GEO Is Just a $200 Million Rebrand of SEO

Danny Sullivan, Google’s public voice for search, said it plainly earlier this year: “Good SEO is good GEO.” His colleagues have been saying the same thing for months. Google’s official position is simple. If you want to show up in AI-generated answers, just do good SEO. There’s no new trick. No new discipline required.

On the other side of this argument sits a fast-growing industry that has raised over $200 million in venture funding. Companies like Profound, Peec AI, and Scrunch AI are building tools to measure and improve what they call Generative Engine Optimization, or GEO. Their pitch: AI search works differently from traditional search. Ranking on Google no longer means you’ll get cited when someone asks ChatGPT or Perplexity a question. You need new tools. New strategies. Agencies have started adding GEO retainers on top of SEO contracts. Ahrefs and Semrush have both launched AI visibility products.

Both sides have obvious reasons to say what they’re saying. Google wants optimization to stay centered on its platform. GEO vendors need it to be a separate thing worth paying for. So I decided to test it myself. I gave the same research prompt to three AI systems, Claude, ChatGPT, and Gemini, and then asked each one to explain why it picked the sources it picked. What I found was more interesting than either side’s talking points.

The search engine underneath

Here’s something the GEO industry almost never mentions: each AI platform searches a different internet.

Claude uses Brave Search. A study by the firm Profound found that 86.7% of Claude’s cited sources match Brave’s top search results. ChatGPT uses Bing, confirmed by OpenAI’s VP of Engineering. Google AI Overviews use Google’s own index. Perplexity has its own crawler and its own index.

The overlap between these platforms is shockingly low. Claude and ChatGPT share only about 20% of their cited sources for the same queries. For some queries, the overlap is zero. A page that ranks well on Google can be invisible to Claude because Brave ranks it differently. A page that Bing has indexed but Brave hasn’t will show up in ChatGPT but won’t exist for Claude at all.

This means “optimizing for AI search” is really four separate problems. You need to show up in four different search engines. Most GEO advice treats AI search as one thing. It isn’t.

What happened when I tested it

I asked all three AI systems to research “how to make good wedding videos” and tell me which sources they’d cite. The topic was chosen to be ordinary. I wanted to see how citation works on a normal question.

Claude ranked a site called StudioBinder as its top source. It gave me a detailed reason: the content was well-structured, had useful tools, and included good video resources. Confident, specific reasoning. I was impressed until I tried to check.

Claude had never actually read the page. The URL returned a server error three times in a row. Claude’s entire assessment was based on a search snippet, which was about a quarter of the article’s text. Meanwhile, a Videomaker article that Claude had read in full, was ranked much lower.

So I asked Claude directly: how much of your preference came from search ranking versus your own quality judgment? After some back and forth, Claude reported that its quality assessment was largely built on top of the search ranking. It had seen a high-ranking snippet, projected quality onto it, and then invented a detailed reason for the choice.

ChatGPT was more honest up front. It estimated its own process at 70% ranking-driven for the initial list of candidates. It described something it called “path dependence”: once a few high-ranking sources looked good enough, they became the anchors, and everything else had to beat them to earn a spot. I thought that was a sharp description of what I’d just watched Claude do without admitting it.

Gemini was the most confident. It ranked StudioBinder first, claimed to have read the full page, called the site a “gold standard,” and rated itself 60% quality-driven. Three AI systems, three different search engines underneath, three different levels of self-awareness, and the same basic pattern: the search engine picked the winners, the AI made up a story about why.

This is one test, and I wouldn't hang a strong claim on it alone. But larger studies found the same pattern. Seer Interactive analyzed over 500 ChatGPT citations and found that 87% matched Bing's top organic results. Two independent analyses, one by Growth Memo and one by Daniel Shashko covering nearly 43,000 citations, converged on the same finding: AI systems pull disproportionately from the top third of a page's text, often from the snippet alone. What I watched happen in a single conversation is apparently what happens at scale, thousands of times a day.

