Twenty days after Dr Anna Harrison rebuilt her website for AI engines, a cold email arrived using phrases that existed nowhere except the technical pages she’d written for machines. No human had ever seen them. An AI had read them, retold them to a buyer, and the buyer had come looking.
That sequence: read-retold-chosen, is now how a growing share of buying begins. Buyers open ChatGPT, Claude or Gemini before they ever reach a website, and the answer those tools give decides which businesses make the shortlist and which never come up at all. Roughly a third of buyers already start there. The first conversation about your business now happens without you aware of it.
Harrison works at the centre of this shift. A behavioural scientist and the founder of RAMMP, a brand trust diagnostic platform, she has helped more than 900 businesses strengthen how customers engage with them and recently released a free AI plug-in that runs inside ChatGPT and Claude, showing business owners in real time where trust breaks in their buying journey.
In this Q&A, her argument is blunt: visibility inside AI tools costs little relative to what it returns, the early movers gain an advantage that compounds, and structured content surfaces fast. Her own site appeared in ChatGPT answers within days.
A common objection to investing in AI visibility is that click-through rates from LLMs are still tiny and aren’t replacing the drop in organic traffic. If the revenue isn’t showing up yet, what is the business case for investing now?
The click-through-rate framing assumes the AI is a traffic referrer like Google was. It isn’t. The LLM is increasingly the destination. The user gets the recommendation, forms an impression, and approaches you already pre-sold, or already dismissed. The conversion you can’t measure is the conversation that never happens because ChatGPT recommended your competitor. That’s not a CTR problem. It’s a share-of-mind problem, and it doesn’t show up in your analytics.
Second, the Day-20 cold email I showed isn’t anecdotal, it’s the leading indicator. The language LLMs surface to humans becomes the language humans use back to you. If your category vocabulary ends up in the LLMs’ answers, you become the category. If a competitor’s vocabulary wins, they become the category and you spend the next three years explaining the difference.
Third, and this is important, the cost to deploy is low. Forty pages of structured content plus JSON-LD takes a sprint, not a quarter. Compare that to a six-figure SEO retainer producing diminishing returns. The business case isn’t “this will replace your traffic.” It’s “this is cheap insurance against being structurally invisible in the only channel growing 30%+ year on year, while every other channel holds flat.” If LLM traffic stays small forever, you’ve spent a sprint. If it doesn’t, you’ve built moat.
For a business with limited time and budget, SEO and AEO compete for the same resources. How should a resource-constrained company balance the two?
If your focus right now is on SEO, I’d pause that for a sprint and get this up and going, or run them in parallel for a month if it’s an industry where you can’t lose the SEO traction. You’ve got a solid case to make, because the earlier you do it, the faster you’ll start showing up. If you think about the nature of LLMs, the unit of consumption is a web page, and they consume all those pages into a giant soup. The earlier your page goes in, the more chance it has of being propagated throughout, and the longer it’s in there, the more it works for you.
Two terms come up repeatedly in this work, ontology and knowledge graph, and they are often used interchangeably. Is an ontology the same as a knowledge graph?
Close, but not identical. An ontology is the schema, the rules defining what types of things exist (Standards, Definitions, Methodologies, Decision Controls, Comparisons) and how they relate to each other. A knowledge graph is the instance, the actual filled-in network of those types with your specific content.
The ontology is the architecture. The knowledge graph is the building. When I built RAMMP’s, I defined the ontology first, then populated the graph. Doing it in that order keeps the structure clean. You can build your own ontology here.
Building the pages is one task; getting AI crawlers to actually read them is another. How do you get the LLMs to read these pages?
You don’t trigger them. You make the pages easy enough to find that they self-trigger. Three mechanisms do the work. The page is on the public web and crawlable, with no auth wall and not blocked in robots.txt. The JSON-LD schema in the header signals that this is structured content worth eating. And llms.txt points crawlers at it.
Beyond that, LLM crawlers run continuously. Anthropic, OpenAI, Perplexity and Google all run their own, and there’s no submission process. The fastest acceleration is to be referenced from a high-authority source such as Wikipedia, Reddit or a trade publication. That pulls crawlers in faster than waiting for them to find you organically.
In traditional SEO, page depth and sitemap position mattered. Are pages that aren’t in the sitemap still accessible to AI tools, and do LLMs prioritise pages higher in the hierarchy?
Yes, fully accessible. LLM crawlers don’t rely on sitemap position the way old SEO crawlers did. They prioritise three things: a well-formed JSON-LD schema in the header, technical precision in the content, and clean cross-references between pages.
A deeply nested page with crisp structured data will outperform a top-level page full of marketing copy. Hierarchy matters less than schema quality.
Some pages end up serving both humans and AI crawlers, which can mean similar copy in more than one place. Is there a risk of duplicate content hurting SEO or AI authority when a page serves both audiences?
For some of the pages inside the knowledge index, we’ve recognised that a page is customer-facing and worth deploying, so we’ve added the design elements and linked it. There are some pages that serve two purposes.
As you go through this process, if some of your comparison pages or decision control pages double up with something customer-facing, you put a bit of extra energy into that page. Make it a human-facing page, but make sure you have all your definitions and those elements included in the header, with a JSON-LD schema that ties it back to the knowledge graph.
Reviews, forum threads and other user-generated content sit outside a brand’s control. How does user-generated content affect your ability to control the AI narrative?
I think it’s orthogonal to this issue. You have no control over user-generated content, and it’s also not structured data, so its weighting as it’s consumed by the AI carries far less priority than something structured in this way.
For consultants, advisors and founders, personal profiles feed into how AI tools describe a person. Is LinkedIn important for AI?
Yes, for two reasons. First, LinkedIn pages and posts get indexed, especially company pages and long-form articles, and they carry authority weight because LinkedIn is a recognised trust source. Second, individual LinkedIn profiles are where LLMs verify who someone is. When ChatGPT or Claude is asked about a person, the LinkedIn profile is often the canonical source it cites.
For personal-brand businesses like consultants, advisors and founders, keep your LinkedIn profile current with the same technical language you use in your knowledge index. The cross-reference between your site’s About page, your knowledge index and your LinkedIn profile is one of the strongest authority signals you can build cheaply.
This work sits between marketing and technical disciplines, which makes ownership unclear. Whose job is it, the developer, the digital marketing agency, or the content and marketing team?
If you’re working with an agency that’s doing great SEO work for you, the next question to ask is whether they can even do this piece of work, because it’s quite a different skill set. The skill to create these standards, methodologies and definitions is a lot closer to someone who can do technical or architectural design in a tech space than to someone who can write marketing copy. At the end of the day, if you’re running a business and it’s your business, it’s your job to find someone who can do this piece of work for you. It’s pretty crucial.
Before you can build pages around the right questions, you have to know what people are asking. How do you figure out what keywords and topics are being searched for inside AI?
This comes back a little bit to SEO. As a practical example, if you search for “best wineries in New Zealand” in Google and scroll down to the “people also ask” questions, that’s your starting point. I would surface those questions straight on the website.
We have a set of query-capture pages, and these come from the “people also ask” questions. We take the question from Google, build a page around it, and integrate it with the knowledge schema we’ve created. Everything connects. We drop a new node into the graph, connect it back to the other nodes, and ship it
