Generative Engine Optimization (GEO): What It Is, and What Held Up When We Measured It
On this page
TL;DR
TL;DR: Generative engine optimization (GEO) is the work of getting your brand cited inside an AI answer rather than ranked in a list of links. We put the six recommendations that appear in almost every GEO guide through 48 measured runs across ChatGPT, Perplexity, Gemini and Google's AI Overview. Not one of them survived unchanged. Every count is published below.
What generative engine optimization is
Generative engine optimization is the practice of shaping your content and your public footprint so that AI systems name you when they answer a question.
The distinction that matters is not technical, it is structural. A search engine returns a list and lets the reader choose. A generative engine returns one answer and chooses on the reader's behalf, then cites a handful of sources underneath. There is no second page. If your brand is not inside the answer, the reader never learns you exist.
That single difference is why the tactics diverge. Ranking is about being one of ten acceptable options. Citation is about being the one the model reaches for. We cover the split in more depth in GEO vs SEO.
You will see the same idea sold under other names: answer engine optimization (AEO), AI search optimization, LLM optimization, AI visibility. As of early 2026 there is no agreed technical distinction between them, and in practice they describe the same work.
Why the advice is unreliable right now
The field is roughly two years old. Almost every guide you will read, including the ones ranking above this page, is a synthesis of other people's claims rather than anything the author measured. The current second result for this term runs past 5,000 words and cites six external sources for its statistics without collecting any data of its own.
That is not dishonest, it is just early. But it means the standard advice has been copied between guides many times without anyone checking whether it holds. So we checked.
How we measured this
Everything below comes from one dataset, collected by hand.
| Questions | 12 buyer questions a real B2B SaaS customer would ask |
| Engines | ChatGPT, Perplexity, Gemini, Google's AI Overview |
| Runs | 48 (12 questions × 4 engines, one run each) |
| Collected | 6–7 August 2026 |
| Market | United States, English |
| Recorded | Every brand named, in order, and every visible source, tagged by type |
| Total | 322 citations across 193 distinct sources |
| Licence | CC BY 4.0, raw JSON and CSV published |
The questions covered four intents: recommendation ("best CRM for startups"), comparison ("HubSpot vs Salesforce"), how-to ("how to choose a CRM") and alternatives ("cheaper alternatives to HubSpot"). A June round using six of the same questions gives us an eight-week comparison.
The limits, before the findings
These belong at the top, not buried at the bottom, because they change how much weight each number below can carry.
- One run per engine-question pair. A single run cannot separate a stable pattern from a lucky draw. AI answers vary between runs. Treat every count as directional.
- Source counts are a lower bound. ChatGPT and Google's AI Overview hide some sources behind a "+N" badge. We recorded only what was visible, so per-engine totals are not strictly comparable.
- Seven of the 48 runs showed no sources at all. That is a finding in itself, covered below, but it also means averages need reading carefully.
- One sector, one market, one moment. B2B software, United States, English, one week in August. None of it transfers to consumer, local, healthcare or non-English queries without measuring again.
- Twenty citations were names, not links. Some engines named a publisher without giving a resolvable address. Fifteen of those twenty came from Gemini alone, which makes Gemini's profile the least reliable in the set.
The six recommendations, measured
1. "Write comparison pages, that is how AI finds you"
This is the single most repeated piece of GEO advice, and it is the one that broke hardest.
On all three "X vs Y" questions, ChatGPT cited nothing at all. Not a thin list of sources. Zero. It answered from memory. It did the same on both how-to questions. But on every "best X" question it cited between four and nine sources.
So ChatGPT is not a reluctant retriever. It retrieves heavily. It simply does not bother going to look when you ask it to compare two products it already holds an opinion about.
This is specific to ChatGPT, and that matters. On those same comparison questions Perplexity and Google's AI Overview both retrieved normally, and Gemini retrieved on two of the three.
What to do with that: comparison pages still earn citations in Perplexity and AI Overviews. But if ChatGPT is the engine you care about, a comparison page is not the lever. Its "best X" and "alternatives to X" answers are where it actually goes looking.
