Is ChatGPT recommending your jewellery business? A monthly check you can run yourself

The way to find out is to run the same check every month. Write a fixed list of real customer questions. Ask each one of each assistant. Record every answer as named, cited or absent. Then compare the results from one month to the next. One answer proves nothing. A pattern across months shows you where you stand, and whether anything you changed made a difference.

This is the method I use, in steps you can follow tonight. You need a spreadsheet and a few evenings.

Why one try tells you nothing

Ask an assistant the same question twice and the answers can differ. Change the wording slightly and they differ more. Ask from another country, on another day, or from an account with a long history, and they differ again. A screenshot of one good answer, or one bad one, is an anecdote.

A fixed list, asked the same way every month, reduces that noise. You are not trying to learn what an assistant “thinks” of your business. You are checking whether, across a fair spread of questions, you show up, and whether that changes over time.

I built a monthly check like this in October 2026. It puts a fixed list of customer questions to ChatGPT with web search switched on and to Gemini with Google Search switched on, records each answer, and repeats on a schedule. What follows is the same method, done by hand.

Step 1: write 10 to 25 questions in a customer’s words

Fewer than ten and one odd answer swings the whole picture. Many more than twenty-five and the job gets long enough that you will skip it. Somewhere in between is right.

Spread the questions across six types:

• Bridal. What people ask before they buy a ring.
• Gifts. Occasions, budgets, who it is for.
• Materials. Metals, stones, what suits everyday wear.
• Care. Cleaning, repairs, resizing, looking after a piece.
• Price. What things cost and what is reasonable.
• Local. Anything with a town or “near me” in it.

Here are generic examples only, to show the shape. Write your own.

• Bridal: “best UK jeweller for a sapphire engagement ring”
• Gifts: “jewellery gift for my mum’s 60th birthday under £300”
• Materials: “is 9ct gold or sterling silver better for earrings I wear every day”
• Care: “how to clean a ring with small stones at home”
• Price: “how much should I spend on a gold wedding band”
• Local: “where can I get a ring resized in [your town]”

Four rules keep the list useful.

1. Type each question the way a customer would. Not “bespoke jewellery provider UK”, but the slightly clumsy thing people actually type.
2. Leave your business name out of most of them. A customer who already knows your name does not need an assistant to find you. Do not add a “site:” search either. Keep one or two questions that do name you, to see whether the assistant knows you exist and gets the basics right.
3. Take the words from real life. Your enquiry emails and messages, what people ask at fairs and at the counter, what your reviews mention. Paraphrase and strip out names. Never paste a customer’s actual message into a sheet you share.
4. Number the questions, date the list and freeze it. Add new questions at the bottom. Never edit an old one, or you can no longer compare this month with last.

Step 2: choose the assistants and keep the conditions the same

Start with two: ChatGPT with search switched on, and Gemini. Add others later if your customers use them.

Then keep the conditions as fixed as you can.

• Use a new chat for every question. An earlier question can colour the next answer.
• Treat personalisation as a variable. The apps carry your history and memory, which can colour an answer. Where the settings let you, switch memory off, or use a private window. My own check runs through the developer interfaces, which carry no history, so my answers and an app’s answers will not match exactly. Neither is wrong. Just use the same route every month.
• Set the place. If an assistant lets you choose a location, set it to the UK, or to your town for the local questions. One of the two I use cannot be given a location at all, so I note that beside its results.
• Ask once per question per month. If you want to see how much an answer wobbles, ask a handful twice on the same day and record both, then go back to once.
• Keep Google’s own AI summaries separate. Those appear in Google’s results, and my check does not cover them. Search your questions in Google on a separate pass and note whether a summary appears and whether you are linked in it.

Step 3: the recording sheet

Copy this into a spreadsheet. It is a template. The cells are empty on purpose, and it holds no results.

Record the sources in the order the answer lists them. The “first three sources cited” column matters most, even on the days you are absent. It tells you who is winning the question, and what sort of page they use to do it.

Step 4: how to read a result

Each answer gets one of three labels.

