Generative engine optimization

What GEO is, and what six AI models actually do

Generative engine optimization (GEO) is the practice of getting a brand accurately described on the sources AI assistants read, so that when someone asks about a category the generated answer has a reason to name it. It optimises toward systems that write one answer instead of returning a list of links.

Almost everything written about GEO cites the same two or three third-party studies. This page does not. We run live checks: real buyer questions, put to six assistants with the web open, every answer and every source kept. What follows comes from 363 recorded answers carrying 5,148 citations, and you can open any of them.

GEO, AEO, AI SEO, LLM SEO: one job, several names

The vocabulary is unsettled because the field is new. These all describe the same work, and each name emphasises a different part of it.

AI SEO: The broad one. Any work aimed at getting a brand into the answers assistants give.
AI search engine optimization: The long form of the same thing, used when people mean search specifically rather than chat.
Generative engine optimizationGEO: The term the industry settled on for optimising toward systems that generate an answer instead of returning links.
Answer engine optimizationAEO: Emphasises the output: there is one answer, and the question is whether you are in it.
LLM SEO: Names the thing doing the reading. Same work, framed around the model rather than the search box.
AI visibility: The measurable side: whether, and how often, a brand actually appears.

One distinction is worth keeping straight, because it splits the search results in two: “AI SEO” is also used for using AI to do your SEO work, which is a different subject. This page is about the other meaning, getting into the answers.

How a generative engine builds an answer

Worth walking through slowly, because every GEO decision falls out of it. When someone asks a buying question, four things happen in about ten seconds:

1. The question is expanded. The engine does not search your buyer’s words once; it fans out into several rewritten queries. “Best CRM for a small sales team” becomes searches about pricing, about alternatives, about complaints. You are not optimising for one query string; you are optimising for the cloud of queries around an intent.
2. Pages are retrieved and read. This is where the spread between engines is enormous: in our corpus the average number of distinct sites an engine exposes per answer runs from under one to nearly forty. What gets retrieved is decided by ordinary search infrastructure, which is why classic SEO is not dead, it has been demoted to a qualifying round.
3. Brands are lifted from those pages. The engine names companies the retrieved pages describe. It has no opinion of its own to consult beyond what it already believed, and your own site rarely gets to be the witness: the answer leans on third parties saying you exist and what you are for.
4. A shortlist is written, and it is short. Twenty retrieved pages collapse into an answer naming a handful of companies. This compression step is where visibility is won or lost, and it favours brands that are described the same way in many places, because agreement across sources reads as confidence.

One more path matters: when nothing is retrieved at all, the model answers from memory. That is not rare in our corpus, and it is why the last section of this guide is about the memory layer, not just retrieval.

How we measured this

Stated plainly so you can judge it, because most numbers in this field arrive without a method attached.

The questions are not ours. They are built from terms Google Keyword Planner reports real monthly volume for in each category, so they are things buyers type rather than things we would like to be asked.
Asked live, with the web open, to 6 assistants: ChatGPT, Claude, Gemini, Perplexity, DeepSeek and Grok. Not a cached prompt database.
Every answer is kept whole, with the sources the model actually read attached, and published. Nothing here is a summary you have to take on trust.
The numbers recompute as more checks run, so this page cannot drift away from the transcripts behind it. Last recalculated 29 September 2026.

The finding that should change what you do

The industry talks about “optimising for ChatGPT” as though winning there carries you elsewhere. Across 62 questions that more than one model answered, it does not.

22.5

brands named per question, counting every model together.

0.9

of those are named by every model that answered. Fewer than one.

23.3%

average overlap between any two models’ lists of brands.

Ask six assistants the same buying question and they will name roughly 22.5 companies between them, of which under one is common to all of them. Being the answer in ChatGPT tells you almost nothing about Claude, and a tool that checks a single model is measuring a fraction of your exposure.

Watch the disagreement happen

Percentages are easy to nod along to, so here is the thing itself: a real question from our corpus, one full recorded answer, and what every other model named when asked the same thing.

A real recorded answer, not an illustration

“Which AI video generator should I use for creating content?”

from “ai video generator” · 246,000 searches a month · 6 models answered

Grok29 sources · 28 distinct sites

There isn’t one best AI video generator for every kind of content. In 2026 the usual split is: Google Veo 3.1 for quality and sound, Kling for volume on a budget, and Runway when you need real creative control. OpenAI’s Sora app was shut down in 2026, so don’t build a workflow around it.

