GEO SEO in 2026: What Actually Works to Get Cited by AI

GEO SEO in 2026: how ChatGPT and AI Overviews choose their sources, what gets set on the page, and how to measure your citations.

TL;DR

GEO is optimization for responses generated by ChatGPT, Perplexity, Gemini, and Google's AI Overviews. Bad news for methodology sellers: 80% of the work is straight SEO, and Google states in black and white that no special optimization is needed. What actually changes comes down to three things: the link between organic ranking and citation has collapsed, AI robots don't render JavaScript, and half your visibility happens off-site. The llms.txt file is useless: I have it live on my site, and nobody reads it.

What is GEO, exactly, and how does it differ from SEO?

GEO stands for optimization to appear in an AI-generated response, while SEO aims for a position in a list of blue links.

The fundamental difference isn't technical, it's structural. In SEO, you fight for a rank. In GEO, you fight to be retained as a source by a model that reformulates, cross-references three or four pages, and only cites what serves its sentence construction. You don't win a position, you win a mention.

Direct consequence: the metric changes. Average rank no longer means much. What counts is the share of responses where your brand appears across a given set of questions, and the share of those responses that cite you with a link.

Second consequence, less discussed: you lose control of context. In a ChatGPT response, your page lands next to three competitors in a paragraph you didn't write. You're no longer optimizing a page, you're optimizing the probability that a model finds you useful.

Third consequence, the most uncomfortable: traffic doesn't always follow citation. Being named in a response without a clicked link exists, and it shows up in none of your reports. You have to accept working toward visibility that's partially unmeasurable, which runs counter to fifteen years of SEO habit.

How do AIs actually choose their sources?

Models don't choose in a vacuum: they run searches, pull a few pages, and cite the ones where a passage directly answers the question asked.

The mechanism is roughly the same everywhere. The model rephrase your question into several searches, queries an index (Google's, Bing's, or its own), retrieves a small number of documents, and writes. What's at stake here is retrieval, not ranking. A page ranking 14th on a reformulated query can easily end up cited, while the #1 result on the original query gets ignored.

Data confirms this shift. Ahrefs analyzed 863,000 SERPs and 4 million URLs cited in AI Overviews: only 38% of citations come from the top 10, versus 76% a year earlier. The rest splits between positions 11–100 and pages absent from the top 100. BrightEdge, on its own dataset, goes even lower.

I take two things from these numbers. Being first is no longer a guarantee of citation. And not being first is no longer grounds for exclusion, which opens a door to mid-tier sites on specific questions.

What gets set on the page, concretely

On the page, four settings have observable effect: the answer phrase, heading structure, FAQ, and structured data.

The answer phrase under each H2. This is the highest-ROI setting, and the simplest. Each section heading poses an implicit question, the first sentence answers it in one complete, standalone sentence, fifteen to twenty-five words. A model that pulls your page finds a citable unit without having to stitch three paragraphs together. I've applied this rule to all my articles for a year, and it's the only writing change where I see a net effect on how my phrasing gets reused.

Headings that repeat the question, not the concept. "How much does an SEO consultant cost" beats "Pricing." Queries reformulated by models sound like spoken questions, your headings should sound like that too.

A real FAQ at article end. Five to eight questions, answers of two to four sentences each, each one standalone. It's the format closest to what a model tries to produce.

Structured data. The Article, FAQPage, and Organization markup won't get you cited by magic, Google is clear about that. It helps resolve ambiguities about author, date, and the entity behind the site. It's hygiene, not a shortcut.

Google states it in its own documentation on AI features and your site: no additional requirements, no special files, no dedicated markup. Existing controls keep working, nosnippet and max-snippet included, if you want to limit your content reuse. I cite this page every time someone sells me a proprietary GEO method.

Does llms.txt do anything? No

The llms.txt file is read by no major AI search engine in production, and large-scale studies find no link between its presence and citations.

I deployed it on my own site across three clean routes, served and cached correctly. I've never observed a request from an AI robot on it, and no trace of effect on citations. This isn't an isolated feeling: Ahrefs reviewed 137,000 sites and concluded that 97% of llms.txt files are never read.

