AI prompt tracking

Summary#

Ask ChatGPT the same question twice and you will get two different answers. That makes a single AI visibility reading interesting but not comparable. Tracked prompts fix the questions: you save a list of prompts for a brand, and from then on the AI tools and the weekly background tracker send that exact list every time. The result is a trend you can actually read, an export you can hand to a client, and an email when the number really moves.

On a paid plan the weekly rescan and the alert are both on by default. You do not have to save a prompt or create an alert for either to happen; doing so only changes what is asked and how large a move has to be.

Purpose#

AI visibility is measured by asking a model a neutral question and checking whether your brand comes up. The question decides the answer. If the wording drifts between runs, a change in your score might mean the market moved, or it might mean you asked something else.

Saving prompts removes that variable. It also lets you measure what you actually care about. Auto-generated prompts are competent but generic; your own can name the buying situation you win in, the objection you lose on, or the comparison a sales team keeps hearing.

Requirements#

• A signed-in account. There is no plan requirement, and none of the screens described here cost credits.
• A brand or domain you have measured at least once. A brand enters the weekly tracker only after you run Visibility Overview, Prompt & Question Research or Prompt Tracking on it. Saving a prompt on its own does not start tracking.
• A paid plan for the weekly rescan and its automatic alert. On Free the brand is not enrolled, and the tool runs that would enrol it cost credits Free does not have.
• For alert emails, the notification switches Email Notifications and SEO Score Alerts both on. See Notification preferences. Both are on for a new account, so this only matters if you turned one off.

Permissions#

Any signed-in account on any plan. Everything here is per account: your alerts are yours, and a teammate does not inherit them. See What teammates share.

Prompts are stored against the brand name rather than against a screen, so the same saved set is used by every tool that measures that brand, and by the background tracker. On the Tracked Prompts tab you see and can delete the prompts saved under your own account, plus any that were saved without an owner.

Navigation Path#

The prompt, comparison and alert controls live in one panel, opened from a floating round button in the lower-right corner of the dashboard labelled Tracked Prompts.

Dashboard → Tracked Prompts (floating button) → AI Tracking Settings

The panel has four tabs: Tracked Prompts, Compare Brands, Alerts and Help.

The tools that consume these prompts live under Dashboard → AI Visibility, in particular Prompt Tracking. See Prompt Tracking.

Step-by-Step Guide#

Save prompts for a brand#

1. Open the Tracked Prompts floating button, then stay on the first tab.
2. Type the brand or domain into Brand or domain. The field's example is e.g. nike.com. Press Load or the Enter key.
3. The box below fills with the prompts already saved. Before you load anything it reads Enter a brand and click Load to see its tracked prompts. A brand with nothing saved reads No custom prompts yet for this brand. Auto-generated prompts will be used.
4. Type a question into Add a prompt. The example is e.g. best note-taking app for teams. Press + Add or Enter.
5. The status line under the form confirms with Added., or Add failed. if the save did not land. A prompt is stored up to 500 characters; anything longer is trimmed.
6. Remove a prompt with the × at the end of its row. The status line reads Deleted.

Compare brands#

1. Open the Compare Brands tab.
2. Enter up to five brands, separated by commas. The example is e.g. nike.com, adidas.com, puma.com. More than five are ignored.
3. Press Compare. The chart plots each brand's visibility percentage over the last 30 days, one line per brand.
4. With nothing measured yet the status line reads No trend data yet for any of these brands. Run AI Visibility on them first, or wait for the scheduled tracker.

The comparison is drawn from measured history only. It never estimates a missing day, so a brand you started tracking last week has a shorter line.

Set a visibility drop alert#

1. Open the Alerts tab.
2. Enter the brand in the first field, example brand (e.g. nike.com).
3. Set the threshold in the number field next to it. It accepts 10 to 99 and starts at 15, meaning "email me if visibility moves by 15 points or more, up or down". Anything under 10 is rejected with Threshold must be 10-99. Below 10 points is measurement noise. Adding a brand that already has an alert adjusts that alert rather than creating a second one.
4. Press + Add Alert. The status line reads Alert added.
5. Each saved alert appears as a row reading <brand> — alert when visibility moves by <N> points or more (last measured: <N>%). Before the first scan the tail reads (awaiting the first scan). Remove it with the ×. A row you delete stays deleted; it is not recreated the next time you run a tool on that brand.

Validation messages you may see: Sign in to add alerts., Enter a brand., Threshold must be 10-99. Below 10 points is measurement noise.

Export the trend#

Run one of the AI visibility tools on a brand. Its trend chart carries two buttons, Export CSV and Export JSON, which download the last 90 days of measurements as metricvault_<brand>_trend.csv or .json.

The CSV columns are captured_at, brand, engine, prompt, brand_appeared, position, sentiment — one row per prompt per engine per run, which is the raw material behind every chart.

If you press an export button before running a tool, you get Run a Tier 2 tool first so the export knows which brand to export. The export needs to know which brand you mean, and it takes that from your last run.

What happens behind the scenes#

The background tracker looks for work about every 6 hours. A brand is due when its newest measurement, scheduled or from a tool run of your own, is at least seven days old. On each pass it:

1. Takes up to two due brands, most overdue first, with brands owned by a paid account ahead of everything else. A brand that misses a pass is more overdue on the next one, so it moves up the queue rather than being dropped. An attempt counts the same as a measurement for this, so a brand the tracker cannot measure waits its week like any other instead of taking a slot on every pass. See the note on prompt generation below.
2. Uses up to eight prompts per brand. Your saved prompts win. If you have saved none, it generates a neutral set instead: a mix of category recommendation, comparison, problem-solving and pricing questions.
3. Sends every prompt to every AI engine that is configured, and records for each answer whether your brand appeared, roughly where in the answer it appeared, and whether the surrounding language read positive, neutral or negative.
4. Writes one history row per prompt per engine, which is what the trend, the comparison and the export all read.

What counts as a meaningful change#

Delta alerts are evaluated at the start of every pass, against the scans that have completed so far. An alert compares the newest scan against the previous one and emails you, in either direction, only when both of these hold:

• the score moved by at least 10 points, or your own higher threshold, and
• that move represents at least 3 individual prompt-and-engine results changing.

The second condition is the one that keeps the mailbox worth reading. A scan is eight prompts against every engine configured, so a single result flipping is worth about 6 points on two engines and 3 on four. A threshold alone would fire every time one answer changed its mind, and because the prompts are regenerated on every pass the two scans are not even asking the same questions. Requiring a real share of the sample to move filters both.

In practice, with four engines a 10-point move is reported; with two engines it takes about 19 points; with a very small sample nothing short of a collapse is reported at all. A thin sample has to move a long way before we are willing to call it news.

At most one email per alert per 24 hours. The subject is AI Visibility drop on <brand> - down <N> points or AI Visibility rise on <brand> - up <N> points, and the body says how many of the sample's results changed.

Setting a higher threshold on an alert makes it quieter. Setting a lower one cannot make it noisier: 10 points and 3 results are the floor whatever the row says, which is why the form does not accept less.

None of this consumes your monthly credits. The background tracker does its own model calls and does not draw on your allowance.

Reading the numbers#

Screenshots#

![screenshot](placeholder: The Alerts tab showing one saved alert row reading "nike.com — alert when visibility moves by 15 points or more (last measured: 62%)".)

Tips#

• Eight prompts per brand is the measurement budget. Saving more than eight means only the first eight are used on a scheduled pass, so put your most important questions first.
• Keep prompts stable. Every edit resets the comparability of the trend from that day forward, in the same way that changing a survey question breaks the series.
• Track your two closest competitors as brands in their own right. Comparison is far more informative than your own line alone.
• Export monthly. The export is the only copy of the raw per-prompt, per-engine detail you can keep outside the app.
• A threshold of 15 is a reasonable first alert. Set it lower once you know how much your brand's number moves on a quiet week.

Best Practices#

• Write one prompt per buying situation rather than several wordings of the same question. Duplicated intent wastes the eight-prompt budget.
• Include at least one comparison prompt and one problem-first prompt. Brands often appear strongly in one and not at all in the other.
• Review sentiment, not only appearance. Being mentioned as the expensive option is a different problem from not being mentioned.
• Give a new brand a week before drawing conclusions. One pass is a reading; a week is a trend.

Common Mistakes#

• Expecting prompts alone to start tracking. A brand joins the weekly tracker when you run an AI visibility tool on it. Save the prompts, then run the tool once.
• Naming your brand in the prompt. It guarantees a mention and destroys the measurement.
• Reading a one-day drop as a trend. Model answers vary. The alert threshold exists so that ordinary variation does not email you.
• Assuming every engine is measured. Only engines whose connection is configured take part, and the engine breakdown on the results screen shows which ones ran.
• Expecting the export to cover the whole history. The export buttons pull the last 90 days.

Limitations#

• Up to two brands are measured per scheduled pass, most overdue first. The scheduler looks for work about every 6 hours, so up to eight brands a day. Beyond that, a brand waits longer than seven days for its rescan; it is not dropped, and each missed pass moves it further up the queue.
• Up to eight prompts per brand per pass.
• The panel compares up to five brands over the last 30 days.
• There is no per-prompt history view in the app. The export carries that detail.
• Prompts cannot be edited, only added and deleted.
• Alerts are email only. There is no Slack or webhook delivery for visibility changes.
• An alert can fire at most once every 24 hours per alert, so a move that continues over several scans produces one email per day at most.
• A move under 10 points, or one carried by fewer than 3 prompt-and-engine results, is never emailed. It is still in the trend chart and the export.

Troubleshooting#

FAQs#

Do tracked prompts cost credits? No. Saving, loading, comparing, exporting and alerting are all free, and the background tracker does not draw on your monthly allowance. Running an AI visibility tool by hand costs whatever that tool costs. See How credits work.

How often is my brand measured? Once a week, on every paid plan, without switching anything on. A brand is due seven days after its newest measurement, and the tracker picks up due brands about every 6 hours. See Automation limits and frequency.

Can I force a measurement now? Not directly. Running the tool on that brand takes a live reading immediately and also moves the brand to the front of the scheduled rotation.

Why is my visibility 100% one day and 40% the next? Because model answers genuinely vary, especially for broad prompts. A stable prompt set narrows this, and the trend line matters more than any single day.

Does a drop alert tell me which prompt caused it? No. The email names the brand, the previous and current percentages and the size of the drop. Open the tool or the export to see which engine and prompt moved.

Can my teammate see my tracked prompts? They see prompts saved under their own account, plus any saved without an owner. Alerts are private to the account that created them.

What happens if I delete every prompt for a brand? Measurement continues using auto-generated prompts, and the panel says so. Your history is not deleted, but the questions change, so treat that point as a break in the series.