Run a fixed set of 10 to 15 prompts about the client, three times each, in clean ChatGPT sessions once a week. Log whether the brand is named, whether the facts are right and which sources are cited. Weekly suits a client in active work; monthly is enough when nothing is changing. Compare rates over four weeks, not single answers.
Tool vendors such as Ahrefs already explain how to track prompts with software. What they leave open is the human routine: which prompts, what to write down, when to check and what to do when the answer is wrong. That routine is the same whether the results land in a spreadsheet or a paid tool: twelve frozen prompts, a run protocol, one row per answer, a weekly slot in the diary and a written rule for what happens when the answer is wrong.
Why weekly, and why not every day by hand
ChatGPT's answers move for reasons you cannot see: a new page gets indexed, a listing changes, the model rewords. OpenAI's own help page says search results and citations can be incomplete, outdated or incorrect, so any single answer is a weak signal. A weekly rhythm balances two needs. It is frequent enough to catch a wrong fact within days and to see whether a fix landed. It is infrequent enough that a person can do it in about an hour and a half without drifting into watching noise.
Daily manual runs produce more rows but not more insight, and they invite overreaction to one odd answer. If a client needs daily checks, that is the point at which paying for a tool starts to make sense. For clients in quiet months, drop to monthly and keep the same prompt set so the history stays comparable.
Step one: build the fixed prompt set
The prompt set is the instrument. Change it casually and your trend line breaks, so agree it with the client, freeze it, and version it. Use four types, three prompts each, giving 12.
Here is an illustrative set for a small driving school client, which serves learners and nervous returners. Replace the details for your own client.
| # | Type | Prompt |
|---|---|---|
| 1 | Name | What does [school name] do, and is it a good choice? |
| 2 | Name | What do customers say about [school name]? |
| 3 | Name | How much do lessons cost at [school name]? |
| 4 | Category | Which driving schools are best for nervous learners in [town]? |
| 5 | Category | Where can I take intensive driving lessons near [town]? |
| 6 | Category | Recommend a driving instructor for someone over 40 in [town]. |
| 7 | Question | How many lessons does a nervous learner usually need? |
| 8 | Question | What should I do before my first driving lesson? |
| 9 | Question | Is an automatic lesson cheaper than a manual one? |
| 10 | Comparison | [school name] versus [competitor]: which is better for beginners? |
| 11 | Comparison | Is a big driving school better than a one-instructor school? |
| 12 | Comparison | What should I ask before booking a block of lessons? |
Why these types? Name prompts test whether the model knows the brand and has the facts right. Category prompts test whether the brand appears when a stranger describes the need, which is the real commercial test. Question prompts test whether the client's content is cited as a source. Comparison prompts show how the brand is positioned against rivals. Ask the client for the three competitors that matter and fix them in the set. Keep wording natural, the way a customer would write, not keyword-shaped.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep two: fix the run protocol so results are comparable
Most monitoring noise comes from changing how you run, not from the model. Write these rules down and follow them every week.
- Same account and plan each time, recorded in the log. If you use a work account with saved memory, personalisation can colour answers. OpenAI's help page says a temporary chat asks you to choose a personalised or unpersonalised mode before the first message, that the unpersonalised mode does not use memory, custom instructions or plugins, and that neither mode creates or updates memories. Pick the unpersonalised mode for every monitoring run; it cannot be changed once the chat has started.
- A new chat for every run. A follow-up inside an old chat inherits its context.
- Make sure web search is on. OpenAI's help page says ChatGPT may search automatically when a question benefits from current information, and can be set to search manually. Note whether sources appeared; a run with no search is a different kind of answer.
- Three runs per prompt, same day. Twelve prompts times three is 36 runs. That is a lot to do by hand, so do the runs in one sitting and paste results straight into the sheet.
- Record where you were. OpenAI says ChatGPT can use location information for local results, so note the device, location setting and date. Do not change them mid-series.
- Do not edit the prompts. If you must change one, give it a new number and start a new row series.
Keep the log free of anything private. Record the client's brand and the model's answers; do not record your own other chats, and use the work account only for monitoring so nothing personal sits in the same history.
