# What Is GEO? How We Make Our Site Citable by AI Search

> GEO makes content easier for AI search to cite. What the research and Google say, plus what we shipped on our own site and how it scored.

- URL: https://noise-and-signal.com/en/insights/what-is-geo-generative-engine-optimization
- Author: 翁睿承 (諾訊科技 Noise & Signal)
- Published: 2026-09-27
- Tags: Website Development
- Language: en

---
GEO — generative engine optimization — is the practice of making your content easier for generative search engines like ChatGPT, Google AI Overviews and Perplexity to understand and cite. The term comes from a 2024 research paper, and Google's own position is that optimizing for AI search is still SEO, with no special files or markup required. In September 2026 we used our own website as a test bed and shipped everything we could. The result: 86 for technical foundations, 5 for brand authority. Here's the full story, including the limits.

## Where the term comes from

GEO was introduced in "[GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735)" by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD '24 (the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining) in Barcelona in August 2024. Three points matter:

- **Definition.** GEO is a black-box optimization framework that helps content creators improve how visible their content is inside generative engine responses.
- **Headline number.** The abstract states that GEO "can boost visibility by up to 40%" in generative engine responses.
- **The caveat people skip.** The authors also found that the effectiveness of these strategies varies across domains, so optimization has to be domain-specific.

So "up to 40%" is a benchmark result on the paper's own GEO-bench, not a promise any site gets by applying a checklist. (Checked September 2026.)

## What Google says

Google's [AI features documentation](https://developers.google.com/search/docs/appearance/ai-features) is blunt: you don't need to create new machine-readable files, AI text files or markup to appear in AI Overviews or AI Mode, and there are no technical requirements beyond normal indexing eligibility. A [companion guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) says that from Google Search's perspective, "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." (Checked September 2026.)

What about llms.txt? Its [own spec site](https://llmstxt.org/) calls it a proposal: a Markdown file at the site root that gives language models background information and links. It isn't a W3C or IETF standard, although the site notes that thousands of sites already publish one. (Checked September 2026.)

Our working definition: **GEO doesn't replace SEO — it's SEO plus three extra concerns.** Can AI crawlers fetch your pages? Can passages be quoted on their own? Is your brand mentioned anywhere off your own site?

| Area | Classic SEO focus | What GEO adds |
|---|---|---|
| Crawling | Can Googlebot crawl and index it? | Are GPTBot, ClaudeBot, PerplexityBot and other AI crawlers allowed? |
| Content | Keywords, headings, internal links | Does each page answer the question up front? Can paragraphs stand alone? |
| Structured data | Rich results | Makes it unambiguous who wrote it and which company it represents |
| Off-site | Backlinks | Mentions on LinkedIn, Wikidata and third-party platforms |

## What we shipped on our own site (September 2026)

Our SEO/GEO changelog for September 1–22, 2026 covers 64 commits. Here's what went live:

| Item | What we did | Why |
|---|---|---|
| robots.txt | 16 rules explicitly allowing GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended, Bingbot and others; /admin and /api/ disallowed | Tells AI crawlers exactly what they may fetch |
| Sitemap | Generated dynamically, covering routes like /faq, /estimate and /glossary, with hreflang alternates | One place for search engines to find every page in both languages |
| llms.txt and llms-full.txt | llms.txt lists services, cases and articles by section; llms-full.txt carries the full text | A site guide for AI agents |
| Markdown versions of articles | Requests with `Accept: text/markdown` receive the article as Markdown | Plain text is cheaper for agents to read than HTML |
| RSS | One RSS 2.0 feed each for Chinese and English, with author fields | Makes content easy to subscribe to and aggregate |
| hreflang and canonical | Every page goes through one metadata helper | Avoids a Next.js trap where a child page's openGraph object silently replaces the parent's |
| IndexNow | A script submits every sitemap URL in batches | Bing supports IndexNow; Google doesn't, so the sitemap still matters |
| Structured data | Site-wide Organization (tax ID, address, sameAs) and WebSite; Article, FAQPage, HowTo and Person on articles; DefinedTerm on glossary pages; WebApplication on /estimate | Helps AI identify the company and the author |
| New pages | /faq, /glossary, /process, /about/weng-rui-cheng, /estimate | Fills in the content people ask about most |
| Performance | Deferred a three.js background animation and fixed a fade-in on article text | See below |

The performance fix is worth a closer look. In the September 16 audit (Lighthouse, mobile simulation), LCP was 5.9 seconds on the homepage and 5.3 seconds on articles. The culprit wasn't images. A three.js background mounted on every page was eating 650–1,100 ms of main-thread time, and a fade-in animation meant the server-rendered article text started fully transparent until JavaScript hydrated. After fixing both, re-tests on September 22 and 25 showed LCP of 2.5 seconds on the homepage and 2.3 seconds on articles.

## The results: 86 for tech, 5 for brand authority

On September 16, 2026, we ran a full GEO audit on our own site. Overall score: 56/100.

| Category | Weight | Score |
|---|---|---|
| AI citability and visibility | 25% | 85 |
| Brand authority signals | 20% | 5 |
| Content quality and E-E-A-T | 20% | 49 |
| Technical foundations | 15% | 86 |
| Structured data | 10% | 50 |
| Platform optimization | 10% | 58 |

Platform readiness scores: Bing Copilot 70, Google AI Overviews 66, ChatGPT Search 56, Perplexity 52, Gemini 44.

