Key Takeaways
- Record the before picture before you touch anything. A baseline is the only measurement you can never go back and take. Once a page is rewritten, the answer that page produced last week is gone for good.
- Let the assistants pick your competitor set. The businesses named in the answers are the ones you are actually losing to. The list in your head is a different list, and it is usually wrong in interesting ways.
- Pin five variables or the re-run proves nothing. Question wording, which assistants, the competitor set, what counts as a mention, and the timing. Leave one loose and you cannot tell a change from noise.
- Count names, not vibes. For each question, write down every business named in order. A tally of names across a fixed question set is a number. An impression of how the answers felt is not.
- Run it more than once before you call it a baseline. The same question asked twice can produce different names. Two or three passes on separate days turn a snapshot into something you can compare against.
Most people start fixing their AI visibility the same week they discover they have a problem. They rewrite the services page on Tuesday and claim three profiles on Thursday.
Then, two months later, somebody asks whether it worked, and there is no way to answer. The before picture was never taken, and it cannot be taken now.
Here is how to spend one hour capturing that picture properly, why the competitor set should come out of the answers rather than out of your head, and the five things you have to hold still.
How to baseline your competitors in AI answers before you change anything
Write a fixed list of buyer questions, put them to a fixed set of assistants, and record every business named in every answer, in order. Do it twice on separate days. Then start work.
That is the whole method, and its value is entirely in being done first. A baseline taken after the first fix is not a baseline. It is a measurement of a business that has already moved.
An hour is enough. The discipline is in the setup rather than the effort, and the setup is what the rest of this post is about.
Why the before picture is the one you cannot go back for
Every other number in this work can be recovered. Your traffic is in your analytics. Your reviews are on the platforms. What an assistant said about you last Tuesday exists nowhere except in a conversation that has closed.
Assistants are also not stable in the way a search results page is. Ask the same question twice and the names can differ, which is a property of how these systems generate text rather than a fault in your question, and it is why checking properly means several passes rather than one.
So the baseline is not one snapshot. It is a small set of them, taken close together, that lets you say what the ordinary answer looked like before you started. Without it, any improvement you report later is a story rather than a finding.
Let the answers pick the competitor set
Ask ten buyer questions and write down every business named. That list is your real competitor set for this purpose, and it will not match the list you would have written from memory.
Two surprises are common. Businesses you have never heard of appear repeatedly, usually because they published something specific that reads well. And the rival you think about constantly turns out to be absent, which tells you they have the same problem you do.
Some of what you see will be structural rather than about any one business. The Princeton study on generative engine optimization tested content changes against generated answers and found that adding citations, quotations and statistics moved visibility measurably, which is part of why the same handful of well-sourced pages keep supplying the names. That is a pattern worth recognizing in your own results, and it is examined at length in why AI recommends your competitors and not you.
Fix the set at the baseline. If a new name shows up in a later run, that is a finding worth recording, and it goes in a new column rather than quietly joining the old ones.
The five variables you have to pin
A comparison is only a comparison if one thing changed. Five variables can drift between your baseline and your re-run, and each one is capable of producing an improvement that never happened.
Pin the exact question wording. Pin which assistants you ask. Pin the competitor set. Pin what counts as a mention. And pin when you ask, because running a baseline in a quiet week and a re-run the day after your industry had a news cycle compares two different worlds.
Four latches closed and one left open is enough to make the re-run unreadable, because the loose variable can produce the whole change on its own. The open rail here is timing, which is the one people forget, and the illustration is of the method rather than of any measured run.
What counts as a mention deserves its own sentence. Named in the body of the answer is a clean rule. A link with no name, a description that happens to fit you, and a name that only appeared after you prompted for it are three different things, and mixing them is the fastest way to record progress that is not there.
Different assistants also behave differently by design, which is why the list of tools is a pinned variable rather than a detail. Perplexity's own help center describes a product built around retrieving and citing live sources, while a chat assistant answering from training data is doing something else entirely. Comparing a Perplexity baseline against a ChatGPT re-run measures the difference between two products.
| Column | What goes in it | Why it is there | What it shows on the re-run |
|---|---|---|---|
| Question | The exact wording, pasted not paraphrased | The wording is half the answer | That you asked the same thing twice |
| Assistant and date | Which tool, which day, which pass | Tools differ and answers drift | Whether a change is real or a different product |
| Names given, in order | Every business named, first to last | Position carries most of the value | Movement into or out of the first three |
| You: named or not | A plain yes or no, by your own rule | This is the headline number | Your count out of the fixed question set |
| What the answer cited | Any source named or linked | It shows which pages are supplying answers | Whether your own pages ever start appearing |
| Notes | Hedges, refusals, wrong details | The failures are diagnostic | Whether a wrong picture of you has been corrected |
What the baseline tells you on day one
Before you change anything, the baseline has already earned its hour. Three things fall out of it immediately.
First, your count. Named in 3 of 25 answers is a fact you can act on and compare against. That figure is illustrative here rather than a typical result, because the honest range across niches is wide.
Second, the shape of who is winning. If three names take most of the slots, you are looking at a concentrated market and the work is displacement. If the names are scattered, nobody has been claimed yet and the work is arriving first.
Third, the failures. Every hedge, refusal, and wrong detail in your baseline is a specific repair with an address, and those are usually the fastest things to fix on the whole list.
From there it is a monthly rhythm: change one thing, wait, re-run the pinned questions, and compare. Reading the difference honestly is its own skill, and telling real improvement from noise is where most self-run measurement quietly goes wrong.
If you would rather have the baseline handed to you, the free check runs 25 real buyer questions from your niche through ChatGPT at $0 and emails a plain-English readout within the hour. It runs as soon as you ask for it.
Questions we hear the most
What is an AI visibility baseline?
A written record of which businesses get named when a fixed set of buyer questions is put to a fixed set of assistants, taken before you change anything. It is the reference point every later measurement gets compared against.
How many questions do I need for a baseline?
Enough that one odd answer cannot swing the result, which in practice starts around twenty. We use 25, split across problem-first, comparison, qualifying, and branded questions, because that spread covers how buyers actually ask.
Which competitors should I include?
The ones the assistants name, plus any you specifically want to track. Take the businesses that appear in your baseline answers, not the list you carry in your head, because those two lists are rarely the same.
What counts as being mentioned?
Named in the body of the answer. Decide that once and hold it. A link with no name, a category description that fits you, and a name you had to prompt for are all different things and should not be counted together.
How often should I re-run the baseline?
Monthly is enough for most coaching and consulting practices. Answers move slowly, and re-running weekly mostly measures the noise in the answers rather than any change in your visibility.
Can I baseline for free?
Yes. A spreadsheet, a fixed question list, and an hour will do it, and doing it by hand teaches you more about your own market than reading a report about it. The cost is the hour, every month, forever.
What does the AnswerHalo free check give me as a baseline?
It asks ChatGPT 25 real buyer questions from your niche and emails a plain-English readout: how often you were named, which competitors were named instead, and the one fix to start with. It costs $0.
Can I run the free check today?
Yes. It runs as soon as you ask for it, and the readout lands within the hour between 7am and 5pm Central. Writing down your ten most likely buyer questions costs nothing and is the part most people skip anyway.
Your buyers are already asking. Find out what AI tells them.
The free check asks ChatGPT 25 real buyer questions about your niche and sends you the report within the hour, at $0. If it shows you are already getting named everywhere, we will say so in plain words.
Get my free check →