GEO (generative engine optimization) is the work of raising the odds that ChatGPT, Gemini, Perplexity and Google AI Overviews mention your brand and cite your pages in their answers. GEO builds on SEO rather than replacing it.
The short version: SEO gets your pages into the set an engine reads. GEO aims to make you one of the few sources that make it into the answer, and to make sure the model knows who you are. We rewrote this entry in October 2026, based on an audit of more than 100 websites and the research we track every week in our newsletter.
Why GEO matters for your brand
Your buyers ask ChatGPT before they ask you. They get one paragraph back. If your brand is not in that paragraph, you are out of the shortlist before anyone visits your site, and you will not see it in Analytics because there was no click.
Here is the finding that changed how we work. In Suganthan Mohanadasan’s analysis of network traffic from 60 ChatGPT conversations, brands the model wrote into its own first search query were mentioned in 68.9% of answers. Brands whose pages were fetched but never named in a query were mentioned in 2.1% (Suganthan Mohanadasan, August 2026). The model picks its shortlist before it searches. That means GEO is two games, not one.
How GEO works
An AI engine knows your brand in two ways:
- Memory (training data). Whatever the model read about you during training. This moves slowly, over months, and it is mostly won off your site: people writing about you, reviewing you and comparing you with others.
- Citation (grounding). When someone asks, the model searches the web, reads pages and bases its answer on them. This moves fast and is largely in your hands: crawlable pages, a clear answer paragraph, structured data.
Two details matter on the citation side. First, engines usually do not open your page live. They read a copy they crawled earlier, so a fix shows up only after the page is recrawled. Second, engines do not share one source. ChatGPT now runs its own family of search indexes, called Labrador internally, alongside several outside providers (Tomek Rudzki, Peec AI, September 2026). “We’re indexed in Bing, so we’re in ChatGPT” no longer holds. Google AI Overviews draws on Google’s own index.
In client work we explain this as three doors:
- Reach: Can AI crawlers get into your site?
- Recognize: Does the engine know you as a distinct entity?
- Believe: Does the engine trust what it says about you?
The SEO and GEO connection
We disagree with “SEO is dead, GEO is here.” Google’s Gary Illyes said normal SEO practices are enough to appear in AI Overviews (Search Engine Land, July 2025). A page that cannot be crawled, indexed or trusted will not make it into any engine’s answer.
GEO adds three things on top of that foundation. You measure citations, not rankings. You track several engines separately. And your reputation off your own site moves to the center of the plan. The term itself comes from research: the GEO paper from Princeton and partners (November 2023, KDD 2024) was the first systematic test of how content methods change visibility in generative engines.
What we’ve seen in the field
Through September 2026, we audited more than 100 websites in seven countries with the GEO Score Card. Every figure is a share of all audited sites, weighted equally:
- About 3 in 4 sites give AI crawlers no explicit access in robots.txt. Some block GPTBot, ClaudeBot or PerplexityBot outright.
- About 87% have no Wikidata entry, and about 30% lack basic Organization schema.
- About half have no independent reputation signal.
- Nearly 9 in 10 have closed at least one of the three doors. Most of the fixes cost nothing.
Full findings and method: The Three Doors of AI Visibility: A Study of 100+ Websites.
One lesson came from our own brand. In August 2026 we asked Gemini about Stradiji. With web search on, it described us correctly. With search off, it did not know us at all, mistook the name for Stradivarius and declined to answer. We were winning the citation game and missing from the memory game.
A practical GEO playbook
- Measure before you write. List the 20 to 25 questions your buyers actually ask. Run them in at least two engines, on different days, in fresh chats, and read the source list under each answer. Log the model name and date with every result. When the model changes, so do the results.
- Split your questions in two. Ask each one with web search on and off. If the answer changes, the question sits on the citation side and can move within weeks. If it does not, it sits on the memory side. Do not promise quick wins there.
- Open the doors in order. Start with robots.txt and your CDN bot settings. Treat OAI-SearchBot (search) and GPTBot (training) as separate decisions. Then entity work: Organization schema, a sameAs chain and a Wikidata item. Reputation comes last.
- Put a stand-alone answer at the top of every key page. Models lift a passage, not a page, and they weigh the title and the first few hundred characters of the body. Don’t bury the answer in paragraph 14.
- Lead with proof, not claims. “We’re the best in the industry” does not get cited. A page that documents an independent ranking, your own measurement or a client result does.
- Show up the same way everywhere. Describe your brand with the same category line on LinkedIn, directories, press releases and speaker bios. Models read mixed signals as an unclear identity.
What not to do: manufacture citations, add fake schema or stack “best of” lists where you rank yourself first. In Lily Ray’s review of 100 B2B queries, the brand that put itself first on its own list was not recommended by the AI answer in 224 of 323 cases, or 69% (Search Engine Land, June 2026). These tactics can look like they work for a while. They carry long-term risk.
Related terms
- AEO (answer engine optimization)
- Grounding
- Query fan-out
- Citation optimization
- Entity optimization
- AI visibility
- E-E-A-T
Frequently asked questions
Is GEO replacing SEO?
No. GEO builds on SEO. A page that cannot be crawled or indexed will not appear in any AI answer.
How long does GEO take to show results?
It depends on the question. Where the model searches the web to answer, page and schema fixes can show up within weeks of a recrawl. Where it answers from memory, change takes months and depends on your reputation off your own site.
How do you measure GEO?
By citations, not rankings. Ask the same questions on different days in at least two engines, then count how often your brand is mentioned and which of your pages are cited. AI answers are personalized, so one test proves little. Repeated tests do.
Is GEO risky?
The shortcuts are. Fake citations, fake schema and self-promotional list stacking count as spam. Real GEO runs on real authority and verifiable proof.
Do I need an llms.txt file for GEO?
No. Google said it will not use llms.txt for AI Overviews (Search Engine Land, July 2025). It does no harm, but it is not near the top of the list.
