How ChatGPT Chooses Who to Recommend
01Nobody is holding a competition
The instinct everybody brings to this is a ranking. You picture the model receiving a question, gathering the candidates in your category, weighing them against each other, and returning the winner. It feels like search, so people optimise for it like search.
Nothing I have observed behaves that way. The answer arrives too fast and too consistently for a contest to be happening, and it usually names the same two or three things whether or not the model goes and reads anything.
The model does not rank candidates. It remembers reputations.
Where this comes from and what it is not
Everything below is observed behaviour, gathered over a year of asking the same category questions in fresh sessions across the products I run. It is not a description of how the system works underneath, because I do not have that and neither does anybody selling you a course about it. You can watch what a model does a thousand times without ever seeing why it did it, and the watching is still worth more than the guessing.
02What actually predicts getting named
Four patterns showed up often enough, across enough different categories, that I now treat them as working assumptions. Each one comes with a way to check it on your own brand rather than taking my word for it.
The four that kept repeating
The strongest predictor of being named is having been describable long before the question was asked. Products that existed, were written about and were explained consistently for years get named without the model going anywhere. Newer things get named when something recent is fetched and read.
Every product I run that gets named regularly can be explained completely in a single clause. The ones that get skipped are the ones where the honest description needs a paragraph and two qualifications.
A claim that appears only on your own site behaves differently from the same claim appearing in places you do not control. Independent sources saying the same thing is the pattern that survives, which is why the listicle and the forum thread keep turning up as citations.
Being unknown is a slow problem. Being contradictory is a fast one. When sources disagree about what you are or what you offer, the safest available move is to name something else instead.
The one that surprised me
Pattern four is the one I did not expect and the one that changed how I work. I assumed the enemy was being unknown. It is not. An unknown brand is simply absent, and absence is fixable with time and output. A contradictory brand is actively avoided, because a model holding two versions of you takes the route that requires no reconciliation. Obscurity costs you a mention. Contradiction costs you the benefit of the doubt.
That is the same finding written from the other direction in the entry about saying one true thing everywhere, and it is why the unglamorous filing work outperforms most content strategies.
You are not competing for a slot. You are being remembered, or you are not.
03Working with a memory instead of a ranking
If the thing you are trying to influence is a reputation rather than a ranking, the work changes shape. Reputations are slow, they are built by other people, and they cannot be bought in a quarter. A ranking is something you climb. A reputation is something other people keep for you.
Five things that follow from that
Write the sentence yourself. Decide the one clause you want repeated about you, then put that exact clause everywhere a machine reads, so there is nothing to reconcile and nothing to invent.
Earn the third party mention before the tenth blog post. One paragraph about you on a source you do not own moves more than a month of your own publishing, because agreement between strangers is the pattern being rewarded.
Be describable before you are impressive. A plain product with a clear boundary gets named ahead of a sophisticated one that takes three sentences to introduce. Simplify the description before you extend the feature list.
Give the browsing path something current. Memory covers the old and established. A dated, maintained page is what gets fetched when the model decides to go and look, which is the lever available to anyone too new to be remembered yet.
Ask, on a schedule, and write it down. Same questions, fresh sessions, once a month. The trend is the only part that means anything, and you cannot see a trend you never recorded.
The honest limit of all this
None of it is a lever you pull with a guaranteed result, and anyone promising otherwise is describing a system nobody outside these companies can see. What you get is a set of conditions that made being named more likely across every product I tested, and a way to measure whether it is working on yours. Read the answer it gives about you before you decide any of this is theoretical.
The reassuring part is that the work is legible. There is no auction, no bid, and no relationship to buy. Nothing about being recommended by a model is purchasable, which is the best news in this entire field. It is earned the slow way, by being one clear thing in every place that mentions you, which is also how a free tool ends up quoted at scale.
Tomorrow: the doorbell almost nobody rings, and the two ways to make it ring itself.
>How does ChatGPT decide which products to recommend?+
>Can you pay to be recommended by an AI assistant?+
>How do I test what AI says about my brand?+
Abd Shanti. "How ChatGPT Chooses Who to Recommend" CITED, Entry 021, Aug 14 2026. unknown.ps/blog/how-chatgpt-chooses/