SEO and GEO: get found by Google and ChatGPT

Agencies sell “GEO” as a new discipline with new tricks. The platforms themselves tell a duller, more useful story. Here’s what they say, what research shows and how solid each piece of evidence is.

Evidence labels: ✅ primary source (platform docs, paper). ⚠ secondary or vendor study: usable, but the maker often sells the measuring tool.

How it really works

• Google has no separate AI ranking. AI Overviews and AI Mode fetch pages via the regular ranking systems and cite them. Google literally calls GEO “still SEO”. ✅
• Query fan-out. Google splits one question into several sub-questions. So you can be cited for a sub-question without ranking for the main question. ✅
• ChatGPT only cites what its own search step fetched. Fetching explains ~32% of the variation in citations, page characteristics only ~2.8%. Google ranking predicts ChatGPT citations poorly. ✅ preprint (arXiv 2610.15033, 2610.15382; October 2026, not peer-reviewed)
• Known brands win. ChatGPT writes brands it knows into its own search queries and also mentions them without a source page. ✅ preprint
• Search and training are separate bots. OpenAI: OAI-SearchBot (search) vs GPTBot (training). Anthropic: Claude-SearchBot and Claude-User vs ClaudeBot. If you block the search bot, at OpenAI you no longer appear in search answers (at most as a navigation link) and at Anthropic your visibility drops. ✅
• Location matters. ChatGPT follows the location from the app down to city level. Test from the Netherlands: set your location in ChatGPT to the Netherlands (a VPN makes no difference there); Gemini, on the other hand, follows your network country. ✅ preprint (arXiv 2610.24081)

flowchart LR
  V["User question"] --> F["Split into sub-questions"]
  F --> Z["Search step fetches pages"]
  Z --> M["Model picks passages<br/>that support the claim"]
  M --> C["Answer with citations"]
  B["Robots.txt / CDN blocks"] -. "no fetch = no citation" .-> Z

Who does it perfectly

Wikipedia

• Does: neutral, sourced, up-to-date pages per entity.
• Result: most cited domain in ChatGPT (16.3%) in Ahrefs’ study of 78.6 million queries. ⚠ vendor study
• Steal this: write one neutral, sourced page per topic that you keep up to date.

Reddit

• Does: real experiences in question-and-answer form.
• Result: according to several vendor studies, among ChatGPT’s most cited sources. ⚠
• Steal this: write like someone who has done it, with the question as the heading.

Ramp

• Does: optimised its content on accounts payable software with Profound’s tool.
• Result: AI visibility from 3.2% to 22.2% in about a month. ⚠ case by Profound, the tool that measured it
• Steal this: comparison pages with criteria, tables and an honest outcome.

On this site you can see the principle in practice: best AI assistant and best way to make money with AI are honest comparisons with a direct answer at the top and structured data.

Do this

1. Let the search bots in (30 minutes)

1. Choose per bot: allow search, you may refuse training.
2. Check your CDN or hosting (for example Cloudflare “AI bots” or “Bot Fight Mode”): they often block silently.
3. Check in your server logs whether OAI-SearchBot, Claude-SearchBot and PerplexityBot get a 200 response.

Example “found, but not trained on”:

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: *
Allow: /

Sitemap: https://<YOUR-DOMAIN>/sitemap.xml

Blocking Google-Extended can cost you Gemini

Google-Extended is not a ranking signal for Search. But Google wrote in October 2026: “Websites that specifically allow Google-Extended are eligible to appear in responses and receive clicks via grounding in Gemini Apps.” ✅ If you block it to refuse training, you then probably no longer appear as a source in the Gemini app (our inference). Note: Google will count the planned shared AI training token as blocking Google-Extended.

