How to do AI prompt tracking: an agent proposes, a person approves
How to do AI prompt tracking with an agent: seed the set by hand, let it propose additions weekly, approve them, keep topics clean, and turn weak prompts into briefs.
A prompt set rots in a predictable way. Week one, it holds 40 carefully chosen questions. By month three, someone has added twenty more from a brainstorm, half of them without topics, two of them duplicates with different punctuation, and nobody remembers which prompts were the original core. The trend lines now mix old and new prompts, and a falling score could mean you got worse or just that you started tracking harder questions.
More discipline rarely fixes this. A routine does, one where an agent does the repetitive parts (finding candidates, filing them, reading results) and a person makes the few decisions that need judgment. This guide sets that routine up. We use Promptwatch for the examples because every step lives in one project there, but the structure works with any tracker that has suggestions and an API.
If you only need the one-time move from Prompt Explorer suggestions into a new monitor, from prompt suggestions to tracked prompts with an agent covers those clicks. This post is about what happens every week after that.
Step 1: seed the first set by hand
Don't let an agent write your first prompts. Suggestions are built from what you already have (your pages, your Search Console queries, your existing topics), so the seed decides the quality of everything proposed later.
Start with the markets. In Promptwatch a monitor's language and location are fixed once it's created, so make one monitor per language and country you sell in before adding anything. Country tracking is on every paid plan. State and city tracking start on Business and are on every agency plan.
Then write 30 to 50 prompts the way a buyer would type them, from sales calls, support tickets, and the comparisons people make between you and named competitors. Give each one a type (Organic, Brand Specific, or Competitor Comparison) and an intent (Branded, Informational, Navigational, Commercial, or Transactional). Our post on how many prompts to track covers sizing in more depth.
Finally, tag this first batch core. The core is the set you never edit or delete. It's what your trend lines come from, however much the rest of the set changes.
Step 2: write the house rules before the agent starts
An agent can file prompts consistently only if "consistent" is written down. Keep it to a short page, and paste it into every agent request:
- A closed list of topic names. No new topics without a person agreeing to it.
- A short list of tags and what each means (
core,launch,competitor,test). - A cap per topic, so one busy topic doesn't eat the plan.
- A number of free prompt slots to keep in reserve.
- What makes a prompt acceptable: phrased as a person would ask it, about something you sell, and not a near-copy of a tracked prompt.
The slot reserve matters because caps are real. Promptwatch's brand plans allow 50 prompts and 6,000 responses on Essential ($95/mo), 150 and 18,000 on Professional ($245/mo), and 350 and 42,000 on Business ($579/mo). Agency plans have unlimited prompts, metered by responses instead. Usage alerts fire at 85% of a quota, and the per-project Usage Overview shows where it went.
Step 3: let an agent propose additions every week
Now give the repetitive research to an agent. In Promptwatch the natural tool is a Scheduled Chat: a saved Agent Chat question that runs on a cadence, reports only material changes by default, and can post to Slack. Brand plans include 10 of them, and our Scheduled Chats setup guide covers the options.
A useful weekly ask, set for Monday morning:
Review new Prompt Explorer suggestions and this week's query fan-outs for our tracked prompts. Propose up to 10 prompts to add. For each one, give the topic from our list, the intent, the monthly volume, and one sentence on why it belongs. Flag any that look like duplicates of tracked prompts.
Scheduled runs are read-only. The agent can analyze and propose but can't add, edit, or delete anything, which is exactly the split you want for a job that runs while nobody watches.
The inputs are good. Prompt Explorer suggests prompts from Search Console, ranking keywords, People Also Ask, manual search, or any page on your site, with a language picker for the prompt language. Prompt Volumes adds monthly volume, difficulty, and citation rates across AI search, Google, and Bing. Query fan-outs show the exact sub-queries each model ran before answering. Those are often the best source of new prompts, and they change. Promptwatch's site: operator report found that on August 8, 2026, site: searches went from 0.37% to 16.8% of all ChatGPT fanout queries, and the report says keyword mappings built before that date are stale. Fan-out mining is a weekly job, not a one-off.
If you prefer your own agent, the MCP server works the same way. Give it a read-only key, which hides the write tools, and let it call listSuggestedPrompts (the newest 50 pending suggestions, ordered by search volume) and listQueryFanouts.
Step 4: a person approves, in about 15 minutes
Approval is the step that keeps the set honest. Run down the agent's list with the house rules open and ask three things of each candidate. Would a real buyer type this? Do we want to win it? Is it already covered by a tracked prompt?
Then apply the decision. On Prompt Explorer, tick the approved suggestions and ask the Agent Chat panel to track them. It asks for any missing value, such as which monitor, before it changes anything. Dismiss the rejected ones so the same suggestions don't come back next Monday. Over MCP with a write key, that's acceptSuggestedPrompts and dismissSuggestedPrompts. For prompts that came from fan-outs rather than suggestions, createPromptsBulk adds them and reports any it skipped as duplicates of live prompts.
On Business, the audit log shows whether each change came from the dashboard, the API, an MCP client, or Agent Chat. That answers the question "who added this?" months later.
Step 5: keep topics and tags clean once a month
Weekly additions drift. Once a month, open the Topics page's Manage tab, where you can check or uncheck a topic's prompts in one place and filter by monitor. The Suggested tab proposes prompts that share words with a topic and shows the matched words, which catches new prompts the agent filed loosely. Nothing saves until you click Save changes. Topic prompt counts leave out prompts on paused monitors, so a topic that looks thin may just be paused.
An agent can do the bulk relabeling. Over MCP, attachTopicsