AI visibility isn't built overnight. It's built by many independent sources, over time.
You Can't Task Your Way Into AI Visibility: Here's What Actually Works
"We need to influence the parametric side."
It sounds like something you can put on a to-do list. It isn't. And once you understand why, it changes how you think about AI visibility search altogether.
This comes out of a piece by Duane Forrester on Search Engine Journal, and he makes a point that most AI marketing conversations gloss right over: AI models "know" your brand in two different ways, and only one of them cares what you publish this quarter.
Two Layers of AI Visibility
Retrieval is the layer everyone's already optimizing for. It's what an AI model looks up the moment someone asks it a question - your website, your structured data, a recent article, a live citation. Most generative engine optimization (GEO) advice lives here, and for good reason: it actually responds to work you do today.
Parametric knowledge is the other layer, and it's a different animal. It's what a model already "knows" about your company before it looks anything up, because that knowledge got baked in during training. You don't build this one by publishing more. You build it by getting other people - journalists, reviewers, analysts, your own customers - to describe your company independently, in their own words, over years.
Most companies budget for retrieval and quietly assume it's doing double duty. It isn't.
Why More Content Isn't the Answer
This is the part that trips people up. Publishing more about yourself barely moves the needle on parametric knowledge. The research on how these models are trained is fairly consistent on this: a topic gets "known" when many different sources describe it in many different ways, not when one source says the same thing a hundred times.
Ten outlets mentioning your company in ten different ways will do more for you here than your own blog cranking out a hundred posts about itself.
And there's no office to submit your brand to. The datasets these models train on are massive. No single company's website makes up a meaningful slice of one, no matter how active the blog is. Worse, a lot of where people actually talk about brands on forums, review sites, Reddit threads, etc. This tends to be under-represented in that training data to begin with. The places your customers hang out and the places these models learn from aren't always the same places.
Who Already Did This Work (Without Knowing It)
Here's the part worth sitting with for a second: almost everything that builds parametric knowledge already sits inside your marketing org. PR, analyst relations, community management, reviews, press coverage. None of it is new work. It just didn't have a name like "AI strategy" attached to it, because that category didn't exist yet.
The catch, and it's a real one, is that none of these functions ever fully controlled the output. A journalist decides what to write. A customer decides what to say in a review, and whether to say anything at all. That's always been the deal with earned media: you can enable it, you can make it more likely, but you can't author it yourself. Turns out that's exactly the kind of content that shapes what a model learns.
There's a timing wrinkle too. What a model knows about you today often reflects work, or neglect, from years ago . . . sometimes done by people who don't even work at the company anymore. Parametric knowledge doesn't move at campaign speed. It moves more like reputation does.
What This Means for Your Strategy
Don't measure AI visibility only by how much you publish. Track it by who's talking about you independently — press mentions, reviews, analyst coverage — because that's the signal that actually matters here.
Treat PR, reviews, and community work as part of your AI visibility strategy, not separate from it.
Run an AI visibility audit. Look at what AI models currently say about your company and compare it to how you'd want to be described. The gap tells you a lot. Whether you're behind on retrieval, on parametric knowledge, or both. And those two problems get fixed in completely different ways.
Set expectations accordingly. Retrieval-layer work (structured data, fresh content, citations) can show results in weeks. Parametric-layer work is closer to a multi-year investment - plan and report on it that way.
This post builds on Duane Forrester's original analysis on Search Engine Journal, worth reading in full if you want the research trail behind it.
Curious about what AI models currently say about your company? Get in touch with RSO Consulting to run an AI visibility audit and find the gap between your retrieval-layer and parametric-layer presence.
Further Reading
For anyone who wants to dig into the research behind this:
Kandpal et al., ICML — how training data volume affects factual recall: https://arxiv.org/abs/2211.08411
Mallen et al. — scaling limits on long-tail factual knowledge: https://arxiv.org/abs/2212.10511
Allen-Zhu & Li — why varied phrasing matters for what models retain: https://arxiv.org/abs/2309.14316
Elazar et al., "What's In My Big Data" — composition of major training datasets: https://arxiv.org/abs/2310.20707
Common Crawl — crawler methodology and coverage gaps: https://commoncrawl.github.io/cc-crawl-statistics/plots/domains
Cheng et al., Johns Hopkins — effective knowledge cutoffs in trained models: https://arxiv.org/abs/2403.12958