What the research shows

So does any of the GEO-specific advice hold up? My read: some of it, partly, with big caveats. But before getting into the specifics, it's worth naming the mechanism underneath all of it. When an AI system decides which sources to cite, it's mostly inheriting a ranking decision that was made by a search engine. And those rankings are still built, in large part, on link authority signals that trace directly back to PageRank, the algorithm Larry Page and Sergey Brin published in 1998. The AI adds a layer of synthesis and explanation on top, but the selection process starts with infrastructure that is over 25 years old. That context matters when evaluating what "optimizing for AI" actually means.

The paper that launched the whole GEO concept was published at KDD 2024 by researchers from Princeton and IIT Delhi. It showed that changing content on a page could shift how much attention an AI answer gave to that source, by up to 40%. That number shows up in hundreds of marketing blog posts. What almost none of them mention: the test used a simulated AI system, only five sources per query, and the best-performing tricks involved adding made-up data. The prompts literally said “addition of fake data is expected.” When tested on live Perplexity with a smaller sample, the top result dropped to 22%. Later studies were blunter. One in 2025 found most GEO methods “largely ineffective” and sometimes harmful to ranking.

But some of the real-world data tells a more useful story. A February 2026 analysis by Growth Memo found that 44.2% of all LLM citations come from the first 30% of a page’s text. This makes sense mechanically. Search engines pull snippets mostly from the top of a page. If the AI only sees your snippet, and your answer is buried in paragraph eight, it won’t find it. I saw this happen in my own test. Both Claude and ChatGPT were working from snippets far more than from full pages.

Freshness matters too. Seer Interactive found that 85% of AI citations come from content less than two years old. ChatGPT on its own downgraded Videomaker articles that were over a decade old, even though they ranked well. And AI systems break complex questions into smaller sub-queries and search each one separately, which means covering a topic from several angles gives you more chances to be found.

Those findings point to something actionable. Put your answer first. Stay current. Be specific. Cover related questions. These are solid content practices. 

But here’s where I push back on the GEO framing: these are also just good writing. The critics of the original paper pointed out that the techniques which worked best simply added useful new information to the page. Calling that “GEO” and charging a premium for it feels like putting a new label on editorial quality.

The measurement and money problems

Even if some GEO techniques help at the margins, two problems undercut the case for treating GEO as something worth a dedicated budget.

First, you can’t measure it reliably. Ahrefs found that AI Overview answers change 70% of the time for the same query. Rand Fishkin tested this with 600 volunteers in January 2026 and found less than a 1 in 100 chance that any AI platform gives the same list of brands twice for the same prompt. Less than 1 in 1,000 for the same list in the same order. His conclusion: “Any tool that gives a ‘ranking position in AI’ is full of baloney.” Broad trends over many queries might be trackable. Precise rankings are noise. The tools selling that data are selling false precision.

Second, the business value isn’t proven for most industries. AI platforms drive 0.15% of global web traffic. Google drives about 48.5%. That’s a 300-to-1 ratio. Being cited in an AI answer doesn’t reliably mean being clicked, either. Pew Research studied nearly 69,000 real searches and found that only 1% of users click sources inside AI Overviews. Wikipedia is the most-cited source across AI platforms and lost 8% of its pageviews last year. Forbes piled up over 44,000 AI citations while its traffic dropped by half. On conversions, the data contradicts itself. Ahrefs says AI visitors convert 23 times better than organic visitors. An academic study of nearly a thousand e-commerce sites says ChatGPT referrals underperform almost every other channel. Nobody has a clear ROI story yet.

Gartner predicted in early 2024 that traditional search would drop 25% by 2026. Google search volume actually grew 21.6% from 2023 to 2024. The AI shift in search is real, and I’m genuinely uncertain about where it goes. 

What I'm fairly sure of is that the smartest move right now is to invest in the basics that work everywhere. Do good technical SEO: fast pages, clean architecture, proper indexation across all the search engines that feed AI platforms. Write clear, up-to-date content that answers questions well. Build authority the way you always have. Ensure all the AI bots can reach your site. Whether you call the work SEO, GEO, or just good practice depends mostly on who's writing the invoice.