The answer is written, not looked up
When you search on Google, you get a list of pages Google had already found and stored. Repeat the search a few minutes later and the list mostly stays the same. An AI assistant does not work like that. It writes the answer again for every question.
The assistant builds the answer piece by piece, and at each step it picks one of several possible ways to continue. Most assistants deliberately allow some chance in that pick; it is part of what makes answers sound natural. The result: the same question, even on the same assistant, can come back as a different text. When the text changes, the businesses named in it can change too.
Illustrative example, not a measurement: you ask "Can you recommend a dental clinic in Ankara that is good with children?" and the answer names three clinics. You ask again in the afternoon. Two of the three are still there, and a name you have not seen before has taken the third place. Which answer is right? Both. Each one is the answer of that moment.
Five engines, five separate answers
Your customer does not only use ChatGPT. Gemini, Claude, Perplexity and the Google AI Overviews summary at the top of a Google search answer the same question too. These are separate products, most of them from different companies. Gemini and Google AI Overviews both come from Google, but they do not have to give the same answer to the same question. In this article we call each of these assistants an engine.
- Some search the web while they write and show their sources. Others lean more on what they learned before.
- Each one reads different sites, different reviews and different lists.
- Google AI Overviews does not appear on every search. For some questions there is no AI summary at all, only the normal results.
So being named on one engine does not mean you are named on the others. The reverse is also true: not showing up on ChatGPT does not mean you are missing from Perplexity.
Wording, language and country can change the answer too
You asked in your own words. Your customer will use theirs. These three questions describe the same need, but they do not have to bring the same list:
- "Which car inspection company in Istanbul can I trust?"
- "Where should I get a used car checked before I buy it?"
- "Can you suggest an affordable car inspection in Kadıköy?"
Language matters as well. A question in Turkish and the same question in English can bring different businesses. The answer can also shift depending on which country the question comes from and which country it is about. If you asked from Turkey, in Turkish, and saw yourself, you still do not know whether a customer in Germany asking in English sees you. We walked through this in our guide for clinics that treat patients from abroad.
Models and their sources change over time
AI companies update their models regularly. A new version can answer the same question in a different tone and with different names. The web does not stand still either: new articles, reviews, guides and business listings on map and directory sites appear, old ones disappear. An assistant that searches the web while answering is reading a slightly different internet every week.
So last month's answer may not be today's answer. Our methodology says it plainly: we do not claim that the same answer will repeat for another account, location, model version or date.
One answer is a snapshot
All of this leads to the same place. A single answer is a snapshot of one moment. It does not show where you stand.
- If you saw your name: good news, but it is not certain to repeat on the next question, on another engine or next week.
- If you did not: worrying, but it does not mean you never appear. You may come up for other wording or on another engine.
- Either way, you do not know where you stand against your competitors. That is the real question.
To see who tends to fill these answers instead, read our field note on local businesses.
For a fair picture, these four things have to stay the same every time. Otherwise you cannot tell whether a difference between two checks came from something you changed or from chance:
How SOV Tracker measures it
On the paid plans, the measurement is set up to keep those things fixed:
- You choose the questions, and the exact same text is asked every time. If even one letter of a question changes, it counts as a new version of that question.
- Each question is tracked for the target country and language you pick. We ask the engines through an API, not from a phone or computer in that country. So we do not claim to show exactly what every user in that country sees.
- Every tracked question is asked once a week on all five engines within the same short time window: ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews.
- Questions you mark as priority are also checked every day on ChatGPT and Google AI Overviews. That daily check (Priority Pulse) covers only two engines, so it is kept out of your weekly SOV Score, explained below.
- A check that failed for technical reasons, returned no AI answer, or was refused by the engine is not recorded as a zero, as if you were not mentioned. It is kept as a separate state. If an engine could not be measured well enough that week, the result says so.
- If your question list or your competitor list changes, the comparison with the previous week is not shown, and the reason is written next to it. Two weeks that measured different things are not compared as if they were the same.
The headline result is your SOV Score (SOV stands for Share of Voice): your share of mentions against the competitors you chose. We count how many times you and those competitors are named in the answers, then work out your part of that total. If you have not added competitors, the SOV Score is shown as "not measured"; the gap is not filled in as zero or as one hundred percent. The full calculation is in how the SOV Score is calculated, and every rule is on the methodology page.
Measuring does not get you into the answers. What it gives you is a trend you can trust more than any single answer. Even a change between two measurements does not, on its own, prove what caused it, but it is far firmer ground than one look at ChatGPT.
Free scorecard first, regular tracking next
You do not need to pay to start. The free scorecard asks ChatGPT five questions that leave your business name out, and shows how many of the answers named you on their own, and which businesses were named instead of you. No signup is needed, and it can be run once every 30 days for the same business.
The headline percentage on the scorecard is Discovery Visibility, and it is not the SOV Score. It does not measure a share against your competitors; it shows in how many of the questions ChatGPT actually answered your name came up. A question that got no answer because of a technical error is not counted as a zero. It is a good way to see whether there is a problem worth looking at. We explain how to read the whole scorecard in our guide to the free scorecard.
If you start a new scan for the same business within 30 days, no new scan runs; you are shown a link to your existing scorecard. A single scan repeated after a short gap can show the normal variation in AI answers more than a real change anyway. To see real change, the same questions have to be asked in the same way, regularly. If you want to follow where you stand against your competitors on five engines, week by week, see the plans.
No one can guarantee an answer
Be careful with anyone who promises to "put you first on ChatGPT". No one outside the AI companies decides what the assistants say, and that includes us. Our job is to measure, and to give you an honest picture of where you stand over time.