There is a second, sharper version of this. On "Notion vs Asana," neither notion.so nor asana.com was cited by any engine. On "HubSpot vs Salesforce," each vendor's own site was cited by Perplexity alone. Being the subject of the comparison does not get you cited on it.
2. "Get cited in one engine and you will show up in the others"
Across 193 distinct cited sources, 78.2% were cited by only one engine. Four sources, or 2.1%, were cited by all four.
Narrow it to same-question agreement and it gets starker. In 48 runs there are exactly two cases of all four engines citing the same source for the same question, and both fall on the same question: zapier.com and project-management.com on "best project management software for small teams."
| Cited by | Sources | Share |
|---|---|---|
| All four engines | 4 | 2.1% |
| Exactly three | 10 | 5.2% |
| Exactly two | 28 | 14.5% |
| Only one engine | 151 | 78.2% |
Set against the brand numbers from the same runs, the picture inverts: the four engines share just 1% of their sources but agree on 32% of the brands they recommend. Same shortlist, completely different reading.
What to do with that: there is no single list to get on. Cross-engine agreement is a property of the question, not of the source. Almost nothing earns durable trust everywhere; a page earns a slot for one query in one engine. Track engines separately or you will misread a win in one as coverage everywhere. Our guides on ranking in ChatGPT and ranking in Perplexity are separate for exactly this reason.
3. "A citation is an asset, it stays put"
We ran six of the same questions in June and again in August. Between the two rounds, 54% of the sources Perplexity had cited were gone.
A backlink sits on a page until someone removes it. A citation is regenerated on every query, from a retrieval set that changes underneath you. Eight weeks is enough to lose half of it.
What to do with that: stop treating a citation as a one-off win to log. It is a position that has to be held. If you are not re-measuring, you do not know whether you still have it. We wrote up how to do that in how to measure AI search visibility.
4. "Get listed on the big review sites"
The standard playbook says pour effort into G2, Gartner, Forbes and the established review press. Here is where the 322 citations actually went:
| Source type | Citations | Share |
|---|---|---|
| The vendor's own website | 156 | 48.4% |
| Independent listicle or niche blog | 85 | 26.4% |
| Review or authority media | 50 | 15.5% |
| YouTube | 16 | 5.0% |
| 14 | 4.3% | |
| Other forum | 1 | 0.3% |
Nearly half of every citation went to a vendor's own site. Small independent blogs were cited 70% more often than the established review media that most software marketing budgets chase.
The split by engine is wider still:
| Type | ChatGPT | Perplexity | Gemini | AI Overview |
|---|---|---|---|---|
| Vendor's own site | 67% | 46% | 43% | 46% |
| Independent blog | 16% | 28% | 43% | 23% |
| Review media | 9% | 21% | 11% | 12% |
| Reddit + YouTube | 7% | 6% | 3% | 18% |
Four genuinely different diets. ChatGPT reads the vendor. Gemini reads small blogs. Perplexity reads broadly across editorial. Google's AI Overview is the only one with a real appetite for Reddit and YouTube.
What to do with that: the advice is not wrong, it is mis-ordered. Your own site is the single largest citation surface in the set and it is the one you fully control. Review listings are worth having, but they are the third priority, not the first.
5. "Block the AI crawlers to protect your content"
BuzzStream studied this directly in April 2026. Among the top 50 news publishers that block GPTBot, 88.2% still appeared in AI citations, in a dataset of 4 million citations drawn from 3,600 prompts.
Two limits on that number, stated plainly: it covers the top 50 news publishers rather than websites generally, and GPTBot is the training crawler, not the live retrieval bot. It is not a claim about your site specifically.
What it does show is that blocking does not do what people expect. The model keeps drawing on what it captured before the block. You lose the visits and keep the citation, in a version of yourself that gets steadily more out of date.
What to do with that: decide crawler policy on your own terms, not on the belief that blocking removes you from AI answers. If you want to check what your own robots.txt currently allows, our AI Bot Checker reads the live file, and allowing AI crawlers in robots.txt walks through the rules.
