AEO, GEO, SEO: what the terms actually mean

Three acronyms, one confused market. Ask five agencies what AEO means and you’ll get five answers, at least two of which are the same answer with a different letter in front of it.

The confusion is not harmless. Businesses are buying services whose names they can’t define, from vendors who can’t define them either, and the resulting scopes of work are impossible to evaluate. So this is the plain version: where each term came from, what each one actually describes, where they genuinely overlap, and the specific places the distinction stops being semantic and starts affecting what you do on Monday.

The short version

SEO — search engine optimization. The practice of getting a page to appear in a ranked list of results a person then chooses from. The output is a link. The unit of success is a click.

AEO — answer engine optimization. The practice of getting your information into a direct answer, where the system responds instead of listing. The output is a statement. The unit of success is being the source that statement is built from — with or without a click.

GEO — generative engine optimization. The narrower, more technically precise case of AEO where the answer is composed by a large language model that retrieves documents and writes new prose grounded in them. The output is generated text. The unit of success is inclusion and attribution inside that text.

The cleanest way to hold all three: SEO is about ranking, AEO is about answering, GEO is about generating. AEO and GEO describe overlapping territory, and in ordinary use they’re interchangeable. SEO is the older and broader discipline that both of the others grew out of and still depend on.

Where the words came from

GEO has a real, datable origin. A paper titled “GEO: Generative Engine Optimization,” by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande, was first posted to arXiv on November 16, 2023 and published at KDD 2024.[] It defined a generative engine as a system that retrieves relevant documents from a corpus like the web and uses large neural models to generate a response grounded in those sources, with attribution.[] The authors built a benchmark of 10,000 queries and tested content-side interventions, reporting up to a 40% relative improvement in visibility on their position-adjusted word count metric, with quotation addition, statistics addition, and citing sources among the top-performing methods.[]

That’s a specific claim from a specific experiment, not a law of nature. But it matters that the term has a paper behind it, because it means GEO refers to something someone actually defined.

AEO doesn’t have an equivalent founding document. It circulated through trade and vendor writing as shorthand for optimizing toward direct answers — the featured-snippet and voice-assistant era first, then the chatbot era. We’re not going to name a coiner, because we couldn’t verify one. Anyone who tells you confidently who invented “answer engine optimization” is guessing.

What’s more interesting is that the terminology has started appearing in platform documentation rather than just marketing copy. In February 2026, Microsoft shipped an AI Performance report in Bing Webmaster Tools and described it as “an early step toward Generative Engine Optimization (GEO) tooling in Bing Webmaster Tools, helping publishers understand how their content participates in AI-driven experiences.”[] A search platform putting the acronym in its own product announcement is a meaningful signal about which word is winning.

Google’s position, and what it leaves out

Google’s public line has been consistent and dismissive of the whole vocabulary. Search Liaison Danny Sullivan has said “Good SEO is good GEO, or AEO, AI SEO, LLM SEO, or LMNOPEO.”[] Nick Fox, Google’s SVP of Knowledge and Information, has said the way to do well in Google’s AI experiences is “very similar, I would say, the same as how to perform well in traditional search.”[] At a Search Central event in Asia Pacific in July 2025, Gary Illyes reportedly told attendees that to appear in AI Overviews you should “simply use normal SEO practices,” and separately that Google doesn’t support llms.txt files and isn’t planning to.[]

Take that seriously. It’s the platform telling you the fundamentals haven’t been replaced, and the evidence broadly supports it. Anyone selling you a parallel technical stack that bypasses ordinary indexing is selling you something Google says it doesn’t read.

But notice the scope of the claim. Google is describing how to perform well in Google’s AI experiences, which are grounded in Google’s index and reachable by Googlebot. It is not a claim about ChatGPT, Perplexity, or Copilot, and it is not a claim that the measurement, the failure modes, or the competitive dynamics are the same. Perplexity’s head of communications, Jesse Dwyer, has publicly cautioned against mapping SEO understanding one-to-one onto answer engines — arguing the biggest mistake is treating them as identical.[]

Two things can be true. The inputs rhyme. The systems don’t.

What actually changes: from ranking pages to grounding answers

Microsoft has written the clearest public description of the underlying shift. In a May 2026 post, engineers on the Microsoft AI team framed traditional search as answering “which pages should a user visit?” and grounding for AI as answering “What information can an AI system responsibly use to construct an answer?” Their summary line: “Search indexing was built to help humans decide what to read. Grounding indexing is being built to help AI systems decide what to say.”[]

That is the distinction, stated by a platform, in one sentence.

The mechanical consequence is that answer engines usually don’t run your query. They run several of their own. Google’s VP of Product for Search, Robby Stein, has described AI Mode’s query fan-out this way: for something like things to do in Nashville with a group, the system “may think of a bunch of questions like great restaurants, great bars, things to do if you have kids, and it’ll start Googling basically.” Its Deep Search mode, he said, can issue “dozens or even hundreds of background queries.”[]

So the target moves. In classic SEO you optimize a page against a query. In answer optimization you’re trying to be the source that shows up repeatedly across a spray of sub-queries you never saw, on a topic you can only infer. Keyword-level thinking doesn’t map onto that cleanly. Entity- and claim-level thinking does.

