How to Build a ChatGPT Prompt Tracking Sheet

Make one tab called Runs. Every answer you collect becomes one row, and every row has the same twelve columns. One row per answer matters: if you summarise as you go ("named 2 out of 3"), you lose the detail you need later to check a count or spot a pattern.

Keep the brand-named and cited columns strictly yes/no, and decide in advance what counts. A mention is your brand or domain named in a form nobody could confuse with something else; a citation is a source link to your own domain. Why the two must stay separate is explained in what ChatGPT citations are.

Here are the three OpenAI rows Leadee's tracker has for CMSmart (cmsmart.net, part of Netbase JSC, the company behind Leadee) from its 2026-10-25 round. They are aggregated per question, because that is how our analysis stored them; a sheet you keep yourself should have one row per run. Cells our analysis did not keep at that level are marked "not stored" rather than filled in.

At engine level the analysis does record more: CMSmart was cited in 3 of its 4 US OpenAI runs, with an average position of 3.67. Across all 25 OpenAI runs for the 7 Netbase JSC brands, a brand was named in 3 (12%, 95% Wilson interval 4.2-30.0%); the round stopped when the OpenAI account ran out of credit. That is an indicator, not a measurement, and it is exactly why the sample-size section below matters. The full brand results are in the CMSmart case study.

Put your prompts on a second tab called Prompts, each with its shape label, so the Runs tab can look them up. Most of the list should be unbranded prompts: questions a buyer asks before they know you. Keep branded prompts as a small separate group, because they make almost any brand look visible.

If you do not have a list yet, build one from sales emails, site search and Search Console first, as described in how to find the prompts your customers ask ChatGPT, then add candidates from the free ChatGPT Prompt Finder. Ten to twenty prompts is a workable start.

Ask each prompt at least three times per round, in a fresh chat each time, and fill in one row per answer straight away. Paste the cited URL exactly as shown. If an answer names a company with the same or a similar name as yours, write it in Notes and leave Brand named as "n" until you have checked it by hand.

The method for asking — which questions, how often, and how to avoid your own chat history steering the answer — is in how to check if ChatGPT mentions your brand.

On a Summary tab, count per prompt shape and per period: runs (n), brand named (k), and the rate k/n. Then add the 95% Wilson interval next to every rate. With k in B2 and n in C2, the lower bound is =(B2/C2+1.96^2/(2*C2)-1.96*SQRT(B2/C2*(1-B2/C2)/C2+1.96^2/(4*C2^2)))/(1+1.96^2/C2); for the upper bound, change the minus before 1.96*SQRT to a plus.

Do the same for cited (y/n) and keep an average position only over the runs where you were named. Leadee's tracker reports the same metrics with the same interval method; the definitions are on our methodology page.

More than most sheets have. The table below is plain arithmetic, not a measurement: it shows the 95% Wilson interval for the same observed rate of 30% at different sample sizes. With 10 runs, a 30% rate is compatible with anything from about 11% to 60%. With 40 runs, about 18% to 45%. Quadrupling the runs roughly halves the width.

In practice: 10 prompts asked 3 times each gives 30 runs per round, and four weekly rounds give 120. That is enough to see a large change, such as 10% becoming 40%, but not a small one. Two periods have only clearly changed when their intervals no longer overlap.

Run the same prompts on the same weekday every week or two, with the same engine and market. Change the prompt list only on purpose, and log the change in Notes. When you fix something on your site or get listed on a page ChatGPT cites, note the date, then compare the periods before and after, with intervals.

The market players column becomes useful over time: count how often each one is named for each prompt shape. It shows who takes the place you want, and the cited URLs show which pages you need to be on. The other ChatGPT guides are on the ChatGPT SEO overview.

A sheet works well for one brand, one market and about ten prompts. It gets hard when you add a second engine, a second country, or a weekly cadence: three engines times two markets times 20 prompts times 3 runs is 360 rows a round, typed by hand. Errors creep in at the same time — a missed citation, an ambiguous name counted as a mention.

At that point automated tracking does the same job through each engine's official API, on a schedule, with the intervals computed for you. Leadee is one option (see pricing); whichever you choose, keep the same column definitions so your old sheet and the new data can be compared.

Pick ten random rows each month and re-check them against the saved answers: is the brand really named, is the cited URL really your domain, is the position right? If more than one in ten is wrong, tighten the rule in your column definitions. Also check that every prompt has the same number of runs per round; uneven runs skew the rates.