AI Prompt Tracking: A Practical Method
How to build the prompt list, decide how often to run it, and log what ChatGPT, Claude, Gemini and Perplexity say about your brand and competitors.
AI prompt tracking is running a fixed, unchanging set of buyer questions through ChatGPT, Claude, Gemini and Perplexity on a regular schedule, and logging whether your brand is named, whether it is linked, and which source got cited instead when it wasn't you. The method is a spreadsheet before it is a tool.
It is the raw data underneath every other measurement in AI brand visibility — AI share of voice, sentiment and any visibility score are all just different ways of summarizing what a disciplined prompt tracking habit already recorded.
On this page
Where do the prompts actually come from?
Sales calls and support tickets, not a keyword tool. The questions a keyword tool surfaces are written for search boxes — short, fragment-like, optimized for how people type into Google. The questions a buyer actually asks a person, or would ask an assistant if given the chance, are longer and more specific: "is it worth paying an agency for this or can my team do it," "what's the difference between X and Y for a company our size," "who else in this space is worth looking at."
If you do not have transcripts to pull from, the next best source is whoever answers the phone or the first sales call — ask them to write down the last ten questions a prospect asked before signing or walking away. That list, written from memory by someone who has the conversations weekly, beats almost any keyword research tool for this specific purpose, because it captures the actual hesitations and comparisons real buyers make, not the search terms a tool infers from aggregate click data.
Resist the temptation to write prompts that flatter you. It is tempting to phrase every question in a way you already know you will answer well, but that defeats the purpose. The prompt set exists to find out what an assistant says when a real, slightly skeptical buyer asks — including the comparison questions where a competitor might come out ahead. A tracking list that only asks favorable questions will report a rosy number that has nothing to do with what buyers actually encounter.
Pull from real conversations
Sales call transcripts and support tickets are the best source, because they are the actual words a real buyer used, not a guess at what they might type.
Cover four question types
Category questions ("best tools for X"), comparison questions naming competitors, questions about you by name, and objection questions ("is X worth it").
Write each one once, exactly
The wording matters and should not change between runs. If you want to test a new phrasing, add it as a new row — do not edit an existing one.
Cap it at a size you'll actually run weekly
Twelve to forty per market. A ninety-question list that gets skipped every third week produces worse data than a twenty-question list that runs like clockwork.
How often do you actually need to run it?
Weekly, at minimum, for as long as you are trying to see a trend rather than confirm a single fact. The underlying answers are not stable: research tracking the same seventy-two questions two days apart found Claude retained only 30 percent of the URLs it had cited in the earlier run, and even Perplexity, the most stable of the four, retained just 65 percent [1]. A monthly cadence will not resolve real movement from that background noise; weekly gives you enough data points to average it out.
Think of it the way a stock price behaves rather than the way a quarterly report behaves. Nobody makes a decision off one minute's tick on a stock chart, and nobody should make a decision off one Tuesday's chat answer either. What you are actually building, run after run, is something closer to a moving average — smoothing out the churn so the real trend underneath it becomes visible. Four weekly runs, minimum, before you say anything to a stakeholder about whether a number moved.
This also means the first month of any tracking programme should be treated as calibration, not results. You are learning what normal week-to-week variance looks like for your specific prompt set and your specific category before you can tell a real signal from noise. Some categories are noisier than others — highly competitive, well-covered categories with dozens of comparison posts in circulation tend to swing more than a narrow, under-covered niche where the same three sources dominate every answer.
30% URL retention on Claude, two days apart
What exactly do you log per prompt, per run?
Four fields, kept identical across every row so the data is actually comparable later: whether you were named, whether you were linked, the sentiment of how you were described if named, and — critically — which URL got cited instead of you when you were absent. That last field is the one people skip and the one that turns a tracking spreadsheet into an actual to-do list, because it tells you exactly which third-party page to go after.
Add a fifth field once you have a few weeks of data: a one-line note on anything unusual about that run. "New competitor appeared for the first time," "our own page was cited for the first time," "answer referenced an outdated price." These notes cost thirty seconds each and they are what make the log readable six months later, when the raw named/linked/sentiment columns alone will not explain why a particular week looked different from the ones around it.
