1. Write the questions your buyers actually ask
Draft fifteen to twenty-five questions in the language a buyer would use before they know your company exists. Do not use your brand name in any of them yet.
Cover four question types. Category questions define the space: "what does a business systems consultant do." Comparison questions force a shortlist: "best partner for a hospitality FF&E project." Local or regional questions test proximity weighting: "AI consulting firms in Virginia." Problem-first questions capture the buyer before they know the category: "our CRM and accounting data do not match, who fixes that."
Pull the wording from real sources where you can — sales calls, inbound emails, support tickets, and the questions prospects ask in first meetings. Invented phrasing produces invented results.
How you know it worked: You have at least fifteen distinct questions and none of them contain your company name. If you cannot reach fifteen without repeating yourself, that difficulty is your first finding: the buyer's language has not been documented.
2. Run every question in a logged-out session
Use a private window, signed out of every account, for all three or more assistants. Personalization and account history will show you a friendlier answer than a stranger receives, which is the answer that actually matters.
Run the identical question wording across every assistant so results stay comparable. Do not refine a prompt because the first answer disappointed you.
Assistants vary their phrasing between runs. Record the first answer you receive and move on rather than re-rolling for a better one.
How you know it worked: Every question has been asked of every assistant, from a session with no account history, using identical wording.
3. Record the answer, the names, and the sources
Build one spreadsheet row per question and assistant, with columns for: the question, the assistant, whether your business was named, every other company named, every source cited with its URL, and how the assistant described the category.
Paste the answer verbatim. Summarizing loses the exact language the system used, and that language is what you will later need to match or correct.
The cited sources matter more than the verdict. They tell you which publications, directories, review platforms, and competitors the system currently treats as authoritative in your category.
How you know it worked: You can sort the spreadsheet by cited source and see which handful of domains shape most of the answers in your category.
4. Now ask the brand-name questions
Only after the unbranded pass, ask directly: "What does [company] do?", "Who does [company] serve?", "Is [company] credible?", and "What is [company] known for?"
Look for four failure types: descriptions that are simply wrong, descriptions that are years out of date, descriptions so generic they would fit any competitor, and hedged answers where the assistant signals it lacks confidence.
A hedge is a finding, not a neutral result. It means the system could not corroborate you well enough to speak plainly.
How you know it worked: You have the machine's own description of your business, in its words, from three or more systems.
5. Trace every cited source to its type
Classify each source you recorded: your own website, a directory, a review platform, a trade publication, a competitor's site, or an unrelated third party.
Count how often your own domain appears versus everything else. In most first audits it is a small minority, which is the point.
Where competitors are cited, open the specific page. You are looking for the structural reason it was selected — a direct answer near the top, a comparison table, a methodology page, published evidence — not for its design.
How you know it worked: You have a ranked list of the sources shaping your category's answers, and you have read the specific competitor pages the assistants chose.
6. Score every question into one of four states
Absent: you were never mentioned. Misrepresented: you were named but described wrongly or with stale information. Generic: you were named with language that would fit any competitor. Cited: you were named and your own content was the source.
Count each state. Those four numbers are your baseline, and they are the figures you will compare against next quarter.
Score by question type as well. Being cited on branded questions while absent on problem-first questions is the most common pattern, and it means you are visible only to people who already know you.
How you know it worked: Four counts, plus a breakdown by question type, recorded with today's date.
7. Convert each state into a specific cause and owner
Absent means you have published nothing that answers that question. The fix is content that resolves the buyer's intent, not more keywords.
Misrepresented means an entity and consistency problem. Your description differs across your website, directories, profiles, and third-party mentions, so the system cannot settle on one account of you.
Generic means a positioning and proof problem. The system can identify you but has no basis to distinguish you, usually because your claims carry no published evidence.
Cited means protect and extend. Note what made that page work and apply the same structure to adjacent questions.
Assign one named owner and one next action per finding. A finding with no owner will still be a finding at the next audit.
How you know it worked: Every gap maps to one of the four causes, one owner, and one action — with no item left as a general instruction to improve the website.
8. Set the re-run date before you close the file
This is a baseline, not a project. Re-run the same question set quarterly, without changing the wording, so the comparison stays valid.
Add new questions in a separate section rather than editing existing ones. Changing the question set breaks the trend line you are trying to build.
Record the date of every run. When traffic or lead quality shifts, this becomes the only record you have of whether your representation in AI answers moved with it.
How you know it worked: A dated baseline file and a calendar entry for the next run, using a question set that will not be edited in between.