2. Make sure your text is in the HTML

1. Render important content server-side. Many AI fetchers run little JavaScript. ⚠
2. No nosnippet or noindex on pages you want cited: Google requires “indexed and eligible to be shown with a snippet”. ✅
3. Don’t switch off the setting for generative AI features in Search Console; otherwise you won’t appear in AI Overviews or AI Mode. ✅
4. Test it yourself:

curl -A "OAI-SearchBot" https://<YOUR-URL> | grep "<A-SENTENCE-FROM-YOUR-TEXT>"

A 200 here doesn’t prove the real bot (own IP ranges) gets through; check your logs too.

3. Heading = the question, first sentence = the answer

1. Rewrite your ten most visited pages: every H2 is a question as your client asks it.
2. Answer that question in the first sentence, only then explain.

Within one site, pages whose title and paragraphs match the question asked were cited more often. FAQ blocks and structured data only helped between sites. ✅ preprint (arXiv 2610.15033). Google does add: you don’t need to chop your text into pieces for AI.

4. Sources, quotes and your own numbers

This trick works mainly when you're not at the top

In the GEO paper (KDD 2024), citing sources, adding quotes and adding statistics were the best methods: +30 to 40% on one visibility measure (word share), 15-30% on the other. Citing sources gave +115% for a page in position 5, and −30% in position 1. Keyword stuffing did worse than doing nothing in the main test. ✅ Caveat: lab test from 2023.

Per page at least: 3 links to primary sources, 1 quote, 1 number with a date. Check your page here:

Your page's text or HTML

Free · runs entirely in your browser · nothing is stored or sent

5. Become a brand others talk about

Mentions beat backlinks

Brand mentions on other sites correlate 0.664 with visibility in AI Overviews, backlinks only 0.218 (Ahrefs, 75,000 brands). ⚠ Correlation, not proof. It does fit the mechanism: ChatGPT mentions brands it already knows. ✅

1. The same name and bio everywhere. Add Person or Organization schema with sameAs pointing to your profiles.
2. Aim for 3 mentions per month elsewhere: guest podcast, trade magazine, someone else’s newsletter, a journalist request.

6. Honest comparisons, without putting yourself at 1

Putting yourself at number 1 works less and less

Lily Ray tested ~100 B2B queries. AI Overviews did cite self-promotional listicles, but left out the publisher’s own brand in 69% of cases (April–June 2026). ⚠ Small sample, not replicated. So make comparisons you have no stake in, or present your own product honestly, including its downsides.

Grey: possible, with risk

Tactic	Risk	Safer
Programmatic SEO (pages from template + dataset)	Google calls mass-producing pages to manipulate “scaled content abuse”, also per fan-out variant. ✅	Only if every page has unique data. Start with 50, scale up only if most get indexed.
AI content at scale	Using AI is fine, “many pages without adding value” is not. ✅	Don’t publish faster than you can edit. Every page contains something a model can’t make up.
Buying an expired domain	“Expired domain abuse” if you fill it with content without value. ✅	Only in the same niche, continuing similar content.
Reddit and forums	Sockpuppets and hidden advertising: ban, and Google says fake mentions “aren’t as useful as they seem”. ✅	Respond as yourself, stating your affiliation.

Risk: only bumping the date

Updating for freshness is fine. Changing the date without a substantive change is misleading, and Google requires structured data to match the visible text. ✅

Black: how it works and why you don’t do it

Tactic	How it’s detected and penalised
Cloaking: showing different text to AI bots than to people	Explicit spam policy, also for text shown only to certain user agents. Algorithmically and via manual actions. ✅
Hidden text or hidden instructions for AI systems in your page	Hidden text abuse. ✅ Researchers publish detection methods, providers invest in them. Risk: de-indexing and reputational damage.
Link networks (PBNs) and bought links	Link spam policy, manual action “unnatural links”. Google’s systems ignore such links: you pay for nothing. ✅
Parasite SEO on a strong domain	Site reputation abuse. Outside the EEA a manual action simply still applies; in the EEA you still get it in Search Console, but that section is ranked separately (since 30 August 2026). A Dutch site is also found outside the EEA. ✅
Back button hijacking	New policy, enforced since 15 June 2026. Can also come from ad scripts. ✅

Source: Google spam policies.

