queries average around 4.5 words. ChatGPT queries average around 60. That gap matters more than most content strategies currently account for.
Short-tail keyword logic produces content that answers simple, direct questions efficiently. It works for traditional search. But in AI search, the queries are longer, more conversational, and often more specific than any keyword research tool would surface. A buyer asking an LLM about global hiring compliance isn't typing "employer of record." They're asking something closer to: "What are the risks of hiring a contractor in Germany without an employer of record, and how do companies typically handle it?" Super specific, and much more nuanced.
The content that gets cited in response to a query like that isn't the content that ranks for "employer of record Germany." It's the content that answers the question with enough specificity and accuracy that the model trusts it.
The brands showing up in AI search are not the ones with the most content. They are the ones whose content is specific enough and accurate enough that an AI system can extract a confident answer from it without hedging.
This is the core argument behind Edgar Allan's approach to Visibility Engineering & Optimization: brand clarity, site architecture, and content structure are not separate workstreams. They are the same problem. A content team producing high-volume output without a clear brand position and structured content architecture is building noise rather than AEO visibility.
The real reason your AI content workflow might be hurting your numbers
There's a version of AI content production that looks sophisticated from the outside but is quietly damaging from the inside. Lots of data sources, lots of automations, and lots of output, but with no clear system for knowing what's accurate, what's on-brand, and what a buyer really wants or needs to read.
Oyster's MCP incident is an extreme version. The more common version is subtler: content that sounds like the brand but is slightly off, pricing that's outdated, or compliance guidance no one in legal has reviewed. For most SaaS companies, those errors are embarrassing. For Oyster, which publishes content about employment rights, maternity leave, and unionization laws across 180 countries, they're an ethics problem. That framing changed how Josephine designed her entire workflow.
Every piece of content her team generates in AirOps routes to Asana as a set of subtasks for brand and legal review before it publishes. That way, the model doesn't decide when something is ready; people do.
What a content operation built for AEO looks like
Josephine's workflow runs in three phases that most content operations collapse into one: listen, decide, produce. Keeping them separate is what makes the whole system work.
Listen at scale.
Every month, Oyster's team runs automated analysis across Reddit threads, G2 and Capterra reviews, and competitor sites to surface what buyers are saying about employer-of-record services. They also run analysis across every Gong sales call, extracting the integrations buyers ask about, the countries they mention, and the concerns that come up repeatedly. All of it flows into a unified monthly report that feeds their content roadmap sessions.
Notice what it doesn't feed, though: a content generator.
As Josephine puts it, "The data informs our roadmap, and then we're making decisions on what we're publishing. The decision-making engine is still human." Listening at scale and producing at scale aren't the same operation. Conflating them is how you end up with an AI slop cannon.
Decide with discipline.
Once the roadmap is set, Oyster decides for each piece whether to create something new or refresh an existing one. Refreshing is often the higher-value move for AEO, because content that already has authority and structure can be significantly improved with updated information, better sourcing, and tighter brand voice, without the indexing lag of a new URL.
Their refresh workflow pulls from three sources simultaneously: their internal brand kit and voice guidelines, their knowledge base (including 1,200-page internal legal training documents their compliance team uses), and a pre-approved list of government sources scraped for recent updates. The output is a structured draft, and it goes to human review before anything else happens.
Produce with constraints.
Content generation happens inside tight constraints: specific models for specific tasks, brand kit variables injected at each step, and keyword and outline analysis pulled from top-ranking results before a word is written. Every step is logged. Every output gets human review before it moves forward.
Josephine's AirOps workflows look intimidating when you see them. They're long and detailed, full of custom variables and conditional logic. They also started simple. "You add complexity where the output keeps failing," she told me, "and you leave simplicity where it keeps working." Eight months of iteration is what made them what they are now.
The content flywheel that compounds
One of the most immediately replicable parts of Oyster's operation is what Josephine calls the content flywheel: a workflow that takes a single webinar or live conversation and turns it into a blog post, a LinkedIn post for Oyster, a LinkedIn post written in the partner's voice, and an email snippet for their newsletter. All from the same transcript. All reviewed before publishing.
The partner post is the part worth pausing on. After every webinar, Oyster's team scrapes the partner's website and LinkedIn, extracts their brand voice, and writes a post as if the partner wrote it themselves, ready to publish with one click. The partner doesn't have to write anything; they just decide whether to hit publish.
