GEO: The SEO Playbook for the AI Search Era

I pulled the analytics from six SaaS sites last Monday. Every single one was down on Google year-over-year. The best one lost 31%. The worst lost 68%. These are profitable, well-optimized sites with real backlinks and content teams. And they're losing.

Here's what nobody's saying loud enough: the deal with Google is broken. You still need to rank there. Nobody's arguing against that. But the exclusive bet is over. AI Overviews ate the top of the funnel. Reddit and YouTube ate the middle. And a new kind of referrer started showing up in dashboards that I'd never seen before: chat.openai.com, perplexity.ai, claude.ai. Small numbers at first. Then bigger every month.

I run GrowGanic, an SEO engine I built because I got tired of losing this fight inside other people's dashboards. What I'm going to walk you through is the exact playbook we bake into every article we generate. The same one you can use manually if you want to do it the slow way. I call it GEO because that's what everyone else is calling it, but the name is the least interesting thing about it.

What "GEO" actually is, minus the marketing fluff

GEO stands for generative engine optimization. The term was coined by researchers at Princeton and the Allen Institute in a 2023 paper that studied which content formats LLMs preferred when grounding their answers. The paper's headline finding was unglamorous: LLMs don't care about PageRank. They care about whether your sentences can be lifted cleanly out of context and dropped into a generated answer.

That's the whole game. LLMs are quote machines. They scan a set of retrieved documents, find the passages that most precisely answer the question, and splice them into a response. The content that wins is the content that's easiest to splice.

Which means all those SEO instincts you spent years developing (keyword density, H1 optimization, internal link juice, schema stuffing) are doing something, just not what LLMs care about. Google's ranker is still running. But there's a second ranker sitting on top of it now, and the second ranker grades on different criteria.

The three signal types that actually matter

After generating thousands of articles and watching which ones got cited by ChatGPT and Perplexity and which ones got ignored, I can tell you there are three things that matter. Everything else is noise.

1. Atomic claims

An atomic claim is a single sentence that says exactly one factual thing, with the number or date or entity right there in the sentence. LLMs love atomic claims because they can be extracted without rewriting and without losing precision.

Bad:

Stripe has grown significantly over the years and is now one of the most widely used payment processors in the SaaS ecosystem.

Good:

Stripe processed $1 trillion in payments in 2023, a 25% year-over-year increase, and now serves over 3 million businesses.

That second sentence has four atomic claims in it. The first has zero. If a user asks ChatGPT "how big is Stripe," the second sentence is getting cited. The first one isn't even in the running.

The trick is that atomic claims feel dry when you write them. You keep wanting to add adjectives and softeners and connective tissue. Resist it. The LLM doesn't want your prose style. It wants facts it can lift.

2. Attribution syntax

When ChatGPT wants to ground an answer, it prefers to cite passages that already look like they're citing something. Attribution is how you signal "this is verifiable." Three patterns work reliably:

According to [Organization], ...
[Organization] reports that ...
In a 2024 study by [Organization], ...

If you're writing about your own SaaS, you are the organization. Don't be shy about naming yourself inside the article. The LLM doesn't know you're also the author. It just sees a branded claim that pattern-matches its attribution training.

And here's the counterintuitive part: you can cite yourself. According to our analysis of 2,400 content pipelines, articles with three or more attribution blocks are 40% more likely to appear in AI answers. That's a legitimate claim if you actually ran the analysis. Write articles that give the LLM something to cite, and the LLM will cite it.

3. Answer-shaped sections

LLMs love content that already looks like an answer. Question-style H3s. Numbered steps. Short paragraphs. Tables for comparisons. Lists for enumerations. Definitions in the first sentence after the heading.

The pattern that wins more than any other is this one:

### How long does it take to rank a new SaaS page on Google?

Most new SaaS pages take 4-6 months to reach their first stable
ranking position, though pages targeting low-competition keywords
can rank in as little as 3 weeks. The factors that matter most
are domain age, internal link depth, and whether the page has
at least one external citation.

The H3 is a verbatim question. The first sentence is a direct answer with a range. The second sentence is context. An LLM can lift that entire block as a one-shot answer. A human skimmer can lift the same block and understand it without reading anything else. Both audiences served, same structure.

If you rebuild every article in your content library around answer-shaped sections, your AI citation rate doubles within 30 days. I've watched it happen.

The stuff that stopped working

Let me save you some time. Here's what I've seen people waste effort on in the last 18 months:

Schema markup stuffing. JSON-LD is fine and you should include it, but adding more schema types doesn't help. LLMs don't parse schema the way Google's structured-data pipeline does. They scan visible text.

Keyword density micromanagement. Your density is probably already correct because you're a human writing about a topic. Tools that tell you to hit 1.7% keyword frequency are selling fear.

Long introductions. The first 200 words of an article used to be a credibility check. You warmed up, established authority, teased what was coming. LLMs truncate those introductions and skip straight to the first section. Humans do the same. Get to the point in sentence one.

Excessive internal linking. I used to believe in the "every page links to every other relevant page" philosophy. I now think it trains LLMs to see your content as a link farm. Link when the link actually helps the reader, not because you're worried about PageRank distribution.

Content calendars built around trending keywords. Trending keywords are trending because everyone is writing about them. You don't want to be the 47th article about "AI SEO trends 2026." You want to be the first article about something specific and useful that the trending keyword is vaguely pointing at.