The engineering was mostly right: all 42 audited pages returned 200, canonicals were correct, each page had a single H1, and full text was server-rendered. **The biggest deduction wasn't code — it was that nobody else was talking about us.** At audit time, LinkedIn, Wikidata, PTT, GitHub and third-party reviews returned nothing for "Noise & Signal" or 諾訊科技, and our sameAs pointed only to Google Maps.

Early search data told the same new-site story. Google Search Console for September 14–19, 2026:

- 9 clicks, 68 impressions, 13.2% average CTR, average position 9.1.
- 30 pages indexed, 21 not indexed; the sitemap listed 114 URLs, but Google knew about only 51.
- Two queries that looked like brand searches actually matched a different company with a similar name (諾歐科技), so they don't count as brand traction.

Bing Webmaster Tools showed 0 clicks and 0 impressions for the same period (still processing a new site), and we had submitted 406 URLs through IndexNow.

## Honest limitations

- **We don't have AI citation data yet.** As of September 27, 2026, we have no publishable numbers on how often ChatGPT or Google AI cites us. This article is about making a site *citable*, not about having been cited at scale.
- **"Up to 40%" doesn't transfer automatically.** It's a benchmark result, and the paper itself says effects vary by domain.
- **llms.txt isn't a Google requirement.** We publish it for other AI agents, not for Google rankings.
- **The samples are tiny.** A few days of early data shows direction, not conclusions.

## Off-site presence: Wikidata, Google Business Profile, LinkedIn

The audit shifted our focus from "tweak more code" to "make the outside world aware we exist":

- **Google Business Profile:** verified, with hours Monday to Friday, 08:00–18:00.
- **LinkedIn company page:** created and added to our structured data, so Organization sameAs now points to Google Maps, LinkedIn and Wikidata.
- **Wikidata:** entry Q141559765 created on September 26, 2026, with statements sourced to Taiwan's official company registry, including our tax ID (60503781) and a retrieval date.

We have no new audit score to publish yet, so we won't claim the brand-authority score has improved. What's clear is that **off-site signals can't be engineered in one sprint** — they take time and genuine mentions from clients and partners. It's the slowest and hardest part of GEO to fake.

## If you're doing GEO on your own site

Based on this round, here's the order we'd recommend:

1. **Get SEO basics right first:** server-rendered content, consistent canonical and hreflang, a correct sitemap, and verified Search Console and Bing Webmaster Tools.
2. **Check robots.txt isn't blocking AI crawlers** from the content you want cited.
3. **Write to be quoted:** answer the question in the first lines, keep paragraphs self-contained, and source your numbers.
4. **Add Organization, Article and FAQPage structured data** so AI knows who wrote what, for which company.
5. **Build off-site presence in parallel:** Google Business Profile, LinkedIn, Wikidata, and real mentions from clients and press.
6. **Treat llms.txt as a bonus**, not a first priority.

For a fuller pre-launch list, see our [SEO/GEO launch checklist](https://noise-and-signal.com/en/en/insights/seo-geo-website-launch-checklist). To compare Taiwan vendors working on this, see [5 SEO/GEO-Ready Website Development Companies in Taiwan](https://noise-and-signal.com/en/en/insights/seo-geo-website-companies-taiwan).

## Bottom line

There's no magic switch for GEO. It's solid SEO, plus content that's easy for AI to read and quote, plus off-site presence that takes time to build. We used our own site as case number one and published what we did, what we scored and where we still fall short. If you want your site built on the same foundations, take a look at our [SEO/GEO website service](https://noise-and-signal.com/en/en/services/website) or [talk to us](https://noise-and-signal.com/en/en#contact) about where your site stands today.

## FAQ

### What is GEO?

GEO (generative engine optimization) is a set of practices that make website content easier for generative search engines such as ChatGPT, Google AI Overviews and Perplexity to understand and cite. The term comes from a paper presented at KDD 2024, which reported that GEO can boost visibility in generative engine responses by up to 40%, while noting that results vary by domain.

### How is GEO different from SEO?

Google's documentation says optimizing for generative AI search is still SEO, and there are no extra technical requirements to appear in AI Overviews. In practice, GEO builds on SEO with more attention to whether AI crawlers can fetch your pages, whether passages can be quoted on their own, and whether your brand is mentioned off-site.

### Do I need an llms.txt file for GEO?

No. llms.txt is a proposal rather than a ratified web standard, and Google says you don't need AI-specific files to appear in its AI features. We publish one as a guide for AI agents, not as a ranking lever.

### What results did Noise & Signal get from doing GEO on its own site?

In a GEO audit on September 16, 2026, our site scored 56/100 overall, with 86 for technical foundations but only 5 for brand authority, because the company barely existed on off-site platforms such as LinkedIn and Wikidata. We have since added a LinkedIn company page and a Wikidata entry to our sameAs, but as of September 27, 2026 we have no publishable data on AI citations.