This is a distribution play, not a volume play. A webinar that stays on Oyster's channel reaches Oyster's existing audience. A webinar that gets posted by the partner reaches an entirely different audience, with zero additional lift on the partner's side. That compounds over time in ways that more blog posts simply don't.
A single piece of original research or a single live conversation, properly processed and distributed, is worth more to AEO visibility than a dozen generic articles. AI systems cite sources that demonstrate real expertise. A properly structured and attributed webinar transcript is primary evidence of expertise, whereas a polished but generic blog post is not.
What this means for how enterprise teams should think about AI content
The teams winning in AI search right now are the ones who have been the most deliberate about what they're optimizing for.
The question worth asking is: what would an AI system need to see from our content to cite us confidently on the topics that matter most to our buyers? That question leads to investment in proprietary data, brand clarity, content architecture, and human review processes that are actually followed. It leads away from volume-first thinking.
At Edgar Allan, this is the argument at the center of our AEO audits and visibility work: most enterprise content teams don't have a content volume problem. They have a brand clarity and content structure problem that more volume won't fix. The brands showing up in AI search answers made it easy for a model to understand exactly what they do, who they do it for, and why they're credible. That clarity starts with brand, runs through site architecture, and shows up in the specificity of the content itself.
If you want to understand how your brand currently appears in AI-generated answers and where the gaps are, our Profound-powered AEO gap analysis is the right starting point.
Watch the full conversation with Josephine Cahill on YouTube. If you want to understand how your brand currently appears in AI-generated answers, start with an AEO audit.
FAQ
How is AI search behavior different from traditional Google search, and why does it matter for content strategy?
AI search queries are significantly longer and more conversational than traditional Google queries. Oyster's team tracks this directly: ChatGPT queries average around 60 words, compared to roughly 4.5 words for a Google search. Content optimized purely for short-tail keywords will often fail in AI search, not because it's bad content, but because it was designed to answer a different kind of question. The content that performs is built to answer complex, multi-part questions with enough specificity that a model can extract a confident, accurate response.
What is the biggest operational risk when scaling content production with AI?
The biggest risk is undetected high-volume output that modifies existing content without clear human review in place. Oyster's experience of having 80 articles quietly rewritten by an MCP tool is an extreme example, but subtler versions show up constantly: outdated pricing surfaced by AI search, compliance information that hasn't been reviewed, brand voice that drifts across a large library. The fix is to design human review into the workflow at the structural level, not as an afterthought.
What makes content more likely to be cited in AI search answers?
AI systems favor content that is specific, accurate, clearly attributed, and structured to make it easy to extract a direct answer. That means concrete examples over general claims, proprietary data over aggregated information, clear entity association (named authors, named companies, named methodologies), and FAQ-style structures that answer questions directly. Content that hedges or relies on implicit rather than demonstrated expertise tends to get passed over in favor of sources that make the model's job easier.
How should enterprise teams think about balancing content volume with content quality in an AEO context?
The volume-versus-quality framing is a distraction. The more useful frame is specificity. A large library of generic content produces less AEO value than a smaller library of highly specific, well-structured content built around the actual questions your buyers are asking. Oyster's approach, running monthly analysis of sales calls and buyer reviews to surface real buyer language before deciding what to write, is a good model. The production is disciplined and reviewed. The roadmap is informed by listening at scale.
What role does brand clarity play in AI search visibility?
Brand clarity is foundational to AEO performance. If an AI system can't accurately and consistently describe what your company does, who it serves, and why it's credible, it will either misrepresent you or omit you from answers where you should appear. This is a brand problem, and content alone can't fix it. The teams performing well in AI search have done the upstream work: clear positioning, consistent messaging across the site, and content that demonstrates expertise rather than just asserting it.
How do you build a content workflow that scales without losing brand voice?
The teams doing this well are building brand voice into the workflow itself, not relying on post-production editing to catch drift. That means a detailed brand kit with writing examples, voice guidelines, and tone specifications injected directly into content generation steps. It also means human review at specific checkpoints, not just at the end, so feedback loops back into the workflow and improves it over time. Oyster tracks which AI-generated pieces perform best on engagement and organic traffic, then trains future models on those examples. The workflow gets smarter through iteration, not through manual correction.