The core and the carousel

You asked ChatGPT the question your customers ask, and there you were. Congratulations — you have learned almost nothing. Ask again next week and there is a good chance you are gone, and that nothing you did caused it.

This is the part of AI visibility that nobody selling it wants to lead with, because it makes the product harder to demonstrate. AI answers are not rankings. They do not hold still. And the size of the movement, once measured properly, is larger than most people assume.

Start with the shortest possible interval. SE Ranking took a fixed set of 10,000 keywords and ran them through Google's AI Mode three times on the same day. Between the three sets of answers, the average overlap of exact URLs was 9.2%. For 21.2% of the keywords there was no overlapping URL at all — three runs, three entirely different sets of sources, same question, same afternoon (SE Ranking).

Now stretch the interval. Profound compared citations for the same prompts one month apart, at around 80,000 prompts per platform. Google AI Overviews had replaced 59.3% of cited domains, ChatGPT 54.1%, Perplexity 40.5% (Profound).

So a single check tells you what happened in one draw from a deck that gets reshuffled. That is a real finding, and it is where most articles on the subject stop. But it is also incomplete, and the larger study that followed shows why.

The movement is not evenly distributed

Sistrix tracked 82,619 prompts across three platforms and six countries every week for 17 weeks, from December 2025 to April 2026 — roughly 1.5 million snapshots. The headline numbers match everyone else's: Google AI Mode swaps out 56% of its cited domains each week, ChatGPT Search 74%. Underneath them sits a structure that changes what you should do about it.

Two different things are happening in the same answer. There is a small set of sources the system keeps coming back to, and there is a carousel of everyone else, cycling through week by week. Being in an AI answer once means you were on the carousel. It says nothing about whether you are in the core, and those are not the same business outcome at all.

The pattern holds even for the most obvious case. In Google AI Mode, a brand's own domain appeared in all 17 weeks for only 43% of queries about that brand. The 12 to 15 domains cited alongside it changed at 70% per week. Strong brands hold their place; the space next to them is reallocated constantly (Sistrix).

The platforms also differ far more than the averages suggest. Google AI Overviews barely move: for 53% of prompts, not one source changed across the entire 17 weeks — the answer text was rewritten each week, but it drew on the same eight domains. ChatGPT is the opposite extreme, and cites only three to four domains per answer against AI Mode's 14 to 16. Fewer slots, and almost all of them rotating.

What gets into the core

Sistrix also classified which types of pages survive. The spread is wide enough to plan around:

Video (YouTube) reached the stable core in 24% of cases, big tech domains 16%, Wikipedia 12%, marketplaces 10%, forums and UGC 3% — and news and media 1.4%. Press coverage is cited and then gone the following week. It still does what press coverage does, but as a way to buy a permanent place in AI answers it is close to worthless.

Content that lasts does better than content that reports. In the German AI Mode data, 85% of core URLs were evergreen against 77% in the periphery, and product and service pages dominated the core while guides and articles rotated through it.

If you have five employees

Three things follow, and none of them require a budget.

Stop treating a single screenshot as evidence — yours or an agency's. If someone shows you one ChatGPT answer with your name in it as proof their work succeeded, they have shown you a coin landing heads. Ask for the same prompts run repeatedly over weeks, and for the share of runs you appeared in.

Put your effort into the pages that can hold a core slot. Your service pages, your prices, your process — the pages that will still be true in a year. That is the profile the core rewards. A news push has a 1.4% chance of lasting, which is a fine thing to know before you pay a PR agency for GEO.

Look at video honestly. A 24% core rate is the highest of any content type in the data, and a phone and a tripod clear the entry cost. It is the one lever on this list where a small firm is not structurally outmatched.

The uncomfortable part is that none of this produces a rank you can frame. AI visibility is a rate, measured over time, and anyone who reports it as a position is either not measuring or not telling you how.