For 25 years, “rank on Google” meant rank in the blue links. The blue links still exist, but a growing slice of search now ends in an AI-generated answer: Google’s AI Overviews, ChatGPT, Claude, Perplexity, Copilot, Arc Search, Brave’s Leo. Sometimes the user clicks through to a source. Often they don’t.

This is the territory people have started calling Generative Engine Optimization, or GEO. The name is awkward and probably won’t stick, but the underlying shift is real. Search behavior has changed. The question is what to do about it.

What’s actually different

Traditional SEO optimizes for click-through from a results page. The user types a query, sees ten blue links, picks one.

AI search compresses that. The model reads dozens of sources, writes an answer, and sometimes cites the sources at the bottom. The “click” is optional from the user’s perspective.

That changes a few things at once:

Quotability matters more than ranking position. The model picks specific sentences, not pages. Clear, scannable, citable facts have a much better chance of being included than ones buried inside long paragraphs.
Definitive answers beat fluff. Models extract sentences they can attribute. “It depends on your situation” doesn’t make it in. “Hreflang requires a self-referencing tag on every page” does.
Brand visibility decouples from clicks. Being cited in an AI answer counts as exposure even if the user never visits your site. Conversely, ranking #1 for a query where Google now shows an AI Overview can mean far fewer clicks than it used to.
Real-time crawlers matter alongside training crawlers. GPTBot trains the model. ChatGPT-User fetches pages in real time when ChatGPT browses the web. They’re different bots and you can allow or block them independently. We wrote about this in Why Robots.txt Matters for SEO and AI Visibility.

The TL;DR: traffic from AI answers is real but smaller than organic search. What you’re optimizing for is being part of the answer, not just being a link near it.

What we’ve actually seen work

We’ve been tracking AI citations for our own site and a handful of client sites since mid-2025. A few patterns keep showing up:

1. Direct, declarative writing. Posts that lead with a clear thesis and define terms early get cited more. Posts that build to an answer over six paragraphs of context don’t.

2. Specific, attributable claims. “Brotli compresses 15-25% better than Gzip for text” is a sentence a model can lift verbatim. “Brotli is more efficient” isn’t. The more concrete and quotable, the better.

3. Structured data and clear semantics. FAQ schema, How-To schema, Article schema with author info. Models seem to use these as hints about what part of the page is the actual answer.

4. Recency. Models with browsing capability prefer fresh sources. A post dated 2024 is much less likely to be cited than one updated for 2026, even if the content is identical.

5. Authority signals that look like SEO authority. Backlinks, brand mentions, domain age, consistent author bylines. The AI ranking signals appear to overlap heavily with the search ranking signals, because the same training data feeds both.

What hasn’t worked, in our experience: keyword stuffing, “GEO-specific” hacks (sentence injection, prompt injection hidden in HTML, comments aimed at LLMs), and aggressive structured data on irrelevant pages. Models seem fine at ignoring all of these. Some risk getting you penalized in classic SEO too.

A practical playbook

If you’re building or auditing content for AI visibility in 2026, here’s the order we’d work in.

Step 1: Decide who’s allowed to crawl you

Open your robots.txt and look at what you’re blocking. The major AI crawlers as of mid-2026:

GPTBot (OpenAI training)
ChatGPT-User (OpenAI real-time browsing)
ClaudeBot, anthropic-ai (Anthropic training)
Claude-Web (Anthropic real-time browsing)
PerplexityBot
Google-Extended (Gemini training, separate from Googlebot)
Bytespider (ByteDance)
CCBot (Common Crawl, feeds many models)

You have a real choice to make here. Allow them all and you maximize visibility. Block training crawlers but allow real-time browsing and you get cited in answers without contributing to training. Block everything and you get neither.

There’s no objectively right answer. Just don’t leave it on default and pretend you decided.

Step 2: Write content that’s quotable

Some practical edits we’ve made on actual client posts:

Replace soft headlines with definitive ones. “How fast should my site load?” → “How fast should my site load? Under 2.5 seconds for LCP.”
Lead each section with a one-sentence answer, then expand. The first sentence is what the model is most likely to lift.
Use specific numbers, dates, and product names. Vague qualitative writing doesn’t get cited.
Add a TL;DR at the top of long posts. Models pick them up.
Use FAQ schema for actual FAQ content. Real questions and answers, not keyword-stuffed.

You can write naturally and still do all of this. We’re not asking you to talk like a robot. We’re asking you to be specific.

Step 3: Build authority that travels

The signals AI engines use to decide what to trust mostly look like classical SEO signals: links from reputable sites, mentions in industry conversation, content authored by named humans with bios.

For agency sites like ours, that means:

Real author bios with credentials, not “Admin”
Consistent author across posts in a topic area
Earned mentions in industry conversations (Slack groups, Reddit, HN, niche newsletters)
Get cited on Wikipedia where appropriate — Wikipedia is heavily weighted in training data

This is just SEO 101 with a slightly different motivation.

Step 4: Track what’s working

The tooling for this is still rough. As of mid-2026, the options are:

Manual checks. Ask ChatGPT, Claude, and Perplexity questions in your domain and see if you get cited. Tedious but honest.
Otterly.ai, Profound, Peec AI, Ahrefs Brand Radar. Track brand mentions across AI answer engines.
Google Search Console. Doesn’t yet break out AI Overview impressions reliably, but the “search appearance” filter is starting to surface them.
Server logs. Look for ChatGPT-User, PerplexityBot, Claude-Web user agents hitting your pages. That’s real-time AI retrieval, and it’s a leading indicator of citations.

We mostly use a mix of GSC, log analysis, and weekly prompt-checks against the top engines. It’s a developing space and nobody has a clean answer yet.

What about traditional SEO?

Traditional SEO isn’t dead. It’s the foundation GEO sits on. The same content that ranks in Google is also the content that gets cited by AI engines, because they were trained on the open web Google has been indexing for decades.

Don’t trade SEO work for GEO work. Do SEO well and you’ll naturally do better in AI answers. The marginal GEO work — being more declarative, using structured data, tracking citations — is on top of solid SEO, not instead of it.

If anything, the basics matter more now. A site with broken links, slow load times, and confusing structure is going to be filtered out by both Google and the AI engines. We have plenty of posts on those basics — start with SEO Page Analysis and Broken Links: The Silent SEO Killer if you haven’t read them.

The honest take

Nobody knows how much of search will end up in AI answers in five years. It might be 20%. It might be 60%. Anyone giving you a confident number is guessing.

What we do know: optimizing for clarity, accuracy, and structure helps you in both worlds. The work isn’t fundamentally different from good content marketing. It’s the same work, with a small set of new considerations on top.

If you’d like a second pair of eyes on your site’s AI visibility, or a content strategy that holds up in both classic search and AI answers, get in touch.

Generative Engine Optimization (GEO)