Topical Authority Injection: How LLMs Pick Winners Before You Even Ask
When you ask ChatGPT, Claude, or Gemini something, they usually don't answer straight from memory. They break your prompt into multiple search queries, pull fresh results from the web, and synthesize an answer from that. This is called fan-out query, and it's completely invisible to the user.
We used Visby's data to analyze over 11,000 prompts across ChatGPT, Google AI Overview, and Claude to see what's actually happening in that invisible step. We ran this on prompts in multiple languages, not just one, and the pattern held consistently across all of them.
Topical Authority Injection
Here's what stood out: ChatGPT adds a brand or institution name that the user never mentioned in about half of the prompts we looked at. Google AI Overview does this in roughly 1 out of 20 prompts, Claude even less. ChatGPT does this about 12x more often than the other two.
There isn't an established term for this yet. We refer to it as Topical Authority Injection.
The difference from classic SEO ranking matters here. In SEO, your competitor shows up after the search happens, on the same results page as you. With Topical Authority Injection, the model has already decided who the "default answer" is before the search ever runs. You're not even in the race unless you're already on that list.
A simple example: ask "reliable online fashion retailers" and ChatGPT will silently inject brand names into the search query that were never in the prompt, big, well-known players in that category. Nobody typed those names. The model just assumes they belong there.
This has a real implication for anyone thinking about AI visibility: if your category already has an entrenched "default list" baked into the model, getting added to it later is a much harder problem than ranking on page one ever was.
We're continuing this research across more languages and categories. If you want to see how your own brand shows up (or doesn't) across ChatGPT, Claude, Gemini, and Google AI Overview, that's what we built Visby for.
This is interesting because I'm trying to solve almost the opposite problem with a relatively new site. We've been investing heavily in original research rather than just publishing more SEO content, including commissioning a U.S. survey and running our own product tests.
What I'd be curious to see is whether strong original data can eventually change that "default list." If an unknown brand becomes the primary source for a topic and starts getting cited elsewhere, does ChatGPT eventually begin injecting that brand into related fan-out queries? Or does the advantage mostly stay with already established entities?
That would make a really interesting longitudinal test.
This makes me wonder about a slightly broader version of the problem.
If the model can introduce an entity that the user never specified during query formation, there seems to be a related question on the execution side: can an agent introduce a decision or action that was not contained in the original command?
I have recently been looking at agent-driven operations where the original instruction reaches an enforcement boundary, but the agent can still make subsequent decisions about how to proceed. That made me realize that the “original command” may not be a sufficient description of what ultimately happens.
So there may be two separate places where hidden choices enter the chain: before retrieval, when the model shapes the question, and later during execution, when an agent shapes the action.
I would be interested to see whether your research distinguishes the injected entities from other decisions the model makes along the same chain. The latter seems potentially important for understanding AI visibility as well as AI control.
The fan-out framing makes the visibility problem much easier to reason about. I’d be curious how you distinguish durable topical coverage from a temporary burst of mentions—especially when a page is relevant but has little engagement yet. A small set of recurring, genuinely useful references seems more defensible than trying to maximize raw volume.
Agreed, recurrence beats volume. One mention is an event, repeated mentions across independent runs are a pattern. Engagement isn't a great filter either, since some genuinely useful pages take a while to get traffic. A small set of prompts checked consistently over weeks tells you more than any one-time spike.
The useful distinction here is that category association is a prior, while page level citation is evidence. To make it actionable, keep a small test set of natural buyer prompts, note the categories and brands that appear without prompting, and then inspect which pages get cited. That shows whether you have a naming problem or an evidence problem. The fixes are different.
Thanks for the contribution - great distinction, really well laid out.
Glad it landed. The split helps when a brand is named but never cited, or cited but never named. Same dashboard, two different next steps.
The 11k-prompt analysis is compelling, but the commercial test seems downstream. Have brands using Visby actually changed their AI visibility or recommendations after acting on the findings, or is that still unproven?
Automation that works with the user's existing tools tends to survive longer than automation that needs its own environment. What is the smallest first step you are optimizing for?
Automation that works with the user's existing tools tends to survive longer than automation that needs its own environment. What is the smallest first step you are optimizing for?