Keep the raw log even after you start reporting a summarized trend. It is tempting to only keep the rolled-up weekly percentage once you have enough history, but the raw per-prompt rows are what let you answer a specific question later — "did we ever get cited for the pricing comparison prompt specifically" — that a summary number cannot answer on its own.
| Manual spreadsheet | Paid tool | |
|---|---|---|
| Setup time | An afternoon | Days to weeks of onboarding |
| Cost for the first month | Free | Usually a monthly fee per brand or per prompt volume |
| Ceiling before it breaks down | Comfortable to about 40 prompts, a few engines | Scales to hundreds of prompts, many markets and languages |
| What it teaches you | Exactly what the metric means, because you build it by hand | Whatever the vendor's dashboard chooses to show first |
Start manual. Move to a tool once the spreadsheet is genuinely too large to run by hand — most companies never get there, and the ones that do are better buyers for having built the manual version first, because they know exactly what question to ask a vendor's sales demo instead of taking the dashboard's word for what it measures.
How does sentiment fit into a prompt tracking log?
As a fifth column, not a separate project. When an assistant names you, note briefly how — neutral factual description, positive recommendation, or something closer to a warning or a comparison that puts you last. This is a short label, not a full analysis, and it belongs in the same row as the named/linked data so you can see the two move together over time. AI brand sentiment covers how to read and act on that column once you have a few months of it, including what to do when the tone is negative but nothing in the answer is factually wrong.
Does entity data affect how consistent your tracking results are?
Indirectly, and more than people expect. A brand whose name, description and founding facts are consistent across its own site and the profiles an assistant might check tends to produce more stable tracking results — less week-to-week churn caused by the model resolving conflicting signals differently each time it is asked. Entity SEO covers the underlying work; the payoff shows up here, as a tracking log with less noise in it and more signal you can actually act on, and fewer weeks where you cannot tell whether a dip is real or just the model guessing differently.
A tracking spreadsheet that changes its own questions is not measuring anything. The discipline is the whole method — decide the list once, run it exactly as written, and only revisit it on a schedule you set in advance.
If you would rather have this run and reported on for you, weekly tracking across every engine you need is the core of our AI visibility audit.
Questions people also ask
How many prompts should I track?
Twelve to forty per market is enough to see a trend and few enough to run by hand for a month before deciding whether to invest in a tool.
Should I write different prompts for each engine?
No. Run the identical wording on every engine. Differences in how each one answers the same question are the finding — rewording per engine hides that.
Can I just check ChatGPT since that's what most buyers use?
You can, but you will be blind to whatever a different assistant says, and the four major engines disagree with each other far more than most people expect — they cite the same source for the same question only a small fraction of the time.
Do I need a paid tool to do this properly?
Not to start. A spreadsheet and a recurring calendar reminder gets you a month of real data, which is enough to know whether the number matters to your business before you pay a vendor for automation.
What's the single most common mistake in a prompt tracking programme?
Editing the prompt list after a bad month. It resets the baseline and destroys the only thing the measurement was for. Decide the list once, run it as written, and only revisit it on a fixed schedule — quarterly, not reactively.
Sources
- I tracked 13,184 AI citations across 4 LLMs: here's what I learned — Orbit Media Studios (checked 23 September 2026)
- AI features and your website — Google Search Central (checked 23 September 2026)
Get a tracked, weekly read on what AI engines say about you
Book a free 30-minute consultation with Ibrahem and Jeremy. We look at your site while we talk, and you leave with a prioritised plan.
Who wrote this
Ibrahem HelalCo-founder
Leads strategy and delivery.
Works on: SEO, GEO
Reviewed before publishing by Jeremy Zava, Co-founder.
- What Is AI Brand Sentiment?AI brand sentiment is the tone assistants use when describing your company — favorable, neutral or hedged. Here is what shapes it and how to check it.
- AI Share of Voice: How to Measure ItAI share of voice is the percentage of a fixed prompt set where you're named or linked instead of a competitor, tracked weekly across four assistants.
- Entity SEO: A Practical GuideEntity SEO is making your company a single, consistent, machine-readable thing across the web instead of a set of disconnected mentions. Here's how to do it.
- AI brand visibility: how to measure and grow itHow often ChatGPT, Perplexity, Gemini and Google AI Mode name or cite your brand, how to measure it on a schedule, what moves the number and what it costs.