What is GEO (Generative Engine Optimization): how to get cited by ChatGPT, Perplexity and AI Overview

The first time I noticed the shift was August 2024. A client with solid classical SEO — top three for their commercial keywords on Google — started losing traffic without anything visibly changing in the SERPs. Same position, same impressions, fewer clicks. We asked ChatGPT and Perplexity questions about their industry. Zero mentions. That was the moment it clicked.

The problem wasn't SEO. The problem was that the answer the user was looking for no longer sat in result #1. It was inside the paragraph the AI returned before the user ever reached the blue links. And in that paragraph, our client didn't exist.

That's GEO. And that's why it matters now.

What GEO actually means

Generative Engine Optimization is the work of getting generative engines — ChatGPT, Perplexity, Gemini, Claude, Bing Copilot, Google AI Overview — to cite, mention or recommend you when someone asks them about your category.

It's not the same as classical SEO, which competes for the ten blue positions. It's not the same as winning a featured snippet either. GEO is about competing for the content of the paragraph the model generates. If a user asks "what tool should I use to audit my site" and the AI answers naming three brands, GEO is the discipline that determines which three brands those are.

You'll also see the acronyms AISO (AI Search Optimization) and AEO (Answer Engine Optimization) floating around. All three point to the same idea: how to show up inside generated answers. AEO predates the LLM era (it was featured snippets, voice search) and is now considered a subset of GEO. For practical purposes pick whichever term and stay consistent across your team.

The shift that already happened

The way users reach a brand has changed over the past two years, and it changed fast.

Year	Typical query	What the user sees
2022	"best SEO audit tool"	10 blue results, one or two clicks
2024	"best SEO audit tool"	AI Overview on top + 10 results, zero or one click
2026	"what tool should I use to audit my site"	Conversation with ChatGPT, three brands mentioned, zero clicks

It isn't a uniform shift — it depends on the vertical, the country, and the intent behind the query. B2B SaaS and digital tooling feel it more sharply. Cheap physical products still see SERPs do most of the work. But the trend is clear and the data supports it: ChatGPT went from hundreds of millions to several hundred million weekly active users in under three years; Google's AI Overview now appears on a wide majority of informational queries in mature markets, and when it shows, the #1 organic result loses a meaningful chunk of its CTR.

The detail that matters: when AI Overview or ChatGPT hands the user three brands as the answer, that user rarely clicks through to verify. They assume the answer is good enough and pick from what the AI offered.

If your brand isn't in that set, you didn't play the game.

How the model decides who to mention

It's worth spending a couple of minutes here, because it changes the tactics that come next.

Large models (GPT-4, Gemini, Claude, etc.) don't store a hand-curated list of favorite brands per category. When someone asks them something, they combine two things.

First, what they learned during training — a massive corpus that includes publicly indexed web pages up to a cutoff date (Common Crawl, Wikipedia, Reddit, GitHub, forums, big blogs, news).

Second, when retrieval is enabled (the model searches the live web before answering), the first results their internal search returns for that query.

On top of those two inputs they apply confidence heuristics: high-authority domains weigh more, sources that agree with each other weigh more, structured markup is easier to parse, sentences with concrete numbers get cited more often than generic prose.

The practical takeaway: for an LLM to mention you, several conditions usually need to hold at once. Your brand needs to be in the training corpus (something built over years, not a month). It needs to show up in the first five to ten sources the model retrieves when it searches live. Your content needs to be structured so the model can extract from it without effort. And mentions of your brand across other sources need to be consistent — same name, same category, same positioning.

That's why GEO is never just on-site. What you do on your site matters a lot, but if nobody talks about your brand on Reddit, in vertical forums, or in trade media, the model has fewer reasons to trust you.

The levers that actually move the needle

There are a lot of tactics out there. These are the ones we see make a real difference.

Citable content, not filler

LLMs prefer to cite sentences they can extract without re-interpreting. If a single sentence from your post makes complete sense out of context, that sentence has a chance of showing up in an answer. If your whole page is three generic paragraphs about being "the leader" or "the most innovative," there's nothing to extract.

What works in practice:

• Numeric summaries at the start of the piece.
• Comparison tables (models parse them better than paragraphs).
• Sentences with specific numbers: not "fast," but "ninety seconds"; not "many variables," but "seventeen axes."
• Short, clean definitions — not paraphrases packed with four adjectives.

It's more a mindset shift than an effort issue. Most existing posts already have the knowledge inside; they're just written in a form the model can't use.

Schema markup that actually describes your entity

Schema.org is still one of the few ways a crawler will take your word for what you are. LLMs use it as a structural signal — not as content, but it helps.

Minimum viable: Organization with sameAs pointing to your official profiles (LinkedIn, X, GitHub, vertical sites). Article with author, datePublished and dateModified on every editorial piece. FAQPage with literal questions you know people ask — not paraphrased ones. BreadcrumbList so the model understands hierarchy. If you have human authors, Person schema with explicit credentials.

The mistake we see often: schema duplicated between <head> and body, or FAQPage with three made-up questions that don't appear on the page. At best that doesn't help, at worst it hurts.

EEAT — visible, not certified

Google formalized EEAT (Experience, Expertise, Authoritativeness, Trustworthiness) years ago. LLMs replicate similar signals because they inherit them from the search engines they use for retrieval.