Step three: the log sheet
One row per run. Columns that earn their place:
| Date | Prompt # | Run | Brand named (Y/N) | Position (1, 2, 3 or 0) | Facts correct (Y/N/Part) | Client URL cited (Y/N) | Other sources | Notes |
|---|---|---|---|---|---|---|---|---|
| 3 Nov | 4 | 1 | N | 0 | n/a | N | Two review sites | Three rivals named |
| 3 Nov | 4 | 2 | Y | 3 | Y | N | Directory | |
| 3 Nov | 4 | 3 | N | 0 | n/a | N | Local guide | |
| 3 Nov | 3 | 1 | Y | 1 | Part | Y | Client site | Price out of date |
(Illustrative rows.) A weekly summary sheet then calculates three numbers per type: the mention rate (rows with Y divided by all runs), the accuracy rate (Y divided by named rows) and the citation rate (client URL cited divided by all runs). In a spreadsheet the mention rate is a one-line formula such as =COUNTIFS(D:D,"Y",B:B,">=4",B:B,"<=6")/COUNTIFS(B:B,">=4",B:B,"<=6") for the category prompts. Illustrative week-one summary:
| Type | Runs | Mentions | Mention rate | Accuracy |
|---|---|---|---|---|
| Name (1 to 3) | 9 | 8 | 89% | 5 of 8 correct |
| Category (4 to 6) | 9 | 1 | 11% | 1 of 1 |
| Question (7 to 9) | 9 | 0 (cited: 1) | 0% | n/a |
| Comparison (10 to 12) | 9 | 2 | 22% | 1 of 2 |
The pattern is typical: well known by name, almost invisible for category questions, and some facts wrong. That tells you where to point work: fix the wrong facts first, then build pages that answer the question and category prompts. How ChatGPT finds a business and what a firm can control explains which levers exist, and SEO for AI search: a client audit checklist for agencies turns them into an audit.
The weekly schedule, with minutes
| Day | Task | Minutes |
|---|---|---|
| Monday | Run all 36 prompts in unpersonalised temporary chats; paste answers and sources into the log | 45 |
| Tuesday | Score the rows: named, position, accuracy, citations; flag every wrong statement | 20 |
| Wednesday | Trace each flagged error to its source; list fixes with an owner | 15 |
| Thursday | Update the summary sheet; write three lines for the client | 10 |
| Friday | Send the note; log any fixes made so next week's rows can be read against them | 5 |
That totals 95 minutes. At an internal cost of $50 an hour, that is about $79 a week per client, or roughly $340 a month, before any fix work. This sum is why many agencies batch clients and why a tool starts to pay off past four or five of them. Time your own first week and replace my estimate with your number.
Prompt-tracking tools, checks and monthly cost
The manual method costs time. Tools shift cost to a subscription and give you history without effort. As read on 4 November 2026, Ahrefs sells prompt tracking inside Brand Radar. Its pricing page lists custom prompt packages from $50 a month (Basic: 2,500 extra checks a month, described as roughly 80 prompts daily on one platform, with overage at $0.020 a check), then $100 for 7,000 checks and $250 for 25,000. It defines a check as one prompt on one platform in one location. The Brand Radar page says every paid Ahrefs plan includes a Custom Prompts quota (5 prompts on Lite, 10 on Standard, 20 on Advanced), that you can pick daily, weekly or monthly cadence per platform, and that Claude uses 8 checks per update rather than one.
Work out checks before you buy. Twelve prompts on one platform, checked daily, is 12 times 30, which is 360 checks a month. Twelve prompts on three platforms is 1,080. If you check weekly instead, 12 prompts times four or five weeks is 48 to 60 checks per platform. A 12-prompt set is more than the Lite quota of 5 prompts, fits inside Advanced's 20, and uses under half of a Basic package's 2,500 checks even when run daily on three platforms (1,080 checks). Prices and quotas change, so confirm on the vendor's page. Semrush also lists AI visibility tracking and custom prompt monitoring on its plans; its pricing page is the place to check current limits. Semrush AI Visibility versus Ahrefs Brand Radar compares the two for an agency, and AI visibility tracking for B2B on a budget covers cheaper routes.
| Option | Monthly cost for one client | Best when | Weakness |
|---|---|---|---|
| Spreadsheet, manual runs | Time only, about $340 internal at the sum above | One to three clients, weekly checks | Slow; human error in logging |
| Prompt tracking add-on | From $50 a month for a package; check whether your account also needs a plan | Four or more clients, daily or multi-platform checks | Prompts still need writing and checking; results still vary |
| Hybrid | Package plus about 30 minutes a week | You want history from the tool and judgement from a person | Two systems to keep in step |
Whichever route you choose, keep the human review. A tool can tell you the brand was named; it cannot tell you the price quoted was last year's.
