AI Search Is Computing Its Own Answers From Structured Data, and That Opens a New Way to Get Found

For twenty years, getting found meant getting ranked. A page earned its way into the top ten, and buyers picked from what they saw. That arrangement is coming apart. Assistants and AI Overviews increasingly assemble a response rather than hand back a list, and the signals they read to do it are not the ones most marketing teams spend their budgets on. The result is a widening gap between how brands optimize and how machines retrieve, and the advantage is moving to whoever feeds those machines cleaner structured data.

Brie Moreau is the Founder and CEO of White Light Digital Marketing, a data-science-first SEO agency that has spent close to two decades in some of the most competitive search markets, from crypto to online casinos. His team ran a study on 11 million AI citations and more than 2,000 hours of analysis to understand how ChatGPT chooses what to cite, then pulled an open web repository of more than 600 terabytes to map the structure underneath. What he found reframes what SEO is even for.

The shift he keeps returning to sits beneath content and links, in how a machine reads a page at all. "Instead of referencing content, it's computing answers. It's a very different shift," says Moreau. Once a system is computing rather than quoting, the question stops being which page ranks and becomes whether a brand's data is legible enough to be used at all.

From retrieving to computing

Moreau traces the change to Google's compression research, an algorithm called Turboquant that Google published in March and has been folding into its own inference. He says it lets the search stack read faster and wider, and points to what he describes as a decoupling between AI Overview citations and the traditional top ten results, a correlation he says has slid from around 80 percent to 30 percent in a year. "Google isn't just relying on the top 10 results," says Moreau. He sees it as part of the move toward autonomy in how search operates.

The mechanism, the way he describes it, is closer to solving a puzzle than fetching a document. "Imagine a crossword puzzle where you have 20 words and only one letter for each word, and somehow it gets it right every time," says Moreau. A model with the right structured inputs can assemble an answer from fragments. "When you write that question and get an AI Overview response, it's using Turboquant, it's using schema. That's what you need to understand," he explains.

Breadth without depth

When Moreau's team scrapes an entire keyword cluster, the top twenty results plus the hundred pages behind each, a pattern holds across industries. Roughly half the sites carry no schema at all. The rest tend to deploy the same handful of tags from a shared vocabulary everywhere and call it done. "They've got the breadth covered but not the depth," says Moreau.

That gap is the opening. He argues most teams treat structured data as a box to check rather than a surface to build on. "You think your schema is deployed. It's not," says Moreau. Because the markup is invisible to readers, he adds, there is no content-rewriting penalty for going deeper. "There's a ton of schema out there that people could be using and aren't," he says, pointing to programmatic work for one client, the online casino Stake.com, that pushes granularity well past the usual template.

Wikidata as a source of truth

Depth alone is not the whole play. Moreau points to an open knowledge base called Wikidata, which assistants lean on when they need a reference point. "Wikidata is basically Wikipedia for LLMs. It's where the LLMs go to reference data," says Moreau. Entities there carry their own identifiers, and a brand can claim or create them, then tie them back to a verifiable source.

Few companies have touched it. Most large brands, he says, have fewer than twenty data points there, if they appear at all, which leaves the field open to anyone willing to do the work. Paired with deeper markup, he claims the upside is steep. "We can easily 200x any website's schema in all industries," says Moreau. He flags a caution alongside it: because Wikidata entries can point to third-party sources, the chain of trust is only as sound as those references, something his own team demonstrated by getting a referenced page changed for a fee. "It's cool, but it's also scary," says Moreau.

When the machine gets it wrong

For all the retrieval mechanics, Moreau says the reason executives actually move is reputation. A head of SEO can love a schema pitch and still watch it stall, until leadership sees the problem firsthand. "All of a sudden the CEO checks and the information about their company is incorrect on ChatGPT, and then everything moves really fast," says Moreau, who puts the hallucination rate he runs into at roughly 37 percent.

He frames the fix less as a growth lever and more as managing reputation off-site, correcting what a model asserts before it hardens into the default answer. The exposure, he argues, is universal. "Every single company is going to have something with some form of negative sentiment," says Moreau. The work of showing up in answers increasingly runs through structured data a model can trust.