Pick by what you’re making

What you’re making	Start here	Why
Realistic clips, product shots, cinematic B-roll with sound	Google Veo 3.1 (Gemini / Flow)	Most reviews call it the strongest overall model: realistic motion plus native synced audio
Lots of short social clips on a budget	Kling AI	Best price-to-quality ratio, strong motion, free tier to test, paid plans often start around $7–10/month
Directed shots, consistent characters, editing after generation	Runway (Gen-4 / Gen-4.5)	Camera controls, motion tools, and a real production workspace—not just a prompt box
Talking-head explainers, ads, training, localization	HeyGen or Synthesia	Presenter avatars, not general scene generation
Playful effects, memes, fast social experiments	Pika	Speed and stylized effects over photorealism
Brand-safe commercial work, especially in Adobe	Adobe Firefly	More defensible licensing and fits existing Adobe workflows
Finished social videos from a script (captions, VO, music)	InVideo AI, CapCut, or similar assemblers	They finish a post; pure generators mostly give you short clips

Practical default

Quality-first content (YouTube intros, ads, brand spots): use Veo 3.1 on a Google AI plan (often cited around $20/month). Expect short clips (commonly about 8 seconds, sometimes extendable) and a higher cost per second than budget tools. PCMag currently ranks Veo 3.1 best overall for realism, control, and usable audio.
Regular posting (TikTok, Reels, Shorts): start with Kling. You’ll get more usable seconds per dollar, and the free tier is enough to see if the look fits your channel.
You already edit and care about the shot: use Runway. It’s the tool people keep when the first generation is close but not finished.

CNET’s 2026 testing points the same way: Veo for cinematic quality, Firefly for professional/commercial work, Runway when you want hands-on control.

What to watch before you pay

Commercial rights. Free tiers are often watermarked and non-commercial. Paid plans are what you want for client or monetized content.
Retries. A “$10/month” plan can be expensive if most generations fail. Budget for several attempts per usable clip.
Audio. Veo’s native dialogue and sound is a real advantage. Many other tools still need music, VO, or captions added later.
Length. Most generators still make short shots. Longer videos are usually several clips edited together, or a script-to-video tool that assembles them for you.
Licensing risk. If a client or platform is strict about training data and IP, prefer Firefly or another tool with clearer commercial terms over the cheapest model.

Test two free tiers with the same prompt (one person, one product, one vertical 9:16 shot) before subscribing. The right tool is the one that needs the fewest fixes for your content, not the one that wins a generic leaderboard.

Sources the model used

The same question, the other models, the brands each one named:

PerplexityGoogleRunwayAdobeSynthesiaHeyGenPika+5 more

ChatGPTRunwayLumaGoogleHeyGenSynthesiaD-ID+1 more

ClaudeGoogleByteDanceRunwayHeyGenSynthesiaInVideo+3 more

DeepSeekGoogle VeoAdobe FireflyRunwayKlingLTX StudioHeyGen+7 more

Line the lists up and the overlap problem stops being abstract: same question, same minute, materially different shortlists.

Read the full transcript, every model, every source

The engines do not read alike either

How many sources each one consults before it answers, and how many companies it tends to name. Per-answer averages across our corpus.

Engine	Answers	Sources read	Brands named
Grok	48	54.2	7.5
Perplexity	63	21.8	6.6
ChatGPT	63	6.9	5.9
Claude	63	6.8	7.8
DeepSeek	63	3	4.9
Gemini	63	1.9	7.6

And 14% of all answers cited nothing at all. In those the model is working from what it already believes about your category, which is the hardest state to influence and the strongest argument for being described widely rather than perfectly in one place.

GEO vs SEO, row by row

Most comparisons of these two are vibes. The differences that change what you should do are specific:

	Classic SEO	GEO
What you win	A position on a list. Tenth place still gets clicks.	A mention in one written answer. There is no tenth place.
Where the signal lives	Mostly on and about your own site.	Mostly on pages you do not own, describing you.
Unit of competition	The page.	The brand, as the web describes it.
How stable the result is	Rankings drift over weeks.	Answers regenerate per question, per model, per day.
How you measure	Rank trackers against known keywords.	Recorded answers to real buyer questions, per model.
Where the two overlap	Being crawlable, fast, and clearly written helps both.	The pages that cite you still have to rank somewhere first.

