# Forget SEO, the next fight is AEO and GEO

> Ranking first stopped being the prize the moment the answer stopped linking anywhere. The new game is whether a model names you, and it is already being gamed with thirteen words on Reddit.

- Author: Abhay Bagda (https://abhaybagda.com/about/)
- Published: 2026-09-16
- Topic: Technical
- Reading time: 8 min
- Canonical URL: https://abhaybagda.com/writing/forget-seo-aeo-geo/

---

In May I wrote that you should stop optimising for AI and go write something worth citing. I still think that was right. I also shipped an `llms.txt` file to this site on Sunday, which makes me look exactly like the consultants I spent a thousand words making fun of. Nothing in that piece has been proven wrong since. More information has just turned up, which is how the next few years are going to go, and a position you took four months ago isn't something you owe any loyalty to.

Here's what actually turned up. Google's Search team published their optimisation guide on 15 May and named `llms.txt` in the mythbusting section, as a tactic that does nothing for AI Overviews or AI Mode. That's still their position today. Eight days before that, the Chrome team shipped Lighthouse 13.3.0 with an `llms.txt` audit inside it, under a brand new category called Agentic Browsing, promoted straight out of experimental and into the default config. It checks that the file is reachable at your root, that it has a heading, that it has links. One team is answering whether the file helps you rank, the other whether an agent that has already landed on your site can find its way around it.

John Mueller, who has spent a year swatting this question away, moved too. On 2 June he said "I don't think anyone knows, it's purely speculative for now (the file has existed for years, yet none of the AI systems use it)". The more useful half of his answer was the next line. "When an AI platform that brings you clients complains that it needs the file for your site, then I'd recommend taking the time to create one." That isn't an endorsement and he'd hate me framing it as one. It's still a long way from Gary Illyes standing up at Search Central Live in July 2025 and saying Google doesn't support it and isn't planning to.

I have had the same conversation with three different friends this year, two of them running small businesses and one freelancing. Someone pulls up Search Console, points at a chart that has been flat or drifting down since early 2025, and asks what they're doing wrong. Usually nothing. Their rankings are fine and their impressions are fine. The clicks stopped arriving because the answer stopped needing them.

## Google already told you where this is going

The numbers aren't subtle. AI Mode, Google's fully generated search surface, passed a billion monthly active users this year, and Sundar Pichai called it "our biggest upgrade to Search ever" at I/O in June. When he was asked on the Hard Fork podcast whether he was comfortable with AI Mode replacing classic Search, he didn't defend the ten blue links. He said "it is very clear as we evolve through these changes, people are responding positively" and described the shift as "a continuum", which is a polite word for a road that only goes one way.

The measurement here is genuinely messy, and anyone quoting you a single clean percentage is selling something. Conductor's Q1 2026 pass over 21.9 million searches put AI Overviews on about 25% of queries. BrightEdge, measuring differently, had it closer to 48% by February. Semrush had it peaking last July and then falling. Pick your methodology and you get a different headline. What isn't in dispute is what happens after the answer renders. Around 60% of traditional searches already end without a click, that climbs to roughly 83% when an AI Overview sits on top, and on AI Mode queries it's about 93%.

So your page can rank first, be perfectly indexed, and still be quietly read aloud to your customer by something that never sends them to you. The one piece of good news in Seer Interactive's 2026 data is that brands cited inside an AI Overview pull roughly 120% more organic clicks per impression than brands that aren't. Being cited is worth real money, which is the same thing as saying that being merely rankable is worth less every quarter.

> The question stopped being whether you rank for the query. It is whether you are inside the answer the person actually reads.

## AEO and GEO are not a file you upload

Here's the distinction most agency decks blur on purpose, because blurring it is how you bill for it.

**SEO is about the index.** Can a crawler reach the page, parse it, and file it under something. This hasn't changed and it's still the entry requirement. A page that can't be indexed or shown with a snippet can't be quoted by an AI surface either, which is exactly what Google's own optimisation guide said in May and still says.

**AEO is about the page being the source.** Someone asks a question, a model assembles an answer, and your paragraph is the one it paraphrases and links. You win that by having a page that answers one specific question cleanly, with the claim stated near the top instead of buried under six hundred words of preamble. That isn't a magic tag doing the work, it's that retrieval splits your page into chunks and scores them separately, so a chunk containing a whole answer beats a chunk containing half of one.

**GEO is about being the entity.** Nobody types "best project management tool site:yourdomain.com". They ask "who should I hire to fix my Snowflake bill" and the model comes back with three names. Those names are assembled from everything the model has read about you across the entire web, weighted heavily towards the places it actually retrieves from at answer time. Your own website is one input among all of those, and usually not the biggest one.

