On this page
Most founders discover their ChatGPT visibility by accident. Someone in a Slack group asks the model for tool recommendations in your category, screenshots the answer, and your product is not in it. That is not monitoring, that is luck. Proper ChatGPT brand monitoring means running a fixed set of buyer questions on a fixed schedule and watching the numbers move.
This guide covers how to build that setup, what to log, and how to tell a real drop from ordinary noise.
What you are actually measuring
Google rankings are positional and mostly stable. Ask for your position today and tomorrow, and you get roughly the same answer. Large language models do not work that way. Ask ChatGPT the same question five times and you can get five different lists of tools, in different orders, with different names.
So the unit of measurement is not a rank. It is a mention rate: across N runs of the same prompt, what share of answers include your brand. A product mentioned in 4 of 10 runs of “best AEO tool for small teams” has a 40% mention rate on that prompt. That number is comparable week to week, comparable against competitors, and it moves when you change something on your site.
Three secondary metrics are worth capturing alongside it:
- Position in list. Being named first in a list of eight is very different from being named eighth, even though both count as a mention.
- Sentiment. “X is a solid budget option” and “X is often criticized for its limited integrations” are both mentions. Only one of them helps.
- Cited sources. When browsing is on, ChatGPT pulls from live pages. Those URLs tell you exactly which content the model trusts in your category, which is the most directly actionable thing in the whole dataset.
Why one-off checks mislead you
The single most common mistake is asking ChatGPT once, seeing your brand, and concluding you are visible. Or asking once, not seeing it, and concluding you are invisible. Both conclusions are drawn from a sample size of one on a probabilistic system.
If your true mention rate on a prompt is 30%, a single check has a 70% chance of showing you nothing at all. Run it three times and you will probably see it once. This is why multi-sampling matters: repeat each prompt several times per measurement period and record the ratio, not the outcome.
Three to five samples per prompt per run is the practical sweet spot. Below three, noise dominates. Above five, you are paying for precision that will not change any decision you make.
Step 1: build the monitored prompt set
Your prompt set is the whole experiment. Get it wrong and everything downstream measures nothing useful.
Write prompts the way a buyer talks, not the way a marketer writes keywords. Four categories cover most of the real demand:
- Category recommendation. “What is the best invoicing tool for freelance designers?”
- Alternatives. “What are good alternatives to Freshbooks for a solo consultant?”
- Problem-first. “How can I automatically chase unpaid invoices without hiring anyone?”
- Comparison. “Compare Wave and Invoice Ninja for a two-person agency.”
Problem-first prompts are the ones founders skip most often, and they are frequently where the softest competition sits. Nobody is optimizing for them yet.
On volume: 20 to 50 prompts is enough for a single product with one clear category. Fewer than 15 and one weird prompt swings your whole average. More than 100 and you are spending money to watch a number you cannot act on. There is a fuller breakdown of that tradeoff in our piece on how the AI visibility score is calculated.
Freeze the set once you have it. Adding or removing prompts mid-quarter breaks your ability to compare periods, which is the entire point of monitoring. Keep a separate “candidates” list and roll changes in on a fixed cadence, noting the date in your log.
Step 2: decide on a schedule and stick to it
The schedule matters less than its consistency. Pick one of these and hold it:
| Cadence | Good fit for | Rough monthly API cost, 30 prompts x 3 samples |
|---|---|---|
| Daily | Competitive categories, active content push | $5 to $15 on a small model |
| Every 3 days | Most indie SaaS | $2 to $5 |
| Weekly | Slow-moving niches, early-stage products | Under $2 |
Costs assume a cheap model tier and short answers. They climb fast if you sample heavily, run several assistants, or use larger models, so put a hard ceiling on the job before you automate it.
Run at the same time of day. Model behavior can shift after provider updates, and a fixed run time makes those shifts visible as a step change instead of smearing them across your data.
Step 3: log every run properly
Whatever you build, each individual run should produce one row containing: timestamp, prompt text, model version, browsing on or off, the full raw answer, whether your brand appeared, your position if it did, sentiment, competitor brands named, and any cited URLs.
