Ask what ChatGPT says about your brand and the honest reply is: for whom? Over the past year the major assistants have become personal by default. ChatGPT remembers past conversations and applies what it learned, its custom instructions, a standing brief that shapes every answer, were expanded to 5,000 characters in July 2026, it can use shared location to sharpen local recommendations, and Gemini increasingly draws on a user's Google context. Two people asking the identical question now receive materially different answers, sometimes naming different brands.
That breaks the most intuitive habit in this field: screenshotting one answer and treating it as the answer. Here is what personalization actually changes, what it does not, and how to think about visibility when the answer is a distribution rather than a fact.
What personalizes an answer
Four layers stack on top of the base model.
Memory. Assistants carry context across sessions: your role, your industry, tools you have mentioned, brands you have complained about. A procurement manager who once discussed a vendor gets future category answers tinted by that history.
Standing instructions. A 5,000-character custom brief is enough to encode strong preferences: favor open-source tools, assume an Australian context, never recommend enterprise software. Every answer passes through it silently.
Session context. What was asked earlier in the conversation reshapes what comes next, which is one reason the same prompt mid-conversation and fresh produce different shortlists.
Signals like location and account context. Location sharing changes local recommendations outright, and ecosystem assistants can draw on mail, calendars, and documents where users allow it.
Underneath all of that sits the volatility that exists even without personalization: identical, logged-out, fresh prompts already return unstable shortlists, with only about 30% of brands persisting from one answer to the next in AirOps' 2026 analysis. Personalization multiplies variance on top of a base that was never stable to begin with.
What this means for brands
Three practical consequences.
Anecdotes are now almost worthless. The founder whose ChatGPT recommends the company proves only that their ChatGPT does; it has watched them talk about the company for a year. The prospect's ChatGPT, primed by different history, may never mention it. Single screenshots, in either direction, are the astrology of AI visibility.
The fundamentals get amplified, not replaced. Personalization tilts answers; it does not invent them. The candidate set still comes from the shared substrate: what the model learned in training and what retrieval finds in the sources everyone shares. A brand consistently described, well-reviewed, and present in the lists and communities engines cite enters more people's personalized answers, whatever their instructions say. The traits in what winning brands do differently are exactly the traits that survive personalization, and the local version, where location sharing makes personalization strongest, is covered in our local guide.
Measurement needs a defensible baseline. If every user sees a different answer, what should tracking measure? The wrong answer is someone's logged-in account, which measures that account's history. The right answer is the neutral baseline: clean, logged-out, memory-free polls of the same questions, repeated over time, per engine, at volume. That is the shared substrate every personalized answer is a variation of, and it is the only version of the question that is stable enough to trend, compare against competitors, and attribute changes to. It pairs with the other axis of answer variance, whether the engine answered from memory or live retrieval, which our memory versus live explainer covers.
This is Diploria's methodology, stated plainly: we poll from clean contexts through official interfaces, store every full answer, and measure the distribution, mention rate, position, sentiment, and share of voice across repeated runs, rather than presenting any single response as the truth. A baseline cannot see inside any individual's personalized session, and nothing can; what it can do is measure the input every session draws from, which is the part a brand can actually change.
How to use this next week
Stop settling debates with screenshots; when someone shares one, ask which account, which day, which conversation. Establish a baseline for the questions that matter in your category and watch the trend rather than the moment. Work the shared substrate, since consistency, reviews, lists, and community presence are what personalization amplifies. And when a wrong claim appears repeatedly in your baseline, fix the cited source, because errors in the substrate reach every personalized variation at once.
The free check runs a clean baseline poll across the engines in under a minute, which makes it a fair first look at what the neutral version of your answer currently says.