The only lever on your own site that has really changed shape is making your claims machine-legible. Here that meant adding two optional fields to the content schema, so a piece can declare what it's about and which questions it answers, and the layout emits both as structured data.

```ts
// src/content.config.ts
faq: z.array(z.object({ q: z.string(), a: z.string() })).optional(),

mentions: z
  .array(z.object({
    name: z.string(),
    url: z.string().url(),          // stable ID: Wikipedia, official site
    type: z.enum(['Thing', 'Organization', 'SoftwareApplication', 'Person', 'Product'])
      .default('Thing'),
  }))
  .optional(),
```

Google killed FAQ rich results for everyone except government and health sites back in 2023, so this buys me no snippet at all. It's there for retrieval, not for Google. The `mentions` array exists so that a system reading this page can tell "Claude" here is a model and not somebody's uncle. That's the entire technical surface area of AEO on a small site, an afternoon of work with no recurring cost attached.

## Anyone can buy their way into the answer for thirteen words

Now the part that should worry you more than your traffic chart does.

In June, three Cornell Tech researchers published a preprint called "Deep-Research Agents Can Be Poisoned via User-Generated Content". Tingwei Zhang, Harold Triedman and Vitaly Shmatikov built an attack they named WARP, and the finding is that you can steer a deep research agent's recommendation by planting a short promotional sentence inside a Reddit comment. Not a thousand-word astroturfed post, just thirteen words appended to an existing comment on a page the agent was already going to retrieve anyway.

"We show that a tiny snippet, just 13 words, of retrieved text on a UGC website like Reddit, Wikipedia, Quora, Facebook, etc. can change AI agents to output spam / scam content pretty consistently," Triedman told 404 Media. He was blunter about why it works. "It really is just that simple. The way that you can attack these systems is usually so much dumber than you think it is."

Zhang, his co-author, named the underlying failure plainly. "It's not thinking about which source you find more credible: a random Reddit comment or an article from a government website. They are treated almost the same by the LLMs." Across 176 queries in the study, once the poisoned page was retrieved, the fabricated business was cited every single time. Reddit alone accounts for somewhere between 54% and 71% of the user-generated content these systems pull in, and roughly a quarter of all their citations come from UGC sites.

So the thing you've seen demonstrated on Twitter, where someone seeds enough posts saying they're the best consultant for X and the model eventually starts telling strangers exactly that, is real, documented, and cheaper than one month of the SEO retainer you're already paying. Microsoft's security team catalogued 31 companies across 14 industries running a related trick in the wild by February. I'm not telling you to do this. I'm telling you your competitor doesn't need your permission, and Zhang's own read on the defence is "really, there's no easy fix."

## What I changed on this site, and what I did not

Four things, in the order I'd do them again.

**I fixed the entity before the pages.** The about page, the homepage description and the JSON-LD `Person` block now say the same sentence about who I am, in the same words, pointing at the same profile. Retrieval systems resolve entities by agreement across sources. Three slightly different bios is three weak signals where you could have had one strong one.

**I added the `llms.txt` route** at [src/pages/llms.txt.ts](src/pages/llms.txt.ts), generated from the content collection so it can't drift out of date. Google Search still ignores it while Chrome's own Lighthouse now audits for it, and Anthropic reads it. Mueller's test is the right one, build it when a platform that actually sends you traffic asks for it, and the only thing I'd add is that forty lines of TypeScript with no maintenance cost is cheap enough to do before anyone asks. That's a very different economic proposition from the $4,000 a month agencies are charging for the same file.

**I stated the negative claims out loud.** The `llms.txt` says in plain text that this is a personal writing archive and that no services are offered here. I closed Nodeflux, I'm employed full time, and I'd rather a model read that off my own site than infer something stale from a three-year-old directory listing. If you don't state what you are not, something else will guess on your behalf.

**I didn't chunk-rewrite a single article.** No 80-word answer blocks, no reformatting my openings into Q&A, no GEO audit. The pages that get quoted are still the ones carrying a number or a scene nobody else has, and that incentive didn't change, it just got sharper. A model pulling from four articles that all say the same thing will paraphrase the lot and cite none of them.

## Go be the entity, not the page

Start with the boring half hour. Search your own name and your company name in ChatGPT, Gemini and Perplexity, and write down what comes back. Most people have never done this and are genuinely surprised by what the model believes about them. Then trace where those answers came from, and you'll usually find it isn't your website at all, it's a directory listing, an old interview, or a Reddit thread you've never read.

That's your real surface. Go fix the wrong bits at the source, get your one-sentence description identical everywhere it appears, and put one page on your own site that answers the exact question a buyer would type, with the answer in the first paragraph where a retriever will actually find it. Then go write the piece only you could have written, because everything above this sentence is table stakes and that last part is the only bit your competitor can't copy in an afternoon.

There's a Darwin line people reach for at this point, about how it isn't the strongest or the most intelligent that survives but the one most adaptable to change. Darwin never wrote it. It comes from Leon Megginson, a management professor at LSU, paraphrasing loosely in 1963, and it has been laundered through slide decks ever since. Which is a fairly on-the-nose demonstration of what happens when a claim gets repeated across enough sources that everybody, humans very much included, stops checking who actually said it.

If you want the research rather than my summary of it, [404 Media's writeup of the Cornell paper is here](https://www.404media.co/it-is-trivially-easy-to-use-reddit-to-manipulate-ai-search-research-suggests/), and Google's own [AI optimisation guide is here](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide). Read them in that order. The second one is a lot less alarming after the first.