You can check by hand and you can check systematically. By hand: you open a fresh window with no history, ask five to ten questions in Hebrew that a real customer would ask — without naming your business — and record who is mentioned and in what position. Systematically: you run the same questions against several engines, several times, and measure mention rate, average position and share of voice against competitors. The difference between the two is not accuracy but stability: one answer is noise, dozens of answers are a measurement.
The step everyone misses
Before the five steps, the sentence that decides whether the check is worth anything:
A question that includes your business name does not measure visibility.
It sounds pedantic and it is the difference between a real number and an inflated one. Ask "מה אתם יודעים על משרד לוי?" ("what do you know about the Levi firm?") and the model will mention the Levi firm — it was asked to. Ask "מי עורך הדין הכי מומלץ לדיני עבודה בחיפה?" ("who is the most recommended employment lawyer in Haifa?") and the answer says something.
The first is called a brand question and it measures reputation: what is said about you once people already know who you are. The second is called a discovery question and it is the only one that measures whether you get found at all.
Five examples of good discovery questions:
- מי עורך הדין הכי מומלץ לדיני עבודה בחיפה? ("who is the most recommended employment lawyer in Haifa?")
- איזה משרד תיווך הכי טוב בבאר שבע? ("which estate agency is the best in Be'er Sheva?")
- מחפש יועץ לקרן השתלמות, מי הכי טוב בשוק? ("looking for an adviser for a keren hishtalmut savings fund — who is the best on the market?")
- למי כדאי לפנות אחרי פיטורים בלי שימוע? ("who should I turn to after being dismissed without a hearing?")
- מי מומחה לביטוח סייבר לעסקים קטנים בישראל? ("who is an expert in cyber insurance for small businesses in Israel?")
And two that will not count, however tempting:
- מה אתם יודעים על משרד לוי ושות'? ("what do you know about Levi & Co.?")
- האם משרד לוי הוא עורך דין טוב? ("is the Levi firm a good lawyer?")
Where to record it
One table is enough. A column for the question, a column for the engine, a column for who was mentioned, a column for the position you appeared in:
| Question | Engine | Who was mentioned | Your position |
|---|---|---|---|
| מי עו״ד מומלץ לדיני עבודה בחיפה? ("recommended employment lawyer in Haifa?") | ChatGPT | Cohen, Levi, Barak | 2 |
| מי עו״ד מומלץ לדיני עבודה בחיפה? ("recommended employment lawyer in Haifa?") | Gemini | Cohen, Avni | did not appear |
After twenty rows like that you have something to count.
What do you measure at the end?
Three numbers, and none of them is a "ranking":
- Mention rate — in how many of the answers you appeared at all.
- Average position — when you appeared, where in the list of names.
- Share of voice — what part of all the mentions in the industry belongs to you compared with the competitors.
The third is the only one that says anything about the market, and it is also why step 5 is not optional. The full explanation of the three metrics is here.
Two things that are easy to misread
"Nobody was mentioned" is not the same conclusion as "you were not mentioned". If the engine answered in principles without naming names, you did not lose to a competitor — there is simply no answer with names in it yet. That is an opportunity and not a failure, and the strategy is the exact opposite.
If every question you managed to draft includes your name, you have no measurement. Not a low score — no score. It sounds like pedantry until you see a score of zero built entirely out of brand questions, and someone drawing a business conclusion from it.
There is a free Google tool, and it is worth starting there
Before you pay anyone — us included — there is a free data source that comes from the provider itself.
Google Search Console has a performance report for AI features that shows how many times links to your site were shown in AI Overviews and in AI Mode, by pages, countries, dates and devices. If your site is verified in Search Console, that is half an hour worth spending before anything else.
Three caveats that are written in the documentation itself: Google is rolling the report out gradually and not every site owner has it yet; a site that has not accumulated enough impressions simply will not see data; and data from Search Labs experiments is not included.
What this report does not see
This is the part that decides whether it is enough for you, and it is not a matter of quality but of definition:
- It counts impressions of a link, not mentions of a name. The metric is how many times a link to your site was shown in an AI feature. That is not the same question as "did the engine name my business in the answer" — you can be mentioned in the text without a link, and you can be linked without the name being said.
