5M ChatGPT prompts analyzed. Here's what matters for GEO:

Jacky from Link Garden came on the MVP podcast with Paul (my co-founder) and Antoine.

He scrapes 1 million prompts on ChatGPT every month. Same prompts, every month, for 5 months. 5 million answers, every brand extracted, every source logged.

Most GEO advice is opinion. This is the closest thing to data I've seen.

Oh and before we dive in, If you're wondering how we're doing our SEO/GEO today:

You can try ChatSEO for free

(no credit card - set up in 10 seconds)

— "Using our own product - SEO/GEO results"

Here's what I took from it:

1. The Reddit play is dying

Late 2025: Reddit was in roughly 80% of ChatGPT answers. 4 months ago (France): 30% citation rate. July: 9%. August: around 5% and still dropping.

In some verticals it never worked. Car insurance: Reddit sits at 3% and hasn't moved in 5 months.

The pattern is simple. Any source that gets too effective gets spammed by SEOs, and the model dials it down. If your GEO strategy is "post on Reddit", you're optimizing for December 2025.

2. Backlinks don't predict citations

From his data: a site with 5 referring domains cited 200 times. Next to it, a site with 200 referring domains cited zero times.

We have an agent at ChatSEO that can get you these backlinks for free, it's what we call "Ninja linking".

Just copy-paste this prompt in ChatSEO:

"Need you to find free backlinks placement for me"

Domain authority is a Google signal. The LLMs don't look at it the same way. Small niche sites get pulled in constantly because they're precise, not because they're big.

3. Sentiment tracking is overrated

He scores millions of named entities from -9 to +9. The average is around 2. A "good" score is 3 or 4. A "bad" score is -1.

The models are consensual by design. Almost nothing is negative. Conor McGregor scores lower than Hitler in his data, which tells you how weird the scale is.

Don't buy a tool for the sentiment score. Buy it for the citation tracking.

4. The €2,000 citation and the €20 citation do the same thing

Citations rotate. A source shows up in March, disappears in April, comes back in June. Around 50 sources rotate on the same prompt, 10 to 15 appear on any given answer.

That means a €1,500 mention on a big media site lasts 1 to 3 months, same as a €20 mention on a small relevant site. The big media premium is mostly paying for a logo.

5. Why citations matter more than you think: the 18 month gap

Two kinds of answers. With sources (the model searched the web) and without sources (the model answered from training data).

If you launched a brand in January 2026, you are not in any training set. It takes 1 to 2 years to get in, 18 months on average. Until then, you only exist when the model decides to search. Some months that's 90% of queries. Some months it's 30%.

Citations today are how you enter the dataset of 2027. That's the real ROI, not this month's answer.

6. Nobody ranks for the prompt. You rank for the fan-out.

This is the one that changed how I think about GEO keywords.

When someone writes a long prompt ("I did judo, hurt my back, saw a doctor, still can't sleep, what do I do"), the model doesn't search for that. It rewrites the prompt into 2 to 5 short sub-queries. Here: "best mattress back pain".

Judo doesn't matter. The doctor doesn't matter. 30 different people write 30 different prompts and the model collapses them all into the same 3 searches.

So "we can't do keyword research for GEO because every prompt is unique" is wrong. The prompts are unique. The fan-outs are not.

What the fan-outs looked like for about a year: "best + [topic] + 2026 + [country]", generated in the user's language and in English.

Copy-paste this prompt (in ChatSEO) to find your GEO queries:

"Find all the GEO queries for my business and tell me which ones I should target today"

What I'd do with this

Stop tracking prompts. Start tracking fan-outs. Your GEO keyword list is the set of short queries the model actually runs, and that list is a lot closer to classic SEO than people want to admit.

That's exactly what we built the GEO agent in ChatSEO to do: take your site, work out which fan-out queries the models run in your topic, and give you the list to go rank for. Because in the end the model still opens Google.

Do your SEO well, get cited on small relevant sites, and re-check the fan-outs monthly. That's the whole game right now.