Last week I wrote about the machine that decides what your customer hears. A study this week complicates that in a useful way. It suggests the machine has often decided before it looks.
A researcher went through 60 ChatGPT conversations and found that the brands ChatGPT named in its own opening search queries were cited about 33 times more often than brands it turned up by actually retrieving and reading pages. In 21 of 27 conversations, it named brands before it fetched anything. Pages it did go and read were cited only about 3 percent of the time. It's one small analysis, so I'd hold the exact numbers loosely, but the shape is worth sitting with. By the time the model goes looking, it mostly already has an answer in mind.
If that's roughly right, a lot of the advice to optimize your own pages for answer engines is necessary without being the whole game. Clean pages and clear claims help the model quote you correctly once you're in the running. They do less to get you into the running in the first place. That part comes from what the model already absorbed about you before the question was ever typed.
And that prior is increasingly built on pages you don't own. AirOps looked at 3.5 billion AI citations and found that creator and social content grew its share by about 140 percent between August 2025 and this June, while citations to brand-owned and product pages fell around 10 percent. YouTube alone rose 158 percent. It's a vendor's analysis, so again, direction rather than decimal. But it lines up with the first study. The model's sense of who to recommend is shaped by what other people say about you, on their platforms, more than by what you publish on yours.
There's a second story running alongside this, about who even gets to read the web. A tool called ShieldFont made the rounds this week. It quietly feeds web scrapers wrong words while human readers see the correct page, and in testing it got more than 90 percent of scraped copies rejected. Cloudflare, meanwhile, is building a way to charge AI crawlers by the visit. Publishers are starting to treat machine access as something to meter or deny, not assume.
Put those together and the job shifts again. You're not only writing pages for a second audience. You're managing a reputation inside a model, and most of the raw material for it sits on other people's sites, in their videos, in reviews and threads you can influence but not control. That's slower work than editing a landing page, and harder to measure. It looks more like earning mentions than producing content.
None of this is settled. The studies are small or vendor-run, and the models change month to month, so I wouldn't rebuild a strategy on any single figure here. But the direction has held for a while now, and it points away from making more and toward being worth citing when you're not in the room.
These are the questions I want to put to the marketers I'm meeting around Vancouver this fall, the ones running real accounts. I'd like to know whether any of them can yet find themselves inside the machine's answer, and what they do when they can't.
Sources
- ChatGPT Already Knows Who It'll Recommend Before It Searches: suganthan.com
- Is Your Brand Missing Out on the Fastest-Growing Source in AI Search? (AirOps): airops.com
- The web's newest weapon against AI scrapers is a font (Ars Technica): arstechnica.com
- How Cloudflare is making AI pay for content (ByteByteGo): bytebytego.com