The practical reading: keep the SEO you have, because retrieval runs on it. Then point the new effort at the row that changed most, which is where the signal lives.

What actually moves a brand into an answer

Be described on the sources that category already cites

Citations in our corpus spread across 1,952 distinct domains, and the 20 most-cited hold only 9% of them. There is no short list to buy onto. What works is being accurately described in many places your category is already being read from.

Get into the comparison pages, not just your own site

Roundups and "best X for Y" pages are what an assistant reaches for when asked to name a few companies. Your homepage is the one source in the world that cannot vouch for you.

Say what you sell in the words buyers use

Assistants match a question to a description. A category label that only exists inside your company gives the model nothing to match, however well written it is.

Be reachable by the crawlers that matter

Several assistants use separate agents for training and for live retrieval, and a blanket block at the CDN can quietly remove you from the second. Worth checking before spending anything on content.

Publish something only you can publish

Original figures get quoted, restated ones do not. This page is an example of the tactic it is describing.

Three things this industry keeps repeating

That you can buy a position in a model

There is no ranking, no submission, no placement product inside any of these assistants. Anyone quoting you a guaranteed citation is quoting you something that does not exist.

That llms.txt is a ranking factor

No major assistant has said it reads one, and none of the citations we have collected trace back to a file like this. Publishing it costs nothing. Believing in it costs you the budget you should have spent elsewhere.

That optimising for one model covers the rest

The measured overlap between any two models on the same question is 23.3%. That is the whole rebuttal.

See where you stand, in about a minute

The same check that produced every number on this page, run for your domain. We find the questions your buyers search, ask them live, and show you the unedited answers with your name highlighted if it is in them.

Run the free check

Free, no account, no card.

FAQ

Frequently asked questions

What is generative engine optimization?▾

Generative engine optimization, or GEO, is the practice of getting a brand accurately described on the sources AI assistants read, so that when someone asks about a category the generated answer has a reason to name it. It optimises toward systems that write an answer rather than return a list of links.

Is GEO different from SEO?▾

The work overlaps, the target does not. SEO competes for a position in a list where being tenth still earns a click. A generative engine returns one answer naming a handful of companies, so there is no tenth place. And the signal sits elsewhere: a search engine reads your site to rank your site, while an assistant reads other people’s pages to decide who to mention.

Can anyone guarantee a position in ChatGPT?▾

No. There is no ranking to buy, no submission form, and no placement product inside any of these models. Anyone selling a guaranteed citation is selling something they cannot deliver. What can be changed is the raw material: what the web says about you, on the sources these systems already read.

Does optimising for ChatGPT cover the other models?▾

Our data says no, and this is the most actionable thing we have measured. Asked the same buying question, any two models overlap on only a fraction of the brands they name, and the set of brands every model agrees on is close to empty. A win in one assistant tells you very little about the others.

Does llms.txt help?▾

There is no published evidence that any major assistant reads it, and none of the citations we have collected trace back to one. It costs almost nothing to publish, so it is not harmful, but treating it as a GEO strategy is mistaking a convention nobody has adopted for a ranking factor.

Is GEO worth it for a small brand?▾

It is where small brands have the most asymmetric upside. An engine answering a precisely phrased need reaches for whoever is described well for that need, not for whoever has the largest domain, and our transcripts regularly show small companies named alongside giants for specific questions. The head terms stay brutal; the specific questions are open.

How is GEO measured?▾

By asking and recording. Put a fixed set of real buyer questions to each model on a schedule, keep the full answers, and track who was named, in what order, citing which sources. Averages and scores hide the mechanism; transcripts show it.

Do Google AI Overviews count as GEO?▾

Yes. An AI Overview is a generated answer assembled from retrieved pages, the same mechanism as a chat assistant with a search tool. The difference is distribution: it sits on top of the search results your SEO already targets, which makes the overlap between the two disciplines wider, not narrower.

Does structured data or schema markup improve GEO?▾

It helps machines parse what a page says, which is worth having and cheap. It does not create a reason to mention you. In our corpus the cited pages earn their place by containing something specific, a comparison, a number, a first-hand account, not by their markup.

How long does GEO take to show up?▾

Publication takes days. The models reflecting it is slower and depends on how often they recrawl the sources in your category, which is why the only honest approach is to measure, change the inputs, and measure again rather than quote a timeline.