- There is no query dimension. The documentation lists four dimensions: pages, countries, dates and devices. Queries are not among them, so you cannot tell which question the impression came from — and the whole guide above is built on exactly that question.
- There are no competitors. The report shows your site. Share of voice, the only one of the three numbers that says something about the market, is not there and cannot be there.
- It covers Google alone. AI Overviews and AI Mode are Google features. On ChatGPT, on Claude and on Perplexity the report is completely silent, and that is a large part of what Israelis actually ask.
So the right order is: the free report first, then the table above. The report tells you whether Google shows you at all; the question list tells you in which questions, against whom, and in what position — and in the engines Google has no opinion about.
When is it worth moving to an automated check?
When the manual check turns into a weekly task.
For a one-off check, or one a quarter, the table above and half an hour do the job. There is no reason to pay for it. The problem starts when you want to see a trend: the same questions, the same wording, every week, in two engines, with a count of competitors. At that point manual work stops being accurate — not out of carelessness, but because people change the wording without noticing, and then the comparison between the weeks no longer says anything.
We did this by hand enough times to end up building it. The check is free and takes about a minute.
Five steps for a manual check
Draft discovery questions
Write five to ten questions a real customer would ask, in Hebrew, without your business name. For example: "מי עורך הדין הכי מומלץ לדיני עבודה בחיפה?" ("who is the most recommended employment lawyer in Haifa?"). A question that includes your name does not count — it asks the model to repeat the name, and it always can.
Ask in a fresh window with no history
Open a new conversation, and if possible in a mode with no history. A model that has already talked to you about your business will mention it again — not because you get found, but because you are in the context of the conversation.
Record who is mentioned and in what position
For every question write down: the question, the engine, which businesses were mentioned, and in what order. The order matters — being first in the list is not the same as being fifth.
Repeat the same questions
Run the same questions again, on a different day. Answers vary from one run to the next, which is why a single answer is noise and not a measurement.
Check the competitors too
Count how many times each competitor was mentioned in the same questions. A number with no competitors beside it is close to meaningless: a 30% mention rate is excellent in an industry where everyone sits around 20%, and dismal against a competitor at 70%.
Frequently asked questions
- How do I check whether ChatGPT mentions my business?
- You ask five to ten questions in Hebrew that a real customer would ask, without naming the business, in a fresh conversation window — and you record who is mentioned and in what position. You repeat it on a different day, because a single answer is noise.
- Why must the business name be left out of the question?
- Because then you are not measuring visibility. A question like "מה אתם יודעים על משרד לוי?" ("what do you know about the Levi firm?") asks the model to repeat the name, and it always can. That measures reputation — what is said about you — and not whether you get found in the first place.
- How many questions does it take for this to be reliable?
- For a manual check, five to ten questions in two engines, twice. That is enough to see whether you are mentioned at all. To see a trend over time you need dozens of questions and a regular run, and that is no longer manual work.
- Which engines are worth checking?
- ChatGPT and Gemini are the engines consumers in Israel actually use, and they are also what we measure. Perplexity is worth a look if your industry is technology. Google's AI summaries — AI Overviews — appear for some of the queries and are worth checking too.
- What do you do if nobody is mentioned in the answer?
- That is not the same conclusion as "you are not mentioned". If the engine answered in categories and principles without naming names, nobody won — and whoever gets into that space first gets in unopposed.
- Is there a free Google tool that checks this?
- There is, and it is worth starting there: Google Search Console has a performance report for Google's AI features, which shows how many times links to your site were shown in AI Overviews and in AI Mode. Three caveats: Google is rolling it out gradually and not every site has it yet, it counts impressions of a link to the site rather than mentions of the business name, and it covers Google alone — not ChatGPT and not Perplexity.
- If Google has a report, why check by hand at all?
- Because the report answers a different question. The dimensions the documentation lists are pages, countries, dates and devices — there is no query dimension, so you cannot tell which question the impression came from, whether you were named in the answer, and who else was mentioned alongside you. The manual question list is what gives you those three things.
- When is it worth moving to an automated check?
- When the manual check turns into a weekly task. For a one-off or quarterly check, a table and half an hour do the job without paying anyone.