Storing the raw answer is not optional. Six weeks later, when your rate has dropped, the only way to understand why is to read what the model actually said then versus now. A boolean “mentioned: false” tells you nothing about the cause.
The competitor column is what turns monitoring into strategy. Every answer that omits you names somebody else, which means every run is also a free competitor sample. Extract those names consistently and you get share of voice for your whole category at no extra cost.
Step 4: read the numbers without fooling yourself
Once a few weeks of data exist, three patterns are worth acting on.
A sustained drop across many prompts. If your rate falls on 2 of 30 prompts, that is noise. If it falls on 20 of 30 in the same week, something structural changed: a model update, a competitor’s content push, or a site change of yours that hurt crawlability.
A competitor climbing while you hold steady. This is the most useful early warning available, and it usually shows up weeks before it shows up in trial signups. Go read the citations on the prompts they are winning. In practice, the answer is often a single comparison page or a fresh review roundup that the model started trusting.
A prompt where you are mentioned but framed badly. Low-effort fixes live here. If the model consistently says your product “lacks an API” and you shipped one last year, the outdated claim is coming from somewhere specific and often citable. Finding and correcting that source is far cheaper than trying to raise a mention rate from scratch.
Resist the urge to change your site the same week you spot a wobble. Give any suspected drop a second measurement period to confirm before you spend engineering time on it.
Manual, scripted, or a tool
Three ways to run this, and the honest tradeoffs:
| Approach | Setup effort | Ongoing effort | Breaks down when |
|---|---|---|---|
| Manual in the chat UI | Minutes | 1 to 3 hours per week | You pass roughly 10 prompts or want competitor tracking |
| Your own script plus a spreadsheet | A day or two | Maintenance when APIs change | You add more assistants or need sentiment and citation parsing |
| Purpose-built tracker | Under an hour | Reviewing dashboards | Mostly a budget question |
Manual is the right starting point, and there is a step-by-step version of it in our guide on tracking your brand in ChatGPT. The scripted route is genuinely viable for a technical founder, as long as you accept that mention parsing is harder than it looks: brand names get abbreviated, misspelled, and confused with similarly named products.
The tool category ranges widely. Enterprise platforms like Profound and Brandlight are built for brand teams with matching budgets, while lighter options such as Otterly and Peec AI target smaller teams. AskAiRank sits in the indie SaaS slice of that market, running your prompt set across ChatGPT, Claude, Perplexity, and Gemini on a schedule and tracking competitors on the same prompts.
Where to start this week
Do not build the whole system on day one. Write 15 prompts a real buyer would ask, run each of them three times in the ChatGPT interface, and record the results in a spreadsheet with one row per run. That is two hours of work and it gives you a genuine baseline.
Repeat it once next week. If the numbers moved and you found yourself wanting to know why, you have justified automating it. If nothing moved and you did not care, you have saved yourself a subscription.
Frequently asked questions
Daily is enough for most indie SaaS products, and weekly is acceptable if your category moves slowly. Anything more frequent mostly buys you noise, since a single prompt's answer varies between runs anyway. The important part is that the interval stays fixed, so week-over-week comparisons mean something.
The chat interface works for spot-checks and is free. Automated monitoring needs API access, because you want the same prompt run repeatedly on a schedule without a human clicking. Expect a few dollars a month in API costs for a set of 20 to 50 prompts on a small model.
Monitor both if you can, and never mix them in the same metric. Browsing-enabled answers pull live pages and reward fresh, crawlable content. Answers without browsing reflect what the model absorbed during training and move much more slowly. They are two different visibility surfaces with two different fixes.
Probably not on its own. With a handful of samples per prompt, swings of 10 to 20% are normal statistical noise. Treat a drop as real when it holds for two consecutive periods, shows up across several prompts rather than one, or coincides with a competitor's rate climbing by a similar amount.
Start with ChatGPT because it has the largest assistant audience, then add the others once your process runs itself. Each assistant sources answers differently, so a strong ChatGPT mention rate tells you very little about Perplexity, which leans heavily on live citations.
