Prompt Tracking

Summary#

Prompt Tracking runs a fixed set of prompts through every connected AI engine, scores how often your brand is named, and stores each run so the win rate builds into a real trend. It shows which prompts you lost, which engine is weakest, and which competitor took your place.

Purpose#

Prompt & Question Research finds the prompts that matter. Prompt Tracking is what you do next: watch them. Presence in AI answers moves without warning when a model updates or a competitor publishes, and the only way to notice is to ask the same questions on a schedule and keep the answers. This tool is the measurement half of that loop.

Overview#

Enter a brand or domain. Metric Vault takes fifteen prompts — your own tracked list if you have one, otherwise a generated category set that never contains your brand name — and sends every prompt to each connected engine: ChatGPT, Gemini, Perplexity, Claude and Google's AI Overview. It records whether your brand appeared in each answer, keeps the verbatim text as proof, and writes the run to your tracking history.

Because every run is stored, the trend chart fills with your own measurements over time. The first run has no curve; by the third or fourth there is something to read.

A background tracker also re-measures saved brands roughly once a day, so the trend keeps building between manual runs and a visibility drop can raise an alert. See AI prompt tracking.

Benefits#

• One credit for fifteen prompts across every connected engine.
• A single win-rate number that survives being put in a report.
• Per-engine win rates show whether a loss is universal or engine-specific.
• Verbatim answer excerpts prove the result rather than asserting it.
• Trend history accrues automatically, including from the background tracker.

Use Cases#

• Weekly standing check. You run the same fifteen prompts each Monday and watch the win rate rather than guessing at AI visibility.
• Post-publication proof. You published three assets for lost prompts; the next run shows which of them changed an answer.
• Competitor watch. The Competitor Replacements block names who is being returned where you used to be.
• Engine-specific diagnosis. You win on four engines and lose on one, which points at that engine's source pool rather than at your content.
• Executive reporting. The win-rate trend is the one AI-visibility chart most stakeholders will actually read.

Requirements#

• A signed-in account and a paid plan; Free covers the zero-credit technical tools only.
• One credit available. See How credits work.
• A brand or domain with enough public presence for category prompts to be generated.

Permissions#

Plan decides access; team role does not.

Cost#

1 credit per run, as printed on the button: Track Prompts · 1 report.

It is a light tool: recorded in your usage but never blocked when the monthly premium allowance is spent, and subject to the 100 light-tool calls per hour fair-use limit. A shared-cache hit still charges 1 credit; the cache holds this tool's data for one day, which is why a same-day re-run may return the earlier measurement. Reopening a saved result is free. See How credits work and Result caching and freshness.

Navigation Path#

Dashboard → AI Visibility → Prompt Tracking

Inputs#

Step-by-Step Guide#

1. Open Dashboard → AI Visibility → Prompt Tracking.
2. Optional but recommended: click the Tracked Prompts button at the bottom right to open AI Tracking Settings. On the Tracked Prompts tab enter your brand, press Load, then add each prompt you want watched with + Add.
3. Close the modal and type your brand or domain into the input marked e.g. nike.com.
4. Click Track Prompts · 1 report, or press Enter in the input field.
5. If you have run this brand before, choose Open saved result (free) or Run fresh (1 credits).
6. Wait while fifteen prompts run against every connected engine — this is slower than a data-only tool.
7. Read the win rate, then the per-engine breakdown, then the prompt table.
8. Optional: on the Alerts tab of the same modal, add the brand and a drop threshold so you are emailed when visibility falls.

Reading the Results#

The header shows the brand with its favicon and a Prompt Tracking / Live LLM pair of badges.

The five KPI tiles.

• Total Prompts — how many prompts ran. Fifteen for a generated set; your tracked list can differ.
• Avg Win Rate — the headline. It is the share of all prompt-by-engine checks your brand won: fifteen prompts across four engines is 60 checks, and winning 27 of them is 45%. Bands are Dominant at 70 and above, Strong from 50, Building from 30, Behind below.
• Top Engine — the engine naming you most often, with its win rate.
• Lost Prompts — prompts where no engine named you. This is the number to drive down; each one is a complete absence, not a partial loss.
• Engines — how many engines were queried. A low win rate with five engines is a different problem from a low win rate with two.