The key word is "visible." Not a course, not a certificate. Just the page making clear:

• Who wrote it, with a name and ideally a photo and a short bio.
• When it was published and when it was last updated.
• Why that person can talk about this — a line or two of relevant credentials.
• Where the data came from, with links to primary sources rather than other blogs repeating the same point.

A page identical to another, but with all of that explicit, wins against an unsigned one. Small effort, big signal.

Brand SERP — the model's first impression of you

Search your brand name on Google. The first three results are the model's first impression of you. If they're mixed reviews, low-quality scrapers or forum threads with unanswered complaints, that's the narrative the AI learns.

Brand SERP work is among the fastest and cheapest things you can do. What helps:

• Your own site, well structured, schema correct, ranked #1.
• Official profiles (company LinkedIn, founder About, team page) at positions 2-5.
• Authentic reviews on relevant vertical sites (G2, Capterra, Product Hunt if it applies).
• If there's been press, make sure it's indexed and ranking.
• If an old negative result is dragging you down, push it out with better-ranked content.

You can shift a lot in a few weeks. Worth doing before throwing serious budget at content.

Distribution into sources the models read

The previous levers are all on-site. This one isn't. And it's where most teams come up short.

Training corpora for the major models over-weight certain types of source. Reddit is the most obvious and the most underused — especially well-moderated topical subreddits, where an honest mention of your brand in answer to a real question is worth more than a backlink. Wikipedia, if your brand has the notability for it (verifiable, with independent secondary press). Stack Overflow and Stack Exchange for technical niches. GitHub for dev tools. Vertical forums with historical authority. Editorial outlets of reference (in SEO: Search Engine Journal, Moz, NeilPatel; in broader marketing: HubSpot blog, MarketingProfs; in tech: Stratechery, Ben Thompson, a16z's content).

The idea isn't to spam. The idea is to be present honestly, with consistent brand mentions, ideally tied to a clear and differentiated proposition. A mention in a relevant Reddit thread where someone replies "I use X for this and it works well because Y" carries enormous weight in the future corpus.

How to measure any of this

Without measurement you're flying blind, and there are three reasonable ways to do it.

The first is the most obvious and least scalable: define ten to twenty prompts that represent how your ideal customer would ask an AI about your category. Run them in ChatGPT, Perplexity, Claude and Gemini. Note whether your brand appears, in what position, in what tone, whether it gets confused with a competitor, what's said about you. Repeat monthly. It's tedious but it's clear.

The second is monitoring mentions of your brand on sources models read. Free with Google Alerts, paid with Mention or Brand24. Every new mention on Reddit, in a forum, or in a niche blog is a vote in the future corpus.

The third is reviewing the brand SERP live: what shows up on Google when someone searches your brand by name. Those results are the first thing the AI sees when it goes to talk about you.

Doing it all by hand works if you have five or ten prompts and one brand. If you want to scale — more prompts, more models, comparisons against competitors, monthly cadence — you need automation.

What we see on almost every site we audit

After looking at thousands of reports, patterns repeat. The most common:

No real summary at the start. The post opens with a paragraph explaining what's about to be discussed, not with data. The model finds no citable sentence in the first five hundred characters and moves on.

Adjectives where there should be numbers. "Fast," "easy," "powerful," "complete." If your proposition can't be expressed as a figure, it's probably not a strong proposition to begin with.

Empty or misconfigured schema. Organization with only name and url, no sameAs, no logo. Or worse, FAQPage repeated on every post with the same three generic questions nobody would actually ask.

Pages with no visible author and no date. The page exists in a vacuum, unsigned, undated. The model can't tell if it's reading something from 2019 or last week.

Internal links with generic anchor text — "click here," "more information," "read." Each anchor is an opportunity to tell the model which entity you're linking to, and most pages waste it.

Brand SERP nobody works on. You search the brand by name and find scrapers, old reviews, weird things. That's the first impression the AI has of you.

And the big one: zero presence on Reddit, in niche forums, in the places where real people ask real questions and the models absorb the answers. It's the lever with the most leverage, and the one fewest teams work on.

If you recognize your own site in three or more of these, the problem isn't content volume. It's citable structure.

Where to start next week

Without trying to give a perfect recipe — every case is different — this is what should reasonably be in place in the next thirty days.

Add a real summary with data at the start of your ten most important pieces. Four to six sentences each, with specific numbers. This alone, done well, moves a lot.

Run your schema through a validator and clean it up. Google's Schema.org Validator is free. Get Organization, Article with author and dates, and FAQPage with real questions — not the ones that sound good in the abstract.

Identify the human author of each post and give them visibility. Photo, short bio, link to LinkedIn. If everything is signed by "the editorial team," no post has a real author from the model's perspective.

Search your brand on Google and look at the top ten positions. Note which negative or irrelevant result bothers you most. That's the first brand SERP battle.

Define ten representative prompts and run them this month in ChatGPT and Perplexity, with results noted down. Run them again next month. The comparison will teach you more than any GEO course.

Identify three to five subreddits or forums where your ideal customer lives. Not to post your content; to start participating honestly. If you don't participate, you don't exist there.

Rewrite the "About" page and the home with specific claims. Drop "leader in," "innovative," "the best solution on the market." Replace every vague adjective with a fact.

Check the consistency of your brand name across all external mentions. Same spelling, no accent variations, no alternate names. Sounds obvious, almost nobody does it.

Do all of this quarterly. It's not a fix, it's a practice.

Questions we get asked a lot