If your AI search plan relies on getting your page found when ChatGPT goes looking, you are working on the part of the process that almost never names a brand. Suganthan Mohanadasan split brands in ChatGPT answers by when they first showed up. Brands that ChatGPT wrote into its own first search query were named in 68.9% of final answers. Brands it only found in the retrieved results were named in 2.1%. The figures come from his own network-traffic captures and he gives no sample size.

The model already has a shortlist before it searches. That shortlist comes from what it absorbed about your category ahead of time through repeated, independent coverage across the web. This is the off-site exposure mechanism, and it is why digital PR carries nearly half the weight in a reliable SEO program.

I was wrong about what the Reddit citation drop proved

In my August 24 post on Reddit, I misread the evidence. I wrote that “your asset just expired” and “The Reddit strategy you bought in July was worthless by August 14.” I read the drop as OpenAI shutting down manipulation. Citations measure what ChatGPT retrieves and links, and they cannot show what the model already knows about a brand. On August 8, site: queries jumped from 0.37% of ChatGPT’s query mix to 16.8% in a day. Promptwatch reported that Reddit held a 3.83% average share of ChatGPT citations from July 18 to August 7. The sharp drop came on August 14, six days after the site: jump. The August 14 to 17 average was 0.52%. Suganthan looked at the same timeline and said: “It looks much more like a retrieval change,” and “I wouldn’t interpret this chart as: ChatGPT suddenly decided Reddit is bad.”

My underlying advice stands. Seeding Reddit threads is inauthentic. Google’s own AI optimization guide says seeking inauthentic “mentions” across the web “isn’t as helpful as it might seem.” You also cannot see when a platform will change its rules. The tactic was a bad buy for the reasons my original post gave.

ChatGPT restricts searches to domains it picks itself

Mark Williams-Cook noted in his Core Updates newsletter this week that ChatGPT now adds site: operators to most of its outside searches. The model decides which domains to search. He concluded that getting a footprint in the training data is getting more important, and digital PR is the biggest lever for it. In July I argued citation counts are the wrong scoreboard, and this is a different point: the shortlist exists before any citation happens.

If your AI search budget goes mostly to tweaking your own pages for AI, move some of it to earning coverage on independent sites. Your own pages still decide whether ChatGPT can use you once it searches your domain, but they do not put you on the shortlist that decides which domains it searches.

Being named by an AI platform drives immediate visits

PPC Land’s write-up of Similarweb’s September 10 State of Ecommerce report tracked what happens when AI recommends a financial product. When an AI answer recommended Capital One, 14.2% of those users visited Capital One within a week, while 3.8% visited American Express, a competitor it did not recommend, in Similarweb’s US desktop data from July to December 2025.

You get onto the model’s initial shortlist with earned coverage on independent sites people actually read. You need data-led PR, expert commentary, and placements on relevant industry sites. Nobody can look inside the weights and trace a recommendation back to a specific set of comments or articles.