“AI doesn't cite the best writer. It cites whoever makes it easiest. And once you see that mechanism, you can read it the same way we've read search engines for years.”
An AI Overview isn't a black box: it's a search engine with rules
When someone asks ChatGPT, Gemini or Google's AI Overview something, they're not reading an improvised opinion. They're reading the output of a process with fixed steps: the question gets split into several smaller ones, a regular old search engine brings back a handful of candidate pages, and the AI copies the passage from one of them it can defend without risk. Google itself already confirmed it: optimizing for generative AI is still optimizing for the search engine.
People call this work different things depending on who you ask: GEO (Generative Engine Optimization) if the focus is the engine generating the answer, AEO (Answer Engine Optimization) if it's the answer itself. We use both names for the same thing — there's no term police, just a mechanism worth understanding.
What's strange is that every answer sounds different. Ask the same thing twice and the wording changes, the order changes, even the example changes. That's why a lot of people assume there's nothing to be done: that it's random, a lottery. It isn't. The wording varies; the list of sources AI actually trusts almost never does.
How AI cites you
Your question travels through a circuit — and it can be read
Behind every AI answer there's still a search engine. Tap each stage to see what happens — and what we do at it.
Every answer sounds different, but the sources AI actually trusts barely change. That circuit can be read — and our job is to get your page into it. A system, yes; autopilot, no.
Why the randomness is misleading
What varies is how the answer gets told. What doesn't vary is where it comes from. Run the same search thirty times, on a different day, on mobile and on desktop, and you'll see the same two or three pages show up almost every time. That's what actually matters: the closed set of sources it never strays from. The exact sentence the model spits out is beside the point.
The pattern behind the randomness
Thirty attempts, almost the same result
Each square is one time we repeated the same question, at a different time and on a different device. In red, the source that came up cited almost every time. The rest is surface noise: the wording changes, the underlying source doesn't.
- Answers before explaining92
- Stands alone without the rest of the page88
- Matches the answer's own format74
- Agrees with the other cited sources69
- Has a fact nobody else has41
Illustration based on the pattern observed across hundreds of AI answers. Source: internal Volcanz analysis.
Look at the order: what weighs most isn't having the exclusive fact. It's how easy you are to quote without risk. A unique fact buried in a long, tangled paragraph loses to an ordinary fact placed up front and clearly bounded. First you need to be easy to cite. Only after that does differentiation matter.
The formula: match the consensus, then beat it
This is the mistake almost everyone makes, in both directions. Publishing something too different breaks the consensus: the AI doesn't recognize your page as a valid answer to the question and it never even makes the shortlist. Publishing a clone of what already exists doesn't work either: you give the AI zero reason to drop the source it already trusts and cite you instead.
The way out isn't picking a side. It's doing both, in the right order: first you look like family (same format, same terms, same tone as the sources that already work), and only once you're a candidate do you add the fact, angle or figure nobody else put on the table.
The formula
Match the consensus, then beat it
AI only cites what it recognizes. And it only picks you if you give it a reason. Tap each layer — and try removing your unique fact.
Match — makes you a candidate
Beat — gets you picked
The four layers below copy what AI already rewards: that makes you a candidate. The fifth is yours — the reason to cite you instead of the usual source. And before publishing, a person reviews what matters. A system, yes; autopilot, no.
Same mechanism, every engine
ChatGPT, Gemini, Claude and Perplexity don't share a database, but they share the same underlying logic: gather candidates and cite the most defensible one. What changes is where each one pulls those candidates from.
| Engine | Where it pulls candidates from | What that means for you |
|---|---|---|
| Google / AI Overviews | Its own search index | If you're not well ranked in Google, you don't even make the shortlist |
| ChatGPT | Bing plus its own trained knowledge | Worth being present on Bing too, not just Google |
| Claude | Its own index plus third-party sources | Reputation off your own site (mentions, press) weighs more |
| Perplexity | Its own real-time crawl | Rewards freshness: recent, clearly dated content |
There's no need to chase all five engines with five different strategies. One well-built page — that answers first, stands on its own, agrees with the good sources and adds a fact of its own — is a candidate on all of them at once. Each engine makes you earn that reputation somewhere different: that's what varies, not the formula.
Same mechanism, every engine
One well-built page, five storefronts
There's no need to write the same content five times. One page that answers first, stands on its own and adds a fact of its own is a candidate on all five engines at once; the groundwork to make each shortlist is what differs.
How we work it
- We capture the same question many times, under different conditions, and log which sources repeat every time and which only show up sometimes.
- From that we work out the format, terms and tone the AI expects for that question — the consensus to match.
- We write the content answering first, in paragraphs that stand on their own without the rest of the page, and add the fact or angle missing from the conversation.
- Before publishing, a person reviews the highest-priority pieces by hand: the system finds the pattern, but the final call isn't automatic.
- We measure whether it starts getting cited and on which engines, not just whether it climbs in Google — related, but not the same thing.
How we work it
A cycle that repeats, not a project that ships once
Each turn of the cycle feeds the next: what we measure in step 5 sharpens what we capture in step 1. That's why it's a system, not a one-off campaign.
None of this is some new three-letter discipline. It's the same question good SEOs have been asking for twenty years — why does the winner win, and how do I win in their place? — applied to a search engine that now also writes the answer for you.
Frequently asked questions
Does this replace traditional SEO?+
No, it needs it. Ranking well in the traditional search engine is still, for most engines, the entry filter into the candidate list. GEO doesn't replace SEO: it builds on it and adds what's needed to go from being on the list to being cited.
Why is my brand invisible in AI search?+
Almost always one of two reasons: your site doesn't answer the real questions people ask AI about your category (price, comparisons, reviews), or nobody else talks about you — without mentions elsewhere, AI has nothing to corroborate you with. Writing more fixes neither. Filling the gap and earning presence beyond your own site does.
Why does my competitor show up sometimes and not others, for the same question?+
Because what varies between answers is the wording, not the pool of candidate sources. If your competitor shows up often, they're probably inside that stable pool. The question that matters isn't "why did they show up this time", it's "am I in that pool at all, even some of the time".
How long until you start getting cited?+
Depends on whether you're already well ranked in search or that foundation needs work first. With a solid base, first citations can land in weeks. Without it, you need to earn that starting position first — and that takes longer.
Do I need to be on all five AI engines at once?+
Not always. It depends on where your customer searches. But the good news is one well-built piece is a candidate on almost all of them at once — what changes per engine is the prior reputation needed to make its shortlist.
Can an automated system do all of this alone?+
It can do the capture and pattern part: that's exactly what gets automated. But deciding what gets published, especially for clients or competitive topics, always goes through human review before it ships. A system, yes; autopilot, no.
We'll help your site make the list of sources AI cites — with a system, not luck.
Talk to us