How I measure Minecraft.fr’s visibility in LLMs ?

On 11 May, I started tracking Minecraft.fr in Search Ai with 399 questions that players actually ask, and then I waited. One snapshot every two weeks, across ChatGPT, Gemini and Claude. Ten snapshots later, I can answer a question everyone is asking but almost nobody measures on their own site: what do AI engines actually do with my content?

I've run this site for fifteen years. It's the largest French-language site in its niche and it ranks very well on Google, which is exactly what makes it a good test case: if SEO carried over to AI automatically, this is where you'd see it.

Search Ai is the GEO tracking tool we build at Pixalione, with Antonia Giarikani leading product. The site being tracked is mine, not a client's. The data covers 11 May to 14 September, and I've kept in the readings I had to correct along the way.

The first surprise comes down to a subtraction

Minecraft.fr sits in Google's top 3 for 303 of those 399 questions. ChatGPT picks it up for just 139.

On 54% of the topics where I lead in classic search, ChatGPT doesn't cite my site at all. Gemini and Claude do better, citing it 73% and 71% of the time, but the gap is still huge. And when I calculate the correlation between my Google position and my position in the AI answer, I get 0.11 on ChatGPT, 0.16 on Gemini and 0.20 on Claude. Next to nothing, three times over.

That number drives everything else. If your Google ranking doesn't predict whether an AI will cite you, you can't infer it from your SEO. There's only one way to know where you stand: measure it.

It's also why I don't believe in GEO audits built on a simple Search Console export. They describe one channel using another channel's data.

Three AIs, three different sites

As of 14 September, on the same 399 questions: ChatGPT cites me on 158 of them (39.6%) at an average position of 1.52. Gemini on 260 (65.2%) at 2.96. Claude on 253 (63.4%) at 3.19.

So ChatGPT is both the stingiest and the most generous. It picks me half as often as Gemini, but when it does, I come first: 113 of my questions show me in position 1.

The explanation is mechanical, and the data bears it out. ChatGPT pulls in 3.5 sources per answer on average, Gemini 5.8, Claude 8.2. The more sources an AI cites, the less weight each citation carries in what the user actually reads.

The practical takeaway: a dashboard that adds up all three engines is meaningless. A spot in a ChatGPT answer and a spot in a Claude answer are not interchangeable.

The mistake I made for a month

At first, I watched ChatGPT and barely glanced at the rest. It's the most widely used AI, so it seemed reasonable, and tracking three curves instead of one took three times as long for what I assumed would be the same insight.

The day I cross-referenced all three, here's what I found. Out of 399 questions, only 111 cite me on all three AIs. 112 on two. 114 on just one. And 62 on none.

Broken down: 56 questions where I only exist on Gemini, 49 only on Claude, 9 only on ChatGPT.

In other words, by following a single AI I was wrong about 72% of my scope. And I was wrong both ways: assuming I was absent where I actually showed up elsewhere, and assuming I'd secured visibility that only held on one engine.

What my overall rate was hiding

For the first few weeks, my KPI was a single percentage: "40% presence on ChatGPT". The kind of number that reassures you and teaches you nothing.

My 399 questions are grouped into twelve categories and mapped to four player profiles. The day I looked at the breakdown, that 40% split in two.

On one side: Potions at 75%, Brand at 70%, Enchantments at 55%, Recipes and items at 53%. On the other: Tech and installation at 17% across 60 tracked questions, Multiplayer and servers at 12%, Culture and meta at 8%.

The profile view tells the same story. A player looking for a survival tutorial finds me in one answer out of two. Someone installing mods, one in five. A server admin, one in nine.

As far as AI is concerned, I'm a survival tutorial site. Nothing more. And it's not a writing or optimisation issue: it's an editorial choice made years ago that the data suddenly makes visible. On those topics, players have always gone to CurseForge, Modrinth and YouTube, and the models have learned to do the same.

An overall rate is just an average of what works and what you don't do.

The drop that wasn't

In late August, my position-1 questions on Claude fell from 108 on 17 August to 40 on 31 August. Two weeks, a 63% loss on the most visible metric on the dashboard.

The first instinct is to look for what you broke. Except I hadn't touched the site during that fortnight: no content, no technical changes, no internal linking. And my overall presence on Claude barely moved over the same period, from 65.7% to 64.4%. I was still cited just as often, simply less often in first place.

It wasn't my site that changed, it was how the model ranks its sources. So I treat it as a model shift, and I don't allow myself to draw any conclusion until it's confirmed over two consecutive snapshots.

When a number collapses, the first question isn't what you did. It's what changed on the other side.

Am I looking at the model, or at my own data collection?

Two points on my curves made me jump before I understood them. On 25 May, Gemini dropped to 28% presence, after sitting at 66% the fortnight before and 68% the one after. On 20 July, ChatGPT had no data at all.

No model behaves like that. These were partial collections, not drops. An AI that falls to zero overnight and bounces back to its usual level two weeks later tells you nothing about the AI. It tells you something about your instrument.

