Why this is not the same as SEO
Classic search optimisation aims at a ranking: be higher up the results list. There is no ranking in an AI answer — there is a piece of text in which a few sources get mentioned and the rest do not. So the question is not what position you are in, it is whether you get in at all.
The process has two steps. First the system pulls together a handful of documents, typically by searching the classic index. Then the model writes the answer out of them and decides which ones it mentions. Those two steps are two different jobs, and most advice conflates them: what helps with the pulling together does not necessarily help with the mention.
The number everyone quotes — and what it actually measures
There is a famous research result according to which certain text modifications can raise visibility in AI answers "by as much as forty percent". That number has been quoted ever since, usually as though it meant forty percent more traffic or more mentions.
What it actually measures: the position-weighted word count of the text inside the answer, in a setup where the document had already been loaded in front of the model. So it does not measure whether you get in, it measures how much gets quoted from you once you are already in. Not clicks, not traffic, and not the chance of the system finding you. That is not the result's fault — it is the fault of whoever passes it on without reading what was measured.
What the controlled studies refute
A peer-reviewed comparative study from 2025 examined dozens of recommended methods across several topic areas. The lessons are uncomfortable.
- Out of fifty-four method-topic combinations, exactly three came out significantly positive. The rest were ineffective or harmful.
- The "sprinkle statistics into the text" advice made things worse in nineteen setups out of twenty-four.
- The question-and-answer block put into an article scored minus 5.7 percent on getting into the answer, in a study across six hundred controlled questions. Meanwhile one large provider's official guide recommends this very trick — without evidence.
- Formatting on its own has no detectable, consistent effect. Bullet lists, short paragraphs and bolding are good for the person reading, not for the machine.
There is one more lesson that frames all of it: if everyone applies the same trick, the advantage melts away. These methods work against each other, they do not add up.
What the same work confirms
That same line of research did identify four factors firmly — across a quarter of a million trials, on six models, with position bias filtered out.
- Topic match. The document should genuinely be about what was asked. This sounds trivial, and it is still the strongest factor.
- A concrete price. If the question carries purchase intent, a written-out price raises the odds of inclusion dramatically. "Contact us for a quote" does not.
- A fresh timestamp. Whether a date is there and how recent it is counts heavily — the next section explains why that is uncomfortable.
- Your order in front of the model. Which of the loaded documents you happen to be. You have no direct influence over this, but it explains why the results swing so much.
The uncomfortable explanation for the fresh date
The date effect is real, but not for the reason we would like. In one study the researchers artificially altered the dates on documents while leaving the content untouched. The model's ranking moved by ninety-five positions, and in a quarter of the cases the preference flipped.
That means the date is not a quality signal but a bias: the model favours text that looks fresher, even when nothing inside it has changed. Two things follow for you. One is that genuine updating really does pay. The other is that rewriting the date without touching the content may work in the short run, but in the long run it is exactly the kind of trick systems eventually filter out — and it will not make your reader any happier either.
The overlap nobody agrees on
A common question is whether doing well in the classic search engine is enough. The answer: we do not know precisely. Three serious measurements exist of the overlap between traditional results and the sources of AI answers, and they contradict each other: 38, 56 and 10-30 percent came out of them, with different methods.
What does follow with certainty is that the overlap is not complete — a good classic position is necessary but not sufficient. A good illustration of how large the uncertainty is: running the same model in two different thinking modes, a quarter of the cited domains were shared. Not two different providers: the same model, two modes.
What is actually worth doing
1. HAVE CONTENT THAT ANSWERS THE QUESTION.
Not "about the topic" — an answer to the concrete question.
This is the strongest proven factor.
2. WRITE OUT THE SPECIFICS. Price, number, deadline, size.
This is the second strongest, and most sites leave it out.
3. UPDATE FOR REAL, and show a visible date.
Do not rewrite the date: rewrite the content.
4. DO NOT BUILD ON TRICKS. Stacked statistics and the Q&A block
MADE RESULTS WORSE in the studies. Formatting is for the reader.
5. CHECK WHETHER THE BOT CAN GET IN AT ALL.
This is the one step that is decisive and not probabilistic.
Common misconceptions
- "There is a proven GEO methodology." There is not. Most of the methods examined were ineffective or harmful in a peer-reviewed study.
- "The forty percent means more traffic." No: it measures word count inside the answer, with the documents already loaded.
- "Put a FAQ block at the bottom of the article." An official guide recommends it too, but the measurement says it lowers your chance of getting into the answer.
- "If I rank well on Google, AI will cite me too." Measurements of the overlap scatter between 10 and 56 percent. It helps, but it is not enough.
What to take away
Getting cited comes down to two things you can influence: that the machine can reach your site, and that the content gives a concrete answer to a concrete question, with concrete numbers. Everything else — the formatting tricks, the inserted statistics, the question-and-answer blocks — is either ineffective or measurably harmful. On the technical side of access the machine readability post goes bot by bot, on measuring it the tracking your brand visibility post gives the metrics, and on the traffic background the zero-click search post has the numbers. The copywriting side is what the writing for humans and AI post walks through, with examples.