You wrote a genuinely good page. Clear, accurate, better than half of what ranks above it. Then you ask ChatGPT the exact question your page answers — and it recommends someone else. Names them, even. Not you.
I’ve done this and felt the specific sting of it. It’s worse than a bad Google rank, because at least on Google you can see where you sit. Here there’s no page two to scroll to. You’re either in the answer or you don’t exist.
So the tempting conclusion is that this is a black box, that the model just likes who it likes, and there’s nothing to do but wait.
But that’s not what the data says. A few researchers actually pried the box open, and it turns out the model has tastes — pretty legible ones. So let’s look at what they found, because it changes what you’d put on the page tomorrow.
What The Princeton Team Actually Asked
The study is called “GEO: Generative Engine Optimization”, by Pranjal Aggarwal and colleagues at Princeton and elsewhere, published at KDD ‘24 (a big data-mining conference — the kind of venue where the methods get poked at hard before they’re let in).
GEO (generative engine optimization) is basically SEO for the answer ChatGPT gives instead of the links Google shows. Which, for you, means the thing you’re optimizing isn’t your rank — it’s whether the model mentions you at all, and how prominently, inside its written answer.
The question they set up is clean. Take the same underlying content; present it a bunch of different ways; and see which presentation makes a generative engine more likely to cite it.
To test that at scale they built something called GEO-bench:
- 10,000 real queries, drawn from 9 datasets
- spanning 25 domains (health, law, tech, history, and so on)
- each answer scored for how visible a given source was in it
So this isn’t three prompts and a hunch. It’s a benchmark big enough that a 30% swing means something.
The Three Things That Made The AI Cite You
Here’s the part worth pinning to your wall. The presentation tweaks that worked best lifted a page’s visibility — its odds of being cited or named in the answer — by a relative 30-40%. Same facts, same page, just dressed differently, and the model started pointing at it.
Three moves did the heavy lifting:
- Adding quotations (pull a line from a third party or a primary source) — about +44%
- Adding statistics (put real numbers on the claim) — about +34%
- Citing sources (name and link where the claim comes from) — about +29%
Look at what those three have in common. None of them is a keyword trick. Each one is a way of saying this claim is backed by something outside my own assertion. A quote points at a person who said it. A statistic points at a measurement. A citation points at a document.
The model, in other words, rewards content that shows its work. And I think that’s the whole story in miniature — the thing being scored isn’t how many times you said “best CRM,” it’s how much evidence you brought.
The nice part is that all three are things you can just do. You don’t need to guess the model’s mood. You add a sourced quote to the paragraph, you swap “many users” for the actual number, you link the study instead of gesturing at it. That’s a Tuesday afternoon, not a replatforming.
Why The Old Trick Backfires
Now the move that people reach for out of habit: keyword stuffing. Cram the target phrase in a few more times, the way SEO rewarded for years.
In the study, that one didn’t just fail to help. It dropped visibility by about 8%.
Which means the old game and the new game pull in opposite directions. Stuffing works on a system that matches strings. A generative engine isn’t matching strings — it’s reading for meaning, deciding which sources it trusts enough to repeat, then writing a fresh answer. Say the keyword ten times and you’ve added no evidence and a faint whiff of spam. The model reads meaning and evidence, not keyword density, so the tactic that used to buy you rank now quietly costs you the mention.
I’d hold this one loosely — it’s a single study, and I could be over-reading one number. But it lines up with everything else here. The three winners add evidence; the one loser adds repetition. That’s consistent enough that I’d stop stuffing today.
One honest caveat the authors are clear about: the size of these lifts varies by domain. Evidence cues help more in some verticals than others — a legal query and a lifestyle query don’t respond identically. But the direction holds across the board. Which is the point: adding a real source rarely hurts, and where it helps, it helps a lot.
Conclusion: Add The Evidence, Then Watch Whether It Lands
So here’s where I’ve landed. If you want a generative AI to name you, stop optimizing for the crawler and start writing like a source worth quoting — pull in a real quote, put the actual number on the claim, and link where it came from. Those three moves bought a 30-40% lift in the study, and they’re all edits you can ship this week.
The one thing I wouldn’t do is add the cues and assume it worked. Domain variance is real, models change, and “I added quotes” isn’t the same as “the model now mentions me.” That gap is the thing you have to watch, not guess at — ask the models your real questions on a schedule and see whether your name shows up more this month than last. That’s the whole reason the tool I work on, HexScope, exists: to tell you whether the evidence you added actually moved the needle, instead of leaving you to squint at ChatGPT and hope.
Good content was always the price of entry. Now the receipts — the quote, the stat, the source — are what get you read back.
Sources
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ‘24). https://arxiv.org/abs/2311.09735