6. "Build backlinks"
This one we did not measure ourselves, and the strongest available evidence is Ahrefs' study of 75,000 brands. It found unlinked brand mentions correlate with AI visibility at 0.664, and backlinks at 0.218.
Correlation is not causation and a single study is not a law. But the gap is large enough to change where effort goes: being talked about appears to matter roughly three times more than being linked to.
What to do with that: a mention in a forum thread, a podcast, a comparison post or a Reddit answer is worth pursuing even when it carries no link at all. We put together seven strategies for improving brand visibility in AI search around this.
What each engine actually does
"Optimize for AI" is not one job, because the four engines do not behave alike. These are the same 48 runs, split by engine.
| ChatGPT | Perplexity | Gemini | AI Overview | |
|---|---|---|---|---|
| Total citations | 43 | 138 | 37 | 104 |
| Answers that showed any source | 7 / 12 | 12 / 12 | 10 / 12 | 12 / 12 |
| Average sources per answer | 3.6 | 11.5 | 3.1 | 8.7 |
| Most in a single answer | 9 | 19 | 6 | 20 |
| Distinct sources drawn on | 35 | 106 | 35 | 77 |
| Share of its sources no other engine used | 54% | 64% | 63% | 55% |
ChatGPT shows its work least often
It displayed sources on only seven of twelve answers. When it did retrieve, it averaged 6.1 sources; across all twelve answers that drops to 3.6. Both numbers are true and quoting only one is misleading, so we publish both. Two thirds of what it did cite was the vendor's own website.
Perplexity reads the most and repeats itself least
It sourced every answer, averaged 11.5 citations, and drew on 106 distinct sources, nearly three times ChatGPT's range. It is also the only engine that showed a repeat favourite: emailvendorselection.com was cited on four separate questions by Perplexity and by no other engine at all.
Gemini cites least and least legibly
Thirty-seven citations, the fewest of the four. More importantly, fifteen of the twenty unresolvable name-only citations in the whole study came from Gemini. It names publishers without linking them, which makes its behaviour the hardest of the four to verify or act on.
Google's AI Overview is the one that reads forums
Reddit and YouTube made up 18% of its citations. No other engine came close, and only Perplexity cited YouTube at all. If community presence is part of your plan, this is the engine it pays off in.
What we could not test
Being straight about the holes is the only thing that makes the rest usable.
- Whether any of this causes anything. We measured what engines cited, not why. A page cited on Monday may be cited for reasons we cannot see from the outside.
- Whether the pattern holds outside B2B software. We have one sector. A consumer or healthcare question set could behave completely differently.
- Whether one run represents the engine. It does not, and we say so above. Repeating each pair five or ten times is the obvious next round, and it costs a day per round.
- Anything about paid placement or model training deals. Invisible from where we sit.
- Our own predictions. In an earlier round we thought we had found a pattern in which kinds of sites get cited. We tested it against more data and it collapsed. We published that too, because a guide that only reports its wins is not measuring anything.
Where to start
In the order the data supports, not the order the guides usually give:
- Make sure the engines can reach you. Everything else is theoretical if your robots.txt blocks retrieval. Check it once, properly.
- Treat your own site as the primary citation surface. It took 48.4% of citations in our set and it is the only one you control outright.
- Pick the engine that matters to your buyers and measure that one. Coverage does not transfer. A win in Perplexity tells you very little about ChatGPT.
- Pursue mentions, not only links. Three times the correlation, on the best evidence available.
- Re-measure on a schedule. Half of one engine's sources turned over in eight weeks. A number you collected in spring is not a number any more.
- Write the pages the engines actually retrieve for. In our set that meant "best X" and "alternatives to X" far more than "X vs Y."
The raw data
All 48 runs are published as JSON and CSV under CC BY 4.0, including every brand in the order it appeared and every source with its type code. If a count above looks wrong, the file is there to check it against.
If you find an error in it, tell us and we will correct the page and say what changed.