How do you check what AI assistants say about your company — for free?

Memory, past conversations and custom instructions can affect a response. If you have already discussed the company in your account, you are not reproducing a new visitor’s conditions. Temporary Chat now offers Personalized and Unpersonalized modes. Choose Unpersonalized before the first message so that chat does not use memory, custom instructions or plugins for personalization, or test without signing in where available. Safety checks may still use limited context from earlier conversations.

Second, by variance. Assistant answers are generated fresh each time, and different surfaces of the same vendor pick different sources. When we ran five GEO buyer prompts twice each on the Gemini app and on Google’s AI Mode on 14 September 2026 — clean sessions, logged with screenshots — exactly one cited URL appeared on both surfaces; every other source differed, even though the answers said roughly the same thing. That mirrors Ahrefs’ large-sample finding of 13.7% URL overlap between AI Mode and AI Overviews across 540,000 query pairs.

A useful initial test needs clean sessions, repeated questions and more than one assistant. The steps below explain how to run it.

Ten prompts are enough for a first check. Choose a mix of questions and phrase them as buyers would, using the words they use to compare products, services and providers.

The constrained prompt is the one owners skip and shouldn’t. The more conditions a question carries — language, city, budget, case type — the fewer candidates the model has, and the more likely a smaller company is to be named. It is also where the gaps are most instructive: a firm visible in English and absent in Spanish has found a concrete page to build, not a vague “visibility problem”.

The protocol, step by step: open a clean session — Temporary Chat with Unpersonalized selected before the first message, or log out; ask one prompt per session, copied verbatim; note for ChatGPT whether it searched the web or answered from the model alone, because those are different mechanisms and your brand can be present in one and absent in the other; take a screenshot from the top of the answer, including any map or row of business cards above the text, and another with the source list expanded; close the session; repeat. Two to three runs per prompt — brand prompts vary less, category prompts more.

Log every run in a table as you go; the columns below are the whole template. From the completed log you can compute the two metrics that matter: citation rate — the share of prompts where an assistant cites your page — and share of voice, the share of brand mentions that are yours among all brands named. Our guide on how to measure AI visibility defines both in detail; you do not need more than these two on day one. For a local business, note the map column carefully: in our study of lawyer searches, 43 of the 63 firms ChatGPT named in its map answers were on the cards. This was the first run’s 11 map answers in our 22 September 2026 study of 72 answers to 12 questions. Some cards were outside the screenshot, so 43 is a lower bound, not a market-wide rate.

An absent company can have several explanations. Treat the following patterns as directions for investigation, not diagnoses established by a single answer.

If the company is absent even when you ask for it by name, check access and public information first. Does robots.txt allow the relevant crawlers, and is the main page text available in the server response? Our crawler-access guide explains these checks. Also inspect indexing and company profiles; absence alone does not prove a technical block.

If the assistant recognizes your brand but omits it from buyer questions, compare your pages with the cited sources. Do they cover the relevant prices, conditions, locations and comparisons? Are independent descriptions available? Our guide on why ChatGPT does not recommend your company examines this situation.

If results differ between repeats or assistants, preserve that difference. A single disappearance does not establish a decline, but persistent absence from one service can warrant investigation. Repeat the same set every two to four weeks and examine several measurement cycles.

Manually checking one brand across 10–20 prompts every two to four weeks takes roughly two hours per cycle and does not require paid tools. That is enough for an initial measurement. Our guide on GEO in-house versus hiring an agency explains which tasks you can continue yourself and which may need a contractor.

Paid trackers become useful when you have too many prompts, assistants or weekly competitor checks to manage by hand. The entry prices on the vendors’ own pricing pages, checked September 2026, are $29/month for 15 prompts across four engines at Otterly, with Gemini and AI Mode as paid add-ons, and $99/month per domain for 25 tracked prompts in Semrush’s AI Visibility Toolkit. These tools measure whether you appear; identifying the causes and deciding what to change still requires analysis.

Our free audit is this same method, run for you: ten prompts in one language where you are currently invisible, plus the crawler-access checks and the specific fixes the log points to. If you have already run the protocol yourself, bring your log — a before-and-after with your own baseline is the most convincing report we can give you.