Not my favorite kind of pub (McSorley’s, the Dead Rabbit, the Full Shilling) but my second favorite: publishers. I like the value exchange that underpins the historical model of digital publishing: you read my article, I put an ad in front of you, we both get what we want. That model’s been under increasing duress as “Google Zero” (people never going to the website the content powering the chatbot draws its knowledge from) has come to be, with increasing amounts of time and attention being dedicated to engagement with LLMs themselves. The clear loser in this is the guy who was gonna make a fraction of a cent putting an ad next to the content you’ll no longer see. Out of this a new discipline has arisen: Generative Engine Optimization (GEO) or AEO (Answer Engine Optimization), a sort of successor to the SEO era that preceded it. The mechanics of needing to know what an agent will say about your brand, and the agent-makers’ desire to not allow itself to be influenced, create a fascinating dynamic.
GEO has been sold to me at every conference this year as a brand-new discipline but diffing into it I find the latest round of a very old game between people who build ranking systems and people who want to be ranked. The rules change every round, but the players and their incentives don’t.
So let’s look at the last round and see what it predicts about this one.
Round one: the content farm
In 2006, Richard Rosenblatt and Shawn Colo founded Demand Media on an elegant idea. Instead of guessing what people wanted to read, the way magazines always had, they’d watch what people typed into Google and make exactly that. Algorithms scanned search queries for long-tail phrases with advertising value. A network of freelancers, reportedly paid around $15 an article, produced the answers at industrial scale. eHow and Livestrong became giants of the how-to and health web. When the company IPO’d in January 2011, it reportedly had over $200 million in revenue and over 100 million monthly visitors.
It was search arbitrage. Buy content cheap, rank it on Google, sell ads next to it. It worked beautifully as long as one company in Mountain View was fine with it.
A few weeks after the IPO, in February 2011, Google rolled out the Panda update, aimed squarely at low-quality content. It touched roughly 12% of queries. eHow’s traffic fell off a cliff, and Demand Media spent the next several years deleting millions of its own articles, trying to convince Google it had changed. Eventually it renamed itself Leaf Group and became an art marketplace. My favored tabloid at the time, Gawker, covered it mercilessly, as the quality of content you get from a fifteen dollar article made to appease the midnight googlings of the inane is not the highest.
The lesson people usually draw is “don’t make junk.” The interesting lesson is structural. Demand Media had built a business on the tolerance of a single counterparty, and that counterparty had every incentive to withdraw its tolerance once the arbitrage started degrading its product. The farm and the search engine weren’t partners. They were adversaries who happened to share a customer.
The adversarial mechanic
There’s a principle I keep coming back to, Goodhart’s law: when a measure becomes a target, it ceases to be a good measure. Every ranking system is a measure of quality, and every optimizer is trying to make it a target. The engine then has to update the measure to keep it meaningful, the optimizers adapt, and round and round we go. Biologists call this a Red Queen race, where you run as fast as you can just to stay in place.
SEO spent twenty years in that race. Keyword stuffing gave way to link schemes, link schemes to content farms, content farms to “E-E-A-T” (Experience, Expertise, Authoritativeness, Trustworthiness). Each optimization worked until it worked too well, at which point it became a target for the engine’s next update.
GEO is the same race with a new referee, or really four or five new referees. ChatGPT, Perplexity, Claude, Gemini, and Google’s AI Overviews each weigh evidence differently, and none of them publish their rules.
Round two: what GEO actually is
Strip away the conference-stage branding and most of what GEO agencies sell today is perfectly reasonable hygiene. Structure your content so a machine can parse it. Use schema markup. Answer the question in the first sentence. Put comparisons in tables. Make sure your brand is described consistently across the web, so a model can recognize you as one entity. The academic GEO research shows that content with citations, statistics, and quotations gets surfaced more often in AI answers. Nothing about that is gaming anything. It’s just clear writing for a new kind of reader.
But there are two moves at the edges that are more interesting, and both rhyme with Demand Media.
The first is placement on borrowed credibility. Generative engines lean heavily on third-party sources over what a brand says about itself. A brand can’t easily become a trusted domain, but it can buy a place on one. Sponsored content on heavily indexed publishers is the obvious play, and I think it’s the strongest commercial idea in the space right now. It’s also the most fragile. How cheap is the influence? Researchers recently showed that a 13-word snippet about a restaurant that didn’t exist, posted on Reddit, was enough to get it recommended by LLMs.
The second is serving the machine a different page than the human sees. Some publishers have begun serving more efficient markdown file versions of articles for agents, which now make up over half of web traffic. These markdowns can include their own versions of ads, “brand fact” blocks that human readers never see. Others sell brands the ability to serve AI retrieval bots an optimized version of their own sites while Google and Bing see the normal one. The industry calls this “agent-first” publishing. In 2011 we had a different word for showing the crawler something other than what the user sees.
Second-order consequences
The first-order consequence of GEO is brands showing up in AI answers. The second-order consequences are the ones worth planning around.
- The referee will respond, and faster this time. Panda arrived about five years after Demand Media’s founding. Perplexity took less than two weeks after the first reporting on TIME’s agent ads to block them from influencing its index and agents, calling the practice deceptive and warning publishers about their trust scores. The response cycle has compressed from years to days. Any GEO strategy whose value depends on an engine not noticing has a half-life measured in quarters.
- Labeled influence gets discounted. Engines trust third-party content because it’s third-party. The moment content is clearly paid, it inherits the credibility of an ad, not an article. Sponsored content on great domains will work best when it looks most like journalism, and that is exactly when engines will start looking hardest. Expect the market to settle into something like native advertising’s history, labeled and tolerated and priced accordingly, not a free ride on editorial trust.
- Infrastructure without demand is theater. A lot of “agent readiness” spending is going into the plumbing. Cloudflare found only about 3.9% of top sites even support serving markdown to agents on request. Ahrefs found 97% of llms.txt files received zero requests in a month. Brands are building on-ramps for traffic that mostly hasn’t arrived. It’s not wasted money exactly, but it’s the 2026 version of every company building a Second Life island or a house in Facebook’s metaverse, if the first ref is too old for you.
- Slop poisons the well, and the well is now the product. Demand Media’s junk degraded Google’s results. Today’s AI-generated content farms can do something worse: they can end up in the training and retrieval corpora of the models themselves. The engines’ incentive to aggressively discount anything that smells synthetic or paid will only grow, which pushes value toward sources that are expensive to fake.
- Measurement becomes the moat. When every optimization is temporary and every referee is opaque, the durable business isn’t the trick. It’s knowing, independently and continuously, which tricks are working and what the machines are actually being fed. The scarce asset in a Red Queen race is accurate information about where the course has moved.
Every paradigm shift tempts people to believe the old rules have been suspended. Sometimes they have. But the rule that whoever controls the ranking eventually polices the ranking has survived every shift I’ve studied. Optimize for the machine, by all means, but design as if the referee is watching, because they are.
Originally published on LinkedIn.