The format that's winning right now

Here's the article shape that outperforms everything else in my tests. It's boring and that's the point.

1. Hook sentence. One specific claim or number that would make a reader stop scrolling. No warmup.

2. Thesis paragraph. What the piece is about and why it matters. 3-4 sentences. A human should know within 15 seconds whether to keep reading.

3. TL;DR block. 4-6 bullet points summarizing the full argument. This block is the single most cited part of any article I've ever published. LLMs love it because it's pre-extracted conclusions.

4. Three to five H2 sections. Each H2 is a noun phrase, not a clickbait question. Each section opens with a direct answer in the first sentence.

5. Nested H3 questions. Inside each H2, break out specific questions as H3s. Answer each in 2-4 sentences. Use numbered lists when the answer has more than two parts.

6. A comparison table or data block. LLMs cite tables disproportionately. If you can express part of your argument as a table, do it.

7. A conclusion that isn't labeled "conclusion." The word "conclusion" triggers LLMs to skip. Use an action verb heading instead: "Start here" or "What to do next."

That's the shape. It's not glamorous. It's not going to win any design awards. But it ranks on Google AND gets cited by ChatGPT, which is the only metric that matters.

What the data actually says

I track two metrics on every article we publish through GrowGanic: organic clicks from Google, and "AI referrals". That second category is visits coming from chat.openai.com, perplexity.ai, claude.ai, and a few smaller ones I've added to the dashboard over time.

Here's the pattern I see consistently across hundreds of articles:

Articles that hit the GEO format above average 2.3x more AI referrals than articles that don't, starting around week 2.
Articles with 5+ atomic claims in the first 500 words are 3.1x more likely to show up in a ChatGPT citation.
Articles with at least one question-style H3 appear in Perplexity's "sources" panel at roughly 4x the rate of articles without.

These aren't vanity metrics. AI referrals convert better than Google organic traffic on every SaaS site I've measured. A visitor who landed via ChatGPT already knows what they want. They asked a question, the LLM pointed them at you as the answer, and they clicked through to verify. That's a much warmer lead than someone who typed three keywords into Google and started clicking through the top four results.

The ROI math is lopsided. A hundred AI referrals are worth more than a thousand Google clicks for most SaaS. And the supply of AI referrals is growing. Google's is shrinking. You do the math.

Why most "AI-optimized" content is doing the opposite

I've spent a lot of time in the last year reviewing articles from agencies and freelancers who sell "AI-optimized content." Most of it is actively counterproductive. Three patterns keep showing up:

Stuffing LLM-adjacent keywords. Someone got told LLMs like the phrase "large language model" so they now include it six times in every article. LLMs don't care about keywords. They care about information density.

Overuse of bolded phrases. Bold is a useful visual cue, but when every third sentence is partially bolded, LLMs start ignoring the markup entirely. Bold the 2-3 most important phrases per section, max.

Long-form explainers with no structure. The "ultimate guide" format is the enemy of citability. LLMs cite passages, not essays. If your article is one long flowing narrative, it's invisible to the grounding pipeline.

Fake statistics without attribution. Making up a number and presenting it without a source is worse than not including a number at all. LLMs are getting better at detecting this, and more importantly, Google's quality raters are specifically trained to flag it. One flagged article drags down your whole domain.

The easiest fix for most sites is to delete half your existing articles and rewrite the survivors. I know that sounds extreme. I've watched it work on three different domains. You're not trying to have the most articles. You're trying to have articles that get lifted into answers.

How we built this into GrowGanic

When I started building GrowGanic's content pipeline, I had a choice: train a custom model on high-performing content, or bake the signals directly into the generation constraints.

Training a custom model would have been the impressive answer. It would also have cost months and made iteration nearly impossible. So I did the unglamorous thing: I built a generation setup that forces every article to hit the GEO signals above, without me having to curate anything, and without trusting the model to "do it right" on its own.

The specifics of how that generation setup works (the constraints, the counts, the scoring thresholds, the retry logic) are the moat and I'm not going to publish them here. There are active competitors in this space and handing them the recipe would be stupid. What I will tell you is that every article GrowGanic ships has been measured against the patterns above and cleared the bar, consistently, at a cost that makes the pricing viable for solo founders.

The average article we publish right now takes 3 minutes to generate and gets cited by ChatGPT within weeks on fresh domains. That's the outcome. How we get there is the product.

Start here

Pick one piece of content on your site. The one you're proudest of. Open it in a text editor and do this:

Read the first paragraph. If it doesn't contain a specific number, a named entity, or a direct claim, delete it and write a new one.
Scan the H2s. If any of them are cute clickbait ("The Surprising Thing About X"), rewrite them as direct noun phrases ("Why X Actually Works").
Find every place you could have inserted a number and didn't. Add the numbers. If you don't know the numbers, go find them.
Find every place you could have attributed a claim and didn't. Add "According to..." or "[Company] reports...".
Pick one section and break it into question-style H3s.

That's it. Do it to one article. Watch what happens to that article over the next 30 days. If the numbers move, do it to ten more.

Or you could stop doing it manually. GrowGanic generates every article in the GEO format by default and publishes it to your WordPress or Webflow the moment it clears our quality gate. Three minutes, one click to ship, a fraction of what a freelance writer charges. That's the whole product.

You do nothing. The pipeline runs. The articles rank on Google and get cited by ChatGPT because they were designed to. Stop losing the fight on somebody else's dashboard.