The overlap is real, and it’s moving

Here’s where the honest version diverges from both the “it’s all the same” camp and the “everything changed” camp.

Semrush’s July 2025 AI Mode study analyzed 5,000 keywords and more than 150,000 citations across Google Search, AI Overviews, AI Mode, ChatGPT, and Perplexity. Overlap with Google’s organic top 10 varied enormously by engine: Perplexity showed over 91% domain and 82% URL overlap, AI Overviews roughly 86% domain and 67% URL, while AI Mode sat near 54% domain and 35% URL — and ChatGPT had the weakest alignment of the group.[]

Read that carefully. For some engines, ranking well in Google is very nearly the whole job. For others, it’s about a third of the job.

The Google-specific picture has also shifted, and the shift is contested. In July 2025, Ahrefs analyzed 1.9 million citations from 1 million AI Overviews and found 76.1% of cited pages ranked in Google’s top 10, with 86% appearing somewhere in the top 100.[] An updated Ahrefs analysis reported in March 2026, covering 863,000 keywords and 4 million AI Overview URLs, put the top-10 share at 38%, with 31.2% ranking 11–100 and 31.0% outside the top 100.[] Ahrefs itself flagged that improved parsing methodology since the first study makes a direct year-over-year comparison unsafe.[]

We’d rather report that caveat than the headline. The trend direction is plausible; the precise magnitude isn’t settled.

What is clearly not the same across engines is which sources get pulled. Profound’s analysis of 680 million citations tracked from August 2024 to June 2025 found Wikipedia accounting for 7.8% of ChatGPT’s citations, Reddit leading Google AI Overviews at 2.2%, and Reddit far more concentrated in Perplexity at 6.6%.[] Semrush’s 2026 AI Visibility Index, built on 126 million U.S. AI search prompts from January through April 2026, found ChatGPT citing an average of about 15 sources per response against roughly 3 for Gemini — and only 36 global brands holding top-100 visibility across all four platforms studied in every month.[]

One page. Four engines. Four different retrieval appetites, four different source biases, four different results. That is not a rebranding of SEO. It’s a portfolio problem.

Where the distinction becomes operational

Three places, specifically.

Access control is now a separate decision. OpenAI documents distinct crawlers with distinct jobs: OAI-SearchBot surfaces sites in ChatGPT’s search features, GPTBot collects content for model training, ChatGPT-User handles live user-triggered fetches, and OAI-AdsBot handles ad-related checks.[] Blocking the training crawler and blocking the search crawler are different choices with different consequences — OpenAI states that sites blocking OAI-SearchBot won’t appear in its search answers.[] Nothing in classic SEO required you to hold a position on that.

Measurement is a different instrument. Google confirmed in June 2025 that AI Mode clicks, impressions, and position data counted toward Search Console performance totals, folded into combined web search metrics; it wasn’t until June 3, 2026 that Search Console got a dedicated generative AI view — and even that groups AI Overviews and AI Mode together rather than separating them.[] Bing’s AI Performance report counts citations and “grounding queries,” and Microsoft explicitly warns that citation frequency “does not indicate ranking, authority, or the role of any page within an individual answer.”[] There is no rank-tracking equivalent for an answer. Pretending otherwise is how reporting becomes fiction.

The payoff currency changed. Pew Research analyzed 68,879 Google searches from 900 U.S. adults in March 2025 and found that when an AI summary appeared, 8% of users clicked a traditional result, compared with 15% when no summary appeared; only 1% clicked a link inside the summary itself.[] If a meaningful share of your visibility now resolves without a click, then a program measured only in sessions will report failure during periods when it’s working, and nobody will be able to tell the difference.

The scale check

Now the counterweight, because the honest version needs one.

SparkToro’s March 2026 research, using Datos clickstream data across millions of desktop devices for calendar year 2025, found that in Q4 2025 Google accounted for 73.7% of U.S. desktop searches across the 41 sites analyzed, while all AI tools combined accounted for 3.2% — with Amazon, Bing, and YouTube each receiving more search activity than ChatGPT.[] In the EU and UK, Google’s share was closer to 80%.[]

Anyone telling you search is over is not reading the data. Anyone telling you nothing has changed isn’t either: Similarweb reported AI Overviews holding a 43%-plus appearance rate on U.S. searches, and a substantial reshuffling of the assistant market between June 2025 and May 2026, with ChatGPT’s share of worldwide AI web traffic falling from roughly 76% to around 53% while Gemini rose from under 9% to roughly 27–28%.[]

The right conclusion is unglamorous. Answer surfaces are a real and growing minority of discovery, they behave differently from ranked lists, and their internal league table is unstable enough that betting on one engine is a bad idea.

How we use the terms

We use SEO for the work of being indexed, understood, and ranked. We use AEO as the umbrella for being usable inside a direct answer. We use GEO when we specifically mean LLM-generated responses, because that’s what the original paper meant and precision is worth something.

We don’t sell them as three products. The underlying work — crawlable pages, unambiguous entities, verifiable claims, structured data, real authority — is largely shared, exactly as Google says. What isn’t shared is the targeting, the instrumentation, and the reporting, and those are the parts that break when an agency treats a new surface as an old one with a new name.

The acronym you use matters less than whether you can say, in one sentence, what you’re optimizing toward and how you’d know it worked. Most people selling AEO right now cannot.