On Google's side, the position is public and consistent: Gary Illyes indicated Google doesn't support it, and John Mueller compared it to the meta keywords tag, which isn't a compliment. The only use that holds up today is technical documentation consumed by code agents.

My advice: if you already have it, leave it, it costs nothing. If you don't, don't spend a development day building it. Put that day into your server response time, the return will be better.

What doesn't get set on the page

Half your visibility in AI responses depends on what other sites say about you, and you can't edit it.

Models lean heavily on pages that aggregate: comparisons, forums, Reddit, directories, "best tools for X" lists. If your brand doesn't appear on any of these pages, no on-page setting will surface it. You could write the best product page on the market, the model will cite the comparison that left you out.

This work looks more like PR than SEO: getting listed in relevant roundups, showing up where your category gets discussed, publishing data that others will reuse. It's slow, it's hard to automate, and it's the only real defensible moat. I described the logic on the tooling side in my SEO consultant tool stack.

The other thing that doesn't get set is perceived freshness. Models favor what looks current, and an article dated two years back on a market that moves every quarter drops out of play without warning. Updating existing content often pays more than publishing something new.

The technical trap nobody watches: rendering

The main AI robots fetch HTML but don't execute JavaScript, which makes any client-side injected content invisible.

Vercel and Merj measured the actual traffic of these robots: GPTBot, ClaudeBot, and PerplexityBot download JavaScript files without executing them. Only Gemini benefits from Google's rendering infrastructure. The full study is public. If your main content arrives after hydration, it doesn't exist for these robots.

I hit an even trickier case on my own site. My CDN cache key ignored the RSC header used by Next.js for preloads. Result: a preload request cached a data fragment for an entire URL, then served it to everyone, robots included, for days. The HTML was perfect locally, the served page wasn't.

The lesson holds beyond my case: always verify what actually comes out of the server, not what your local render produces. A curl on the public URL, no cookies, with the robot's user-agent, answers the question in ten seconds. And watch your 5XX errors: five days of errors served to Googlebot cost me most of my impressions, and no content optimization could fix it.

How to measure AI visibility without fooling yourself

No tool gives a complete picture today, because AI platforms don't expose impressions or positions.

Google Search Console won't help you isolate AI Overviews: Google aggregates these clicks into the standard performance report, no distinction made. You can follow the overall trend, not the AI share.

For citation tracking, the category structured itself in a year. Semrush offers an AI visibility module as a supplement to its subscription, tracking brand mentions in ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews. SE Ranking offers the equivalent at a more accessible price. Search Atlas and Frase attack the topic through content rather than tracking. Specialized players exist too, pricier and more refined. I compared them to classic tooling in my best SEO tools comparison, and the match between the two historical suites is detailed in Semrush vs Ahrefs.

Common trap: all these tools measure a sample of questions you define. Change the list, change the score. Treat these numbers as a trend indicator on a stable list, never as market share.

The free measurement that still works: ask yourself your twenty key questions to ChatGPT, Perplexity, Claude, and Gemini, once a month, and note who gets cited. It's manual, it's a bit long, and it teaches you more about your market than any dashboard.

Where to start if your site is already optimized

If your technical SEO is clean, four projects in this order deliver the best return for time spent.

1. Verify your main content is in served HTML, no JavaScript. It's binary, and it's blocking.
2. Add an answer phrase under each H2 on your twenty most strategic pages. Budget half a day.
3. List the ten third-party pages that answer your target questions, and get yourself referenced on them. This is the long project, start it now.
4. Set your list of twenty tracking questions and measure monthly, with a tool or by hand.

What I wouldn't do: rewrite entire articles "for AI," buy a GEO tool before you have a stable question list, produce llms.txt. Definitely not publish more. On my own site, I decided the opposite this year: fewer articles, thicker ones, better maintained.

Key takeaways

• GEO isn't a separate discipline: Google itself writes that no special optimization is needed for its AI features.
• The link between top 10 and citation collapsed, from 76% to 38% in a year per Ahrefs. Being first no longer suffices, not being first no longer excludes.
• The fifteen to twenty-five word answer phrase under each H2 is the highest-ROI on-page setting.
• ChatGPT, Claude, and Perplexity robots don't execute JavaScript: all client-side injected content is invisible to them.
• llms.txt does nothing today, 97% of files never get read.