Obility Editorial · · 4 min read
The person checking is the least typical user
A familiar moment: a founder asks ChatGPT for the best tools in their category, sees their company named, and relaxes. A week later a prospect mentions that ChatGPT never brought them up. Both can be telling the truth. The founder has spent months in that account drafting launch copy, comparison pages and investor updates, and the assistant has been keeping notes.
Memory is now a standard feature in the major assistants, and each vendor documents what it keeps. Read that documentation as a marketer and one conclusion follows. The account you use for work is close to the least representative place to measure how your brand appears to a stranger.
Each assistant documents what it carries into an answer
OpenAI’s help center says that when Memory is on, ChatGPT can draw on past chats, saved memories, custom instructions, files in its Library and content from connected apps such as Gmail. The same page answers the question that matters here: memory can personalize web searches, because ChatGPT may use relevant details to formulate the search query it runs. With chat history referenced, it may also use the account name shown in the product.
Anthropic’s help center lists what Claude’s memory keeps, and the first item is your role, projects and professional context. Memory is on by default on Claude’s Free, Pro and Max plans. Google’s Search help describes Personal Intelligence in AI Mode, which draws on previous searches and activity saved in your Search history when history and personalized recommendations are turned on. Google says that experience is currently available in the US, in English.
Put those together. If you have asked an assistant to rewrite your homepage a dozen times, you have told it what your company does, who it competes with and which category you believe you belong to. When you then ask for the best tools in that category, that context is relevant in exactly the way these systems are built to notice. None of the vendors says it will favor your brand, and you should not assume it does. You simply cannot tell, from inside your own account, how much of the answer is about you.
Location changes the search before the model reads a page
Memory is one layer. Location is another, and it applies even to people who are not signed in. OpenAI states that ChatGPT collects general location from your IP address and may share that general location with its third-party search providers. It also says ChatGPT typically rewrites a question into one or more targeted queries before sending them, and gives its own example: a request for good restaurants nearby becomes a search for top restaurants in San Francisco. The same page notes that a VPN can change the location it infers.
For a brand, that means a buyer in Manchester and a marketer in Bangalore asking the same question can trigger different searches, read different pages and get different shortlists before any memory is involved. A competitor that dominates review sites in one country can be named for one of them and missing for the other.
Start each check from a stranger’s position
Every major assistant now offers a way to start without your history. In ChatGPT, a temporary chat can be set to Unpersonalized before it begins, and OpenAI notes that the choice cannot be changed once the conversation starts and that temporary chats do not create or update memories. In Claude, an incognito chat is kept out of memory and chat history, and a regular chat can be started with memory switched off from the plus menu, again only before the first message. Google ties AI Mode personalization to having history and personalized recommendations enabled, so run checks with that setting off or from an account that has no history.
Then control location on purpose. Choose the markets your buyers are in, run each check from a network or setting that places you there, and write the market down next to the answer. Keep the prompt wording fixed and record the engine, date, language and whether you were signed in. That is the record Obility’s Answer Engine Insights is built around: prompts grouped by audience, buying stage and market, with the wording and context attached to each observation so that like is compared with like.
A clean session is a baseline, not your buyer’s answer
There is an honest limit to all of this. Your buyers have their own memories, search histories and connected apps, and none of it is visible to you. A clean session does not reproduce what a particular prospect saw. It measures the shared starting point: what an engine says about your category when it knows nothing about the person asking. That is also the part your own pages, your coverage on other sites and your product facts can actually move.
Answers also vary from run to run with identical settings, so one clean answer is still a single draw; count how often you appear across many runs. When a prospect says an assistant left you out, ask for the exact wording and their country, then rerun it clean. A gap that survives the clean run is real work for the content team. A gap that only shows up in one person’s account tells you about that account. Your own assistant remains a fine place to draft. It is a poor instrument for measuring.