The first week: a baseline and a conversation
The first run is not a result; it is a baseline. Say so to the client before you send it. Explain that week one shows where the brand stands today, that numbers will wobble for a few weeks, and that you will report trends from week four. Two things are worth doing in that first week beyond the prompts themselves. First, ask the client which questions customers really ask, because they know the phrasing; swap any of your twelve for their real ones before you freeze the set. Second, agree the list of facts that count as "correct": prices, hours, services, locations, names of key people. Without a written list, accuracy scoring turns into opinion, and the client will dispute a "Part" that you marked on a hunch.
Keep the baseline file untouched. When the client asks in month six whether things have improved, you will want to open the original rows rather than rely on memory. A frozen baseline, a versioned prompt set and a dated list of fixes together are the paper trail that makes your report credible.
Running the routine across several clients
With one client the routine is easy. With six it needs structure. Give every client a copy of the same workbook, named the same way, with the prompt set on one tab, raw rows on another and the summary on a third. Run clients in a fixed order so the weekly slot is predictable, and block the time in a calendar rather than fitting it in. Use one browser profile for monitoring only, kept separate from anything personal, and clear nothing: the point is a consistent setup.
A realistic split for an agency with six clients: one person owns the Monday runs for all six, another scores and traces errors on Tuesday and Wednesday, and the account manager writes the Thursday notes. That separation also catches mistakes, because the person scoring did not run the prompt. At six clients and 95 minutes each, the weekly load is about 9.5 hours, which is the point where the comparison in the tool table above usually tips towards a paid tracker for the volume work and a person for the reading.
Reading the log without fooling yourself
A mention rate based on 9 runs jumps in whole steps of 11 points, so do not read a move from 11% to 22% as a trend. Rules of thumb I use, which are judgement rather than statistics:
- Compare four-week averages, not weeks.
- Treat a change as real only if it holds for at least three weeks and coincides with a known cause, such as a fixed listing or a new page.
- Always read accuracy next to the mention rate. A rising mention rate with falling accuracy is a problem, not a win.
- Note external events: a competitor's launch, a press mention, a site migration.
An illustrative month-one result for the driving school: name prompts hold at about 90%; category prompts rise from 11% to 22% after the client's page on nervous learners goes live and a directory entry is corrected. With nine runs a week that is only a few answers, so the report says "encouraging, not proven" and plans another four weeks. Honest wording protects the relationship better than a confident chart.
When ChatGPT gets the brand wrong
Wrong answers are the most valuable finding in the log, because they have causes you can fix. Triage each one.
| Type of error | Typical cause | Action |
|---|---|---|
| Old price or hours | An outdated page or listing still cited | Update the source; add a dated "prices checked" line on the client's own page |
| Confused with a similar brand | Shared name or category on a directory | State the full name, town and founding year on the About page; fix the directory entry |
| Service not offered is described | An old article or a competitor comparison page | Ask the owner to correct; publish a clear services page |
| Negative claim with a source | A review or forum thread | Reply factually where the platform allows; do not argue or fabricate reviews |
| Claim with no source | The model's own wording | Log it; check again next week; strengthen pages that state the true fact |
The routine for each error is the same: copy the wording and the prompt into the log, open the cited sources, find the origin, fix at the source, note the date, and re-test after one to two weeks. OpenAI's help page says that if a result looks wrong, you can ask ChatGPT to search again using a specific source, date or location; that is useful for confirming a fix has been picked up, but it does not change what other users see. Reporting an answer through ChatGPT's own feedback options is worth doing if it is available, though I found no stated promise of a correction, so rely on fixing sources. For stale facts in particular, how to catch outdated information in AI answers adds checks.