Overall Win Rate repeats the headline as a gauge. Read it as a rate, never as a count — adding an engine changes the denominator, so the percentage can fall while your absolute presence grows.

Per-Engine Win Rate gives one figure per engine with its wins over total prompts. The shape matters more than any single number:

Win-Rate Trend plots your stored history, up to thirty points. It is built only from real runs — your manual ones and the background tracker's — so a new brand shows a short or empty line until history accrues. Judge direction over several points; a single-run swing of a few points is normal model variance.

Sample Verbatim Answers shows the literal text an engine returned. Use it for two things: proving the result to someone who doubts it, and seeing how you were named. Being listed fifth in a rundown is a different outcome from being the recommendation.

Prompt Performance is the per-prompt table — the working list. Each row carries the prompt, its category, a business-impact label, how many engines named you out of the total, the average sentiment and the top competitor returned instead. Sort your attention by the rows with zero engine wins and a high business impact.

If no prompt-level results came back, this block is replaced by No tracked prompts yet: "No prompt-level results came back for this brand. Add prompts to your tracked set, then re-run to populate per-prompt win rates and engine coverage here."

Competitor Replacements names the brands returned where you were not. A name that appears across many prompts is the incumbent answer in your category; study what it publishes before writing anything.

Action Plan closes the report. Its standing advice is sound and worth following: publish a focused asset for every prompt with zero wins; for prompts you win on some engines but not all, read the losing engine's answer and its cited sources to see what is missing; and re-run on a fixed cadence so the trend builds.

An Export control on the result writes the prompt table to CSV with the columns Prompt, Category, BusinessImpact, EngineWins, TotalEngines, AvgSentiment and TopCompetitor.

Examples#

Screenshots#

Tips#

• Define your own tracked prompts before you start measuring. A generated set is a good sample; a curated set is a KPI.
• Fix the prompt list and leave it alone. Changing prompts resets what the trend means.
• Set a drop alert on the Alerts tab so you learn about a loss without opening the app.
• Keep the verbatim excerpts for the prompts that matter most — they are the evidence when someone asks whether AI visibility is real.

Best Practices#

• Run on a fixed weekday so the trend spacing stays even, and let the background tracker fill the gaps.
• Use Prompt & Question Research to discover prompts and this tool to watch them. Discovery and monitoring are different jobs.
• Work the Lost Prompts list, not the win rate. The rate is the score; the lost prompts are the game.
• When a win is lost, read the new answer before assuming the cause. A competitor's new page and a model update need different responses.

Common Mistakes#

• Changing the tracked prompt list mid-quarter. The trend becomes meaningless, because the denominator changed.
• Comparing win rate across brands with different engine counts. The rate depends on how many engines answered.
• Clicking a TRY chip to explore. The chips fill the field and run immediately, spending a credit.
• Expecting a full trend on the first run. The chart is built from stored measurements and starts empty.
• Reacting to a single-run drop. Model output varies; wait for a second confirming point unless the drop is large.

Limitations#

• Fifteen prompts per generated run.
• Only engines Metric Vault has a live connection to are queried, and the report names them.
• The trend chart holds up to thirty stored points.
• Brand detection is a text match, so a brand named after a common word can register false wins.
• The result is cached for one day, so a same-day re-run may not re-measure.
• Sentiment is a coarse positive/neutral/negative read of the answer text.

Troubleshooting#

FAQs#

How do I choose which prompts are tracked? Open the Tracked Prompts button at the bottom right of the app, enter your brand, and add prompts one at a time. Saved prompts are used by Prompt Tracking, Prompt Research and Visibility Overview, and by the background tracker.

Is the win rate the same as the visibility score? No. Win rate here is the share of prompt-by-engine checks your brand won on this tracked set. Visibility Overview reports an overall visibility score across its own measurement. They move together but are not the same figure.

Does the background tracker cost credits? No. Scheduled tracking runs on its own cadence and does not draw on your monthly allowance; only the runs you start from this screen are charged.

Does a cached result still cost a credit? Yes. A shared-cache hit charges 1 credit. Only reopening your own saved result is free.

Is this available on the Free plan? No. It needs a paid plan, then keeps working after your monthly premium allowance is spent because it costs fewer than 3 credits. See Quotas and rate limits.