So I excluded them, and the line on the chart is broken rather than smoothed. It's now a rule I apply every time: an isolated point that contradicts both its neighbours is suspect until proven otherwise, and the proof isn't in the data, it's in the collection logs.

When you take the measurements yourself, part of what you observe is produced by the tool doing the observing.

Citations and mentions: two things I can't measure the same way

A citation is a URL from my site sourced in the answer. A mention is the brand name written in the text, with no link. The first can drive a click; the second builds authority with the model and is what gets you cited from memory when the question doesn't name you. They don't grow together and they aren't won in the same place: citations are earned on the page, mentions are earned off-site.

On Minecraft.fr, I can only track the first. My domain name is also the name of the game, so "Minecraft" appears in almost every answer in my scope, whichever site is cited. A brand mention makes no difference here: across the 399 questions, the string "minecraft.fr" is spelled out in only around fifteen answers.

That's a limitation of my case, not of the metric. For a retailer or a school, whose name has no reason to come up spontaneously in an answer about their field, the indicator is perfectly readable and it's actually one of the most useful. In my case it isn't, so I don't comment on it.

Before you steer by a metric, make sure it's capable of saying no.

You're not up against your competitors

The domains cited most often alongside mine tell the story better than any commentary.

On ChatGPT, the order is what I expected: minecraft.fr 322 times, minecraft-france.fr 155, YouTube 150, Fandom 96. On Gemini, YouTube takes the lead with 302 citations to my 260. On Claude, YouTube is cited 401 times, more often than there are questions in my scope, ahead of my 317.

In an AI answer, a publisher isn't competing with the other publishers in its sector. It's competing with YouTube, Reddit and wikis. That's what the Ahrefs data I published in August on AI Overviews showed, and I'm seeing it again here across 399 questions and three engines. The same lesson in two different places is starting to look like a rule.

Two details I wouldn't have spotted anywhere else. Search Ai also surfaces third-party sites in my ecosystem and flags the ones blocking AI crawlers: digminecraft shuts out GPTBot, ChatGPT-User and Perplexity, while X blocks OAI-FeedFetcher, Perplexity-User and Claude-User. Sites cutting themselves off from part of the answers, and leaving room for me. And eight new domains appeared in my ecosystem in the latest snapshot alone: monitoring I no longer have to do by hand.

What the models say when you ask them about the brand

Sentiment tracking is part of what we review at every snapshot at Pixalione: what a model says about a brand matters as much as the pages it cites. For minecraft.fr, 22 topics come up: 13 positive, 6 negative and 3 where the model declined to take a position.

On the plus side: up-to-date content, French-language, an active community, a 4.5 out of 5 rating from 42 votes, a trust score of 98 out of 100. On the minus side: "interface that may look outdated", "reliance on the community", "limited resources", "limited reach".

I never wrote any of that. The models inferred it from what's said about my site elsewhere on the web. Fifteen years online leave LLMs with the impression of a dated interface, and that reputation can't be fixed with content. It's fixed with a redesign (which I did a few weeks ago!) and with recent editorial mentions.

It's the only lesson from these four months that has nothing to do with SEO.

What I haven't won

I have no proof that any of this drives traffic. That's the honest limit of the exercise, and I'd rather say it than leave you to guess.

A position in an AI answer isn't a Google ranking. There's no results page, just an order of appearance in an answer the user reads without necessarily clicking. Search Ai tells me that I'm cited and at what rank; it doesn't tell me how many people came to the site. I'm measuring awareness within a model, not acquisition.

Finally, answers vary from one user to another and from one session to the next. A snapshot is a sample. That's exactly why you need ten of them, not one.

What I'm not publishing

The list of 399 questions. It's not coyness: in a niche where direct competitors are well known and two of them are clearly working on their GEO, the list of tracked questions is the bulk of the work. I publish the rates, the methods and the mistakes, not the scope.

Where things stand

Four decisions come out of these four months, and I'm sticking to them for the quarter.

1. Open an editorial workstream on Tech and installation and on Multiplayer and servers, my two gaps at 17% and 12%, before optimising any pages that are already cited.
2. Stop steering by a single AI, since 72% of my questions aren't visible on all three at once.
3. Treat any sudden drop as a model shift until it's confirmed over two snapshots, and check the collection logs before reading it as a signal.
4. Work on mentions as much as citations, which means off-site as much as on-page.

Over to you

The number I find most interesting is still the first one in this article. In your sectors, are you seeing the same gap between Google rankings and AI citations, and is it widening or closing since the spring?

There's almost no public data on this kind of measurement at the level of a single site, so I'd welcome any counterexample.

If you want to run this measurement confidentially on a high-stakes site, let's talk at Pixalione.

This article by Clément Reynaud was originally published (in French) on LinkedIn: Comment je mesure la visibilité de Minecraft.fr dans les LLM.