Named. Your business name appears in the text of the answer. The assistant knows you exist and thinks you are relevant to that question. Read what it says about you as well. If it gives the wrong town, price or product, write that in the notes. A wrong fact is worth fixing at its source.

Cited. One of your pages appears in the list of sources. Your page answered that question well enough to be used. Note which page it was. The place in the list is a rough signal at most. Do not read much into first versus third.

Absent. Neither of the above. For a small business this is common, especially at the start. It is a baseline, not a verdict on the quality of your work.

The mixed cases tell you the most.

• Cited but not named. The assistant used your page but did not say who you are. Still a good sign: your page did the work.
• Named but not cited. The assistant got your name from somewhere other than your own pages, such as a directory, a review site or a press feature. Find out where. That page now speaks for you, and it needs to be accurate.
• Absent, with others cited. Open the sources it did use. What do those pages have that yours lack?

Some tools also show pages the assistant looked at but did not use. My own check records those separately. You will only see them where the tool lists its searches, but if you do, note it. It means your page was reachable and another one was chosen.

Judge by question type, not by single questions. Three months of movement in the bridal questions means something. One flip in one question does not.

Step 5: what to do with the gaps

1. List the absent questions by type. If every price question is absent and the care ones are fine, you have a price problem, not an AI problem.
2. Open the sources that were cited instead. Look for what they do. They usually answer the exact question near the top, in plain words, with facts a customer needs: a price range, a timescale, a material, a place.
3. Improve or write one page per gap. One page that answers one question properly beats a general page that touches ten. Start with the two question types where you matter most.
4. Check the plain route to your site is open. See the next section.
5. Wait for the next month. Do not rewrite everything because of one result.

What does not move the answer, and what might

Nobody outside the assistants’ companies knows exactly how they choose their sources, and I will not pretend to. Here is what is established, what I tested, and what is only my advice.

A special file or extra markup for AI. Google says so directly in its page on AI features and your website (last updated 10 December 2025): “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” And: “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.”

In our own test, we checked our own sites. They let the AI crawlers in, with open robots rules, structured data and even a text file written for AI crawlers. Having all of that in place did not, by itself, get us named for most of the buyer questions we asked. That matches what Google says. Do not spend your month on a file.

The one technical switch that can remove you. OpenAI’s page on its crawlers says OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features. A site that opts out of OAI-SearchBot will not be shown in ChatGPT search answers, though it can still appear as a navigational link. Changes to robots.txt take about 24 hours to apply. So type your web address followed by /robots.txt into a browser and look for OAI-SearchBot. If you find it blocked and did not mean to, ask whoever built your site to remove the block.

Being findable in ordinary search. On Google’s page, a page must be indexed and eligible to be shown in Google Search with a snippet in order to be a supporting link. Check that the page you want quoted is indexed, and that nothing in your site settings keeps it out.

What might help. This part is advice, not proof. The monthly check is how you learn whether it works for you.

• A page that answers the exact question, plainly, with the facts a customer needs.
• Accurate mentions on other sites that the assistants already read. Your “first three sources cited” column shows you which sites those are for your questions.
• Bing Webmaster Tools. Bing announced an “AI Performance” report in February 2026 that counts how often a verified site is cited in Microsoft Copilot and in Bing’s AI summaries. It counts citations, not visits, so use it as a second opinion on your monthly sheet and not as a traffic number.

Your first month

• Evening one. Write 10 to 25 questions in six types. Number them, date the list, freeze it. Set up the sheet.
• Evening two. Run every question once on each assistant, in a new chat, with the same settings. Fill in the sheet as you go. Do not stop to fix anything yet.
• Evening three. Read the sheet by question type. Mark the gaps. Open the sources that were cited instead and write down what they do that you do not.
• Week two and three. Write or improve one page for each of your top two gaps. Check your robots.txt and that those pages are indexed. Verify your site in Bing Webmaster Tools.
• First of next month. Same list, same settings, new rows in the same sheet. Compare by question type.

If the first month is mostly “absent”, that is your baseline. The point of the sheet is not a score. It is to know, in a month or three, whether what you changed made any difference.