Rockstar Is Not the Source for Rockstar’s Release Date
One question, three machines, one minute: twenty-seven cited domains and not one shared by any two of them. If your AI visibility number came from a single call, it is a reading, not a measurement.
Somewhere in the next six weeks, in a quarterly review, a slide will go up with a single number on it describing how often a brand appears in AI answers. Someone will have paid for that number. Nobody in the room will ask how many times the question was asked.
One budget line, or two
The person this matters to is whoever has to defend next year's marketing budget, which for most calendar-year companies is being set right now. The ask is a separate line for AI answer visibility. The refusal is almost always the same sentence: answer engines read the web, Google ranks the web, so winning search wins both. One line, one owner, done.
It is a reasonable argument. It is also, as far as we can tell, untested by almost everyone making it. So on October 24 we tested it on one question, in one city, inside a single minute, with every row published.
We picked a question with no ambiguity in the answer. Grand Theft Auto VI ships on November 19, 2026. Rockstar has said so. It is on the store pages. Some phrasing of when does it come out is searched well over a million times a month, which makes it one of the least obscure facts on the internet.
Ask the same machine the same thing three times
ChatGPT got the date right all three times. That part is unremarkable and worth saying plainly, because the first version of this piece claimed otherwise and was wrong.
What moved was where it said the date came from. Across three identical calls it cited seven different domains, and exactly one of them appeared in all three answers. Six of seven were present in some calls and absent from others. A brand measuring its citation share from a single call would have been reading one draw from a shuffled deck.
Perplexity, asked the same string three times, returned the same ten domains in the same order, three times out of three. That is the fourth separate occasion we have measured that split. It is a measurement of these questions on these days, not a league table, and we would not publish it as one.
Domains shared by any two of the three instruments. ChatGPT, Perplexity and Google were given the identical query string within the same minute and returned twenty-seven distinct domains between them, with no overlap in any pair.
Then ask three different machines the same thing
This is the part that bears on the budget argument, and it is blunter than we expected.
| Instrument | Domains returned | Stable across repeats | Shared with the others |
|---|---|---|---|
| ChatGPT — three identical calls | 7 | 1 of 7 | 0 |
| Perplexity — three identical calls | 10 | 10 of 10 | 0 |
| Google — organic positions | 5 | not applicable | 0 |
| Google — the same question reworded | 6 | 1 of 6 held from the first phrasing | 0 |
Read the last row twice. Google was less consistent with itself across two phrasings of one question than Perplexity was across three repeats of the same string. Same engine, same city, same minute, and rewording the question changed five of six results.
Who the machines actually asked
We registered, before the pull, which domains we expected to see: Rockstar's own site, Take-Two's, IGN, GameSpot, Eurogamer. The obvious authorities on a games release date.
Across three ChatGPT calls, that list returned nothing. Of its seven cited domains, exactly one — PC Gamer — was mainstream press. The rest were fan-built wikis and single-franchise sites. On the third call, all four citations were sites that exist to answer questions about this one unreleased game, and nothing else.
Rockstar Games was cited zero times out of three on a question about Rockstar's own release date.
CULT+MATH, run 3063e451, October 24, 2026
Perplexity's top two positions went to the same kind of site, ahead of the BBC, IGN, CNET and USA Today, all of which it also cited further down. Six of the seventeen domains the two engines cited between them exist to answer about this one product.
We are going to resist the obvious conclusion. These sites were cited; nothing in the run measures why, and a correlation between being single-subject and being cited is not a mechanism. One query, one day. It is enough to change what we test next. It is not enough to tell anyone to restructure a website.
What this evidence cannot say
Organic results are a minority of the Google page. We requested depth ten and got six and seven organic positions; the rest of each page was an AI overview, a video carousel and paid units. A statement about organic overlap is not a statement about what a person sees.
Both engines were read at the platform-default budget tier the API serves, which is not the model a consumer is given. Every AI visibility measurement we have found, including our own, has this limit, and most do not mention it.
One query, one city, one day, one product. Nothing here observes a visit, a purchase, or a person. And the strongest result in the run — the zero overlap — came from a comparison we had already run badly earlier the same day, which means we had seen a number before we wrote down what we expected. The study page says so, in the section where the prediction is recorded.
- More than one assistant in the test — a single machine's citation list describes that machine, not the web.
- The same query string, inside the same minute — any wider window and you are measuring drift rather than disagreement.
- A fact with an obvious owner — the result only means something when there is a source everyone would agree should be cited.
- A budget that splits two ways on the answer — if nothing you fund changes, this is trivia rather than evidence.
Take your ten highest-intent buyer questions, put each to one answer engine three times and to Google once, and count how many domains appear in both.
Not a vendor, not a dashboard, and not ten thousand queries. Ten questions you already know your buyers ask, in a spreadsheet, with every cited and ranked domain written down. The three repeats are the part nobody does and the part that decides whether your number means anything: if an engine disagrees with itself across three calls, it cannot be compared to anything, and no overlap figure built on one call is worth defending in a budget meeting.
- What it costs
- About three hours by hand, or roughly three dollars in API calls if you script it. The real cost is choosing the ten questions honestly rather than the ten you already rank for.
- How you'll know
- Overlap above half says one budget line and you should stop asking for a second. Below half says two, and you now have the slide. Either way, within-engine repeat stability tells you whether the number you have been reporting all year was ever stable enough to report.
The question underneath all of this is not about AI. It is the oldest one in measurement: how many times did you look before you wrote the number down? Every finding in this piece exists because the answer was three instead of one. That is not a sophisticated technique. It is the cheapest possible rigor, and almost nothing published in this category currently clears it.
Every row behind this piece — all twenty-seven domains, the per-call presence, the advertising series, the search demand and the comment vocabulary — is published and filterable in the standing study at cultmath.com/research/games/gta, which is re-run weekly and corrected in place, dated. It currently carries one retraction, made in the same pass that produced the measurement that overturned it.
Built on the ADMADS Engine: sourced, dated evidence, with its limits stated. How these pieces are evidenced.
Put the evidence behind the decision in front of you.
Decision Briefs set out the sourced, dated evidence for a decision, with its limits stated. Or commission one built around yours.
See the AI visibility brief Or commission a custom brief →