The Answer Comes Before the Link Now: What GEO Actually Is (and What It Isn't)

Otlume GEO series, part 1 of 10 · Level: Beginner · By Otlume's GEO optimization expert

Here is a number that stuck with me.

Pew Research Center spent March 2025 watching how 900 U.S. adults actually browsed. People who got a Google AI summary clicked a regular search result on 8% of visits. People who didn't, 15%. And the links the summary itself cites? Clicked on 1% of visits.

One percent.

I didn't panic. I got curious. If almost nobody clicks, what exactly is the point of being cited? And if being cited still matters, who decides who gets cited?

Those two questions are why this series exists. They also sit right under a word you've seen a hundred times this year. GEO.

So, from the top.

GEO means Generative Engine Optimization. The phrase came out of a paper by Pranjal Aggarwal and colleagues at Princeton, IIT Delhi and some independent researchers. It went up on arXiv in November 2023 and was published at KDD 2024, one of the big data mining conferences. Their "generative engines" was an umbrella term, covering systems like Bing Chat, Google's SGE (which became AI Overviews) and Perplexity, things that go find sources and then write an answer with those sources stitched in as citations.

Their framing is simple. I still think it's the most useful one anyone has come up with.

A classic search engine hands you a ranked list. A generative engine hands you a paragraph. Some sources in that paragraph get a full sentence. Some get a footnote. Most get nothing at all. So visibility stops being "what position am I in" and turns into "how much of this answer is built on me, and how early do I show up in it." The paper goes as far as defining metrics, like a position-adjusted word count, which is roughly how many words of the answer are credited to your source, weighted so earlier mentions count for more.

That definition matters more than it looks. It quietly swaps out what you're optimizing for.

Being clicked first is no longer the goal.

Being used is.

I can already hear the SEO people. New acronym, somebody's selling a course, give it six months. Honestly, that eye-roll is half right, and I'd rather say so now than pretend otherwise. Google's own docs for site owners say the SEO best practices "remain relevant" for AI Overviews and AI Mode. There are "no additional requirements." No "special optimizations necessary." If your site is crawlable, indexed and written for humans, you've already done most of what any honest GEO guide would ask of you.

So if it's half right, why does the word deserve to exist?

Three things are genuinely different. Different enough that treating GEO as SEO with a new hat will lead you somewhere bad.

The first is that the unit of competition shrinks. On a results page you're competing as a page. Inside an AI answer you're competing as a passage, a claim, a number, a quote. Aggarwal's team tested nine ways of rewriting a source page. Adding quotations, adding statistics and citing sources lifted a source's visibility by roughly 30 to 40% on their main metric. Classic keyword stuffing did approximately nothing, and on Perplexity it came out about 10% worse than leaving the page untouched.

Part 4 will take that apart, because a 2025 benchmark called C-SEO Bench pushed back on a lot of it. For now keep the shape. The model reads your text, then decides whether it's worth borrowing. Keyword density plays no part in that decision.

The second is that you're not the only one describing yourself anymore.

Ask ChatGPT what's a good tool for X and the answer gets assembled from everyone who ever wrote about X. Your homepage is one voice in that room. Review sites are in there. Reddit threads, YouTube videos, news coverage, Wikipedia. Ahrefs studied 75,000 brands and found that how often a brand gets mentioned across the web correlates with showing up in Google AI Overviews at 0.664 (Spearman). Backlinks came in at 0.218.

Correlation, not cause, and Ahrefs is upfront about that. Still, it fits what you'd expect from a system that summarizes the web rather than ranking your page. Part 6.

The third is the one that hurts. It's the 1% I opened with. Clicks are getting thin. Ahrefs ran one 300,000-keyword comparison, then ran it again. In April 2025 they estimated an AI Overview on the page cut the top result's click-through rate by about 34.5%. On December 2025 data the estimate was 58%. Their own phrasing, Google now "keeps" 58 out of every 100 clicks a top-ranking page used to earn.

Which changes the job. Part of the value of sitting inside an AI answer is that the answer itself hauls your name, your number, your framing, out to someone who will never open your site. That's PR work, not traffic work, and it needs different measurement. Part 7.

A quick note on the naming mess, because it does confuse people. You'll run into AEO (answer engine optimization), LLMO, "AI SEO," "AI search optimization." When Bing's Webmaster Tools team shipped a new citation report in February 2026, they described it as "an early step toward Generative Engine Optimization (GEO) tooling." So GEO is what the platforms themselves seem to have picked. For now.

The label isn't what matters. What matters is whether the person using it means "the same SEO checklist, rebranded" or "a different measurement problem with some overlap." I mean the second one.

And if a vendor shows up pitching any of those acronyms, ask one thing. How do you measure it?

If the answer is a screenshot of a single chat, you've already learned everything that screenshot contains.

Okay. That's what GEO is. Now the other half, what it isn't, because that's where the hype hides.