Security

• Search Console is the key to your visibility. Every quarter, review the list of owners and users and remove former staff and agencies.
• Domain and DNS are the foundation. Turn on 2FA or a passkey at your registrar and DNS provider (such as Cloudflare) and enable the registrar lock.
• A shared AI chat is a publication. In 2025, thousands of shared ChatGPT chats and hundreds of thousands of Grok chats ended up in Google. Never put client or strategy data in a share (what not to share).

Myths

• “llms.txt is the new robots.txt.” Google doesn’t use it. ✅ Ahrefs saw in server logs of 137,000 domains that 97% of llms.txt files got zero requests. ⚠
• “You need special AI schema.” Google: no special markup needed. ✅
• “Chop your content into chunks for AI.” Google: “no requirement to break your content into tiny pieces”. ✅
• “This tool sees Google’s internal AI scores.” Google: “No third-party tool has access to our internal ranking or AI systems.” ✅
• “The 2024 Google leak revealed the weights.” The leak contained attributes, not weights. ✅
• “Checking once in ChatGPT tells you enough.” Answers vary per run and per location. ✅

Measuring

• Search Console, Performance: AI Overviews and AI Mode are included in the total figures. ✅
• Search Console, Generative AI performance report: impressions in AI Overviews, AI Mode and Discover. Available for all sites since 31 August 2026, still without click data. ✅
• GA4: create your own channel group with this regex, and place it above “Referral”:

chatgpt\.com|chat\.openai\.com|gemini\.google\.com|claude\.ai|perplexity\.ai|copilot\.microsoft\.com

• Server logs: count hits from ChatGPT-User, Claude-User and Perplexity-User per URL. Free and the most honest signal you have.
• Fixed prompt set: 20 to 50 prompts, monthly, from the Netherlands, multiple runs. Only compare with yourself over time.
• Branded searches in Search Console: people who google your brand after an AI mentioned you.

Clicks really are falling: with an AI Overview, position 1 gets ~58% fewer clicks (Ahrefs, December 2025 data) ⚠. So measure visibility too, not just clicks.

Your first 30 days

For a freelancer selling AI services (chatbots, automations):

1. Day 1: check robots.txt and CDN, connect Search Console and Bing Webmaster Tools.
2. Day 2: curl test on your five most important pages.
3. Week 1: write down 20 questions clients really ask you. That’s your prompt set and your list of headings.
4. Week 1: baseline: test the prompt set in ChatGPT, Perplexity and Google AI Mode, from the Netherlands.
5. Week 2: rewrite your service pages: question as heading, answer in sentence one, one real client result with a number.
6. Week 3: publish one measurement of your own, for example “we tested 5 chatbot builders on 30 Dutch customer questions”, with raw data.
7. Week 3-4: pitch that measurement to two trade magazines, newsletters or podcasts.
8. Day 30: run the prompt set again and compare with your baseline.

Checklist

Checklist

0/4

• Search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot) get a 200, also via my CDN
• My most important text is in the HTML, without nosnippet
• My service pages have questions as headings and the answer in the first sentence
• I have a fixed prompt set and a baseline from the Netherlands

Your progress is only stored in this browser.

Quiz

Test yourself

1. 1What does Google itself say about optimising for AI Overviews?

   It's a separate discipline with its own ranking factorsIt's still SEO: AI Overviews use the regular indexYou need an llms.txt and special schema

2. 2You block GPTBot but allow OAI-SearchBot. What happens?

   You disappear from ChatGPT search answersOpenAI doesn't train on your site, but you can still appear in ChatGPT search answersNothing, OpenAI ignores robots.txt

3. 3For whom does "citing sources" work best in the GEO paper?

   For pages in position 1For lower-ranked pages, such as position 5For everyone equally