SEO for ChatGPT: how to show up when an AI does the recommending

What GEO is (and why it isn't "SEO with another name")

GEO (Generative Engine Optimization) is the set of practices that get your brand to appear inside the answers language models generate — ChatGPT, Perplexity, Google AI Overviews, Claude — when a user asks about your category. The shift from classic SEO is radical but specific: you're no longer competing to land on a list of ten blue links, you're competing to be inside the single answer the AI writes.

Anyone telling you GEO is "SEO with prompts" hasn't shipped either one. They share fundamentals (authority, useful content, clean technical) but they diverge on the piece you optimize, the unit you measure and the cycle you improve on. Not understanding the difference is the first reason 80% of last year's "GEO" budgets moved nothing.

How ChatGPT decides who to recommend

To build a GEO strategy that actually works you have to understand the three mechanisms an LLM uses to decide which brand to include in its answer. These aren't tricks; they're how the tool works.

1. Training data. The model learned from a huge corpus — sites, books, forums, papers — frozen at a cutoff date. Brands that showed up frequently and in positive context inside that corpus enter as "known brands". The lever here is retroactive: what your brand has published and what others have published about you over the past few years weighs more than what you publish tomorrow.
2. Real-time retrieval (RAG, browsing). When the user asks a question that requires fresh information, ChatGPT, Perplexity and Google AI search the web right then, pull pages and use them to compose the answer. The lever here is active: your recent, well-structured, authoritative content competes prompt by prompt.
3. Structured data (schema.org). When the model retrieves your page, it doesn't just read text — it interprets the structured metadata (Organization, Service, FAQ, Article). Schema done right doubles the odds your information gets extracted accurately and quoted literally.

The five pieces of generative visibility

Any serious GEO strategy is built from the same five pieces. Whoever sells you fewer is charging for less; whoever sells you more is charging for smoke.

Semantic architecture: entities, not keywords

Classic SEO optimized for isolated keywords. GEO optimizes for entities: semantic clusters the model recognizes as "the same thing". Your company is an entity; your CEO is another; each of your products is another. The job is to make sure the model associates you with the right entities — category, location, use cases, competitors — and not with the wrong ones.

In practice that means publishing pages that connect your primary entity with every relevant secondary one (entity mapping), using consistent language to describe yourself and keeping a structured profile on Wikipedia/Wikidata when it applies.

Citable content (question-answer format)

LLMs literally quote short, self-contained fragments. A three-line paragraph that answers a concrete question is more citable than a beautifully written chapter. This doesn't mean "write worse" — it means structuring your content so every block is extractable without extra context.

• A clear question as an H2 or H3.
• An initial 40-80 word answer that closes the idea.
• After that, all the long-form development you want for humans.

Schema and structured data

JSON-LD with Organization, Service, FAQPage, Article and BreadcrumbList is the difference between the model "reading your page" and the model "understanding your page". Implementation effort is low (a technical team ships it in days) and the effect on GEO visibility is disproportionately high over the next 12 months.

Authority: external mentions and citations

Generative models still weigh authority — but with a twist: they care where you're mentioned more than how many links you have. A mention in a sector publication, a customer-published case study, a podcast that cites you: each one feeds the corpus the model will use to decide whether to include you.

The lever is slow (8-12 weeks before you see measurable impact) but it's the hardest one for competitors to copy, which makes it strategically the most valuable.

Continuous monitoring (not optional)

You can't optimize what you don't measure. Honest GEO monitoring means: a fixed battery of 50-200 representative prompts, run weekly against the main engines, logging mention rate, relative position and sentiment versus competitors. Without this, everything else is theatre — the equivalent of doing SEO without Google Search Console.

Typical mistakes: what is NOT GEO

The market has filled up with practices labeled "GEO" that aren't. Knowing how to spot them saves you months of budget and reputation.

• Running prompts and posting screenshots. Looking good in a screenshot means nothing. What counts is aggregate position across a stable battery sampled at frequency.
• Buying reviews "so the model reads them". Models detect artificial patterns with growing ease. The risk (penalty, source deindexing) outweighs the benefit.
• Prompt injection on your own site. Invisible text trying to steer the model. Doesn't work in modern versions and, when detected, damages your entity.
• Paid "plugins" or "extensions" inside ChatGPT. Organic ads don't exist right now. If someone offers you "paid featured placement", they're selling smoke.
• Optimizing only for ChatGPT. 30% of generative traffic already comes from other engines. Doing only ChatGPT is the 2005 mistake of "optimizing only for Google" all over again.

How GEO success is measured

There are five real metrics that matter. Any report that doesn't include at least three of them isn't GEO — it's noise on a pretty dashboard.

Where to start (and what can wait)

If you've invested €0 in GEO, the order that delivers the most ROI is clear and cheap. The usual mistake is trying to do everything at once — and ending up with five mediocre things instead of three solid ones.

1. Week 1-2: monitoring. Define a battery of 50-100 prompts and measure your real baseline on the 3 main engines. Without a baseline, no decision is informed.
2. Week 3-4: semantic architecture + schema. The highest-return work: the impact/cost ratio is 10x any other lever.
3. Month 2: citable content. Publish/reformat 8-12 pieces in question-answer format for the prompts where you don't appear.
4. Month 3-4: external authority. Line up 3-5 citations in relevant outlets or channels (podcasts, comparison pieces, external profiles).
5. Month 4+: continuous iteration. Every week, review the battery, adjust content and measure. Phase 1-2 never really ends.

GEO in numbers: the cost of not showing up

"Is investing in GEO really worth it?" is a fair question. Here are the public data points to keep in mind:

• An estimated 15% to 25% of B2B commercial searches already happen on generative engines (ChatGPT, Perplexity, Claude) instead of classic search engines.
• The 2027 projection has that share rising to 40-55% in B2B and professional services categories.
• In emerging categories (AI, GEO, agents), generative engines already beat Google as the first discovery source among B2B buyers under 40.
• The average cost of acquiring a customer via generative recommendation is 60-80% lower than paid SEM, according to early sector studies.

Not investing in GEO right now is the 2026 version of not investing in SEO in 2008. It's not that you'll feel it tomorrow — it's that the cost of entering late, you will.