GEO does not replace search. Not yet, anyway. And the numbers are messier than the conference stage makes them sound.

Gartner predicted in February 2024 that traditional search engine volume would drop 25% by 2026, with search marketing losing share to AI chatbots and other virtual agents. It's October 2026 as I write.

Did it happen?

I can't find clean evidence that it did. Datos, which runs a clickstream panel, went through 2023 data soon after the prediction and found no appreciable shift in total search volume. Google moved from 91.5% to 90.7% of traffic among the sites they compared. Old data. And in fairness to Gartner, a forecast isn't a measurement. But I'd be lying if I called the 25% drop a settled fact. It isn't. When someone quotes it as one, take it with salt.

Traffic from AI is still small, too. Ahrefs keeps a running benchmark from its analytics product. The newest version I could open is dated October 15, 2026, and covers 115,254 sites. In September 2026, AI assistants as a group accounted for 0.43% of those sites' traffic. Google alone accounted for 23.96%. ChatGPT, 0.34%.

Zero point four three percent.

So when someone says AI search is where all your traffic went, ask them which traffic. Clicks from Google are thinning, sure, and AI answers are part of the reason. But those clicks didn't land somewhere else in equal numbers. A lot of them simply stopped happening. Pew found 26% of searches with an AI summary ended the browsing session right there, against 16% without one.

Here's how I hold both facts at once. Attention is moving into AI answers faster than clicks are. Pew's June 2025 survey had 34% of U.S. adults using ChatGPT, about double 2023, and 58% among adults under 30. Alphabet's Q2 2025 earnings remarks put AI Overviews above 2 billion monthly users. People are reading the answers. They're just not walking out of them as often.

That's the situation GEO actually responds to. Not "search is dead." More like this. The conversation about you keeps happening inside a box you don't control, and the box only links out once in a while.

Which leaves the obvious question. If I can't control the box, what can I do? Here's where my bias shows, since this is my job at Otlume.

The most underrated part of GEO is dull. Make sure one place on the web has the facts about you correct, current, plainly written and readable by machines. Your own domain. Not because a model will trust your homepage. It won't, and part 6 covers how it usually trusts other people's version of you more. It's because when a system finally goes looking for the canonical answer to "what does this company do" or "how much does it cost," you want an answer to exist, and you want it to be true.

That's Otlume's whole premise, giving a person or an organization an official site built for people and agents at the same time, with structured data, a sitemap, robots.txt and llms.txt generated for them. What I like is that the product pages don't oversell it. Ask the FAQ whether AI will recommend you and it says, "No one can guarantee that." I'll hold this series to that.

Let me leave you with something you can do this week, because I don't want part one to be all framing.

Open three AI tools you or your clients actually use. Reading in English, that's probably Google with AI Overviews or AI Mode, ChatGPT, and Perplexity or Copilot. In China, use 豆包, DeepSeek or Kimi instead.

Ask each one two kinds of questions. One with your name in it, like "What is [your company] and what does it charge?" One without, the kind a stranger would type, like "What are good options for [the thing you sell] for [the kind of customer you serve]?"

Write the answers down. One screenshot is not a result. Ask each question at least five times, in fresh chats. Part 3 explains why. The short version is a 2026 SparkToro study, which found the odds of ChatGPT or Google's AI handing you the same brand list twice, across about a hundred runs, at under 1 in 100.

I ran a tiny version today on Otlume itself, using an AI search assistant that writes cited answers. I asked "What is Otlume?" and got an accurate one-line description, with all five cited sources being Otlume's own pages. Then I asked the category question, the best platform for an agent-ready personal website with llms.txt and JSON-LD. It recommended three other products. Otlume never came up.

That's the gap in miniature.

Your own site can answer questions about you. It can't, by itself, get you into the answers about your category. And Otlume, brand new, with an empty knowledge base and an empty news page, is an honest picture of where most of us start.

Which I find oddly comforting. The problem is legible. You can see exactly where you're standing.

I keep coming back to an idea from the early PageRank days. Brin and Page saw that the web already held a vote on which pages mattered, written in the links, and all you had to do was read it. Generative engines are doing a version of that with a far richer signal. They read what the web says, not only how it links. Which means the old SEO question, how do I rank, slowly becomes a much older question, the one PR people have asked forever. What do people say about you when you're not in the room?

GEO is that question, aimed at a machine that reads everything.

Over the next nine parts I'll take this from framing to an actual program. Part 2 cracks the black box open a little, how engines choose sources. Part 3 is the honest audit. Parts 4 through 6 cover what you can change, the content itself, the machine-readable layer, and what other people say about you. Parts 7 through 10 are measurement, platform differences, experiments and attribution, and turning all of it into a habit instead of a yearly scare.

I'll try to keep every claim linked and every number double-checked. Where the research fights with itself, and it fights a lot, you get both sides, not whichever one is convenient.

For now, go run those prompts. Five times each. Write them down.

That's your baseline. Everything later builds on it.