A sample message to a site owner who hosts a wrong claim (illustrative, no names):
Hello [first name], I'm writing on behalf of [client]. Your page on
[topic] lists our opening hours as 9 to 5, Monday to Friday. Since March we
have also opened on Saturdays, 9 to 1. Could you update the listing? The
current hours are on our contact page: [URL]. Thank you, [your name]The monthly note for the client
Turn four weeks of rows into one short message. Four parts are enough: the headline numbers by prompt type, the wrong statements found and their status, what changed on the client's side, and the plan for next month. Keep it plain. An illustrative paragraph: "Across four weeks we ran 144 checks. The brand was named in 89% of name prompts and 17% of category prompts. Three wrong statements were found; two are corrected at source, one awaits a directory update. Next month we will publish the nervous-learners page and re-check." Do not promise recommendations; say what you did and what you saw. For turning this into a repeatable client document, see drafting monthly client reports with ChatGPT or Claude.
Mistakes that make the data useless
- Changing the prompts. A reworded prompt is a different test. Version it.
- Logging impressions, not rows. "It seemed to mention them more" is not data. Write down every run.
- Only checking the name prompt. It always looks good and hides the category result.
- Running while logged into a personalised account. Memory and history colour answers; use an unpersonalised temporary chat every time.
- Reporting a single answer to the client. One screenshot can be a flattering or alarming outlier.
Done for a month, the routine yields 144 data points per client, a list of fixable errors and an honest sense of whether the work is moving anything. That is more than most reports built on a single query ever show.
Questions about monitoring ChatGPT mentions
Can I just search my client's name once a month?
A single name search tells you little, because answers vary between runs and wording. A fixed set of prompts, each run several times and logged, gives a trend. Monthly is enough for a quiet brand; weekly suits a client who is changing pages, running PR or fixing wrong facts.
Why run each prompt more than once?
Because the same prompt can produce different answers on different runs, with different sources. Three runs per prompt gives a mention rate rather than a yes or no. Keep the number of runs the same every week so figures stay comparable, and record the date.
Do I need a paid tool?
Not to start. A spreadsheet and 60 to 90 minutes a week handles a dozen prompts. A tool becomes worth paying for when you track several clients, need daily checks or want several AI platforms in one report. Check current prices and limits on the vendor's own page before you quote them.
What if ChatGPT says something false about the client?
Open the answer's sources, find where the claim came from, and correct it there if you or the client control the page. Log the wording, date and prompt. Re-test after a couple of weeks. You cannot instruct the model to remember a correction, so fix the sources it reads.
Further reads
- PR for AI Search: How Agencies Earn Citations in ChatGPT Answers — Earn the third-party sources AI answers cite.
- How to Track Whether AI Search Is Sending You Customers — Check whether AI answers send visitors and enquiries.
- How to Improve AI SEO: Getting Your Firm Cited in AI Answers — A citation checklist for the fixes your log points to.
- Best AI Visibility Service Providers: What a Small Firm Should Ask — What to ask before paying a visibility provider.
- Search Console and AI Overviews: What Agencies Can Report to Clients — What Search Console can report on Google's AI features.
- AI Brand Audit: A Template Branding Studios Can Sell — A ten-day brand audit for a music school, step by step, with the 12-page client deliverable a branding studio can reuse and sell.
- AI Schema Markup Generator: Validate Client Structured Data Before Launch — A pre-launch routine for AI-drafted JSON-LD: prompts that stop invented data, then validator, Rich Results Test and Search Console checks.
- AI Social Media Monitoring: A Monthly Insight Report for Clients — Lead with the monthly insight report: a template, a filled example, alert thresholds, a sentiment spot-check and what listening tools cost for five clients.
- How to Advertise on ChatGPT: Approval, Budget and Reporting for Agencies — Run ChatGPT ads for a client with sign-off, a capped test budget, tracking and a weekly report, based on OpenAI's own Ads Manager documentation.
- Does llms.txt Work? What SEO Agencies Should Tell Clients — What Google says about llms.txt, what evidence exists elsewhere, and the exact wording an agency can give a client who asks for one.
- Best AI Brand Visibility Software: Is AI Mentioning Your Business? — Check by hand first. Paid AI visibility trackers start around $29 a month, and only some small businesses need one. Prices, limits and a test.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Ahrefs Brand Radar and pricing pages, Semrush pricing page, OpenAI Help Centre pages on ChatGPT search and temporary chats (all read 4 November 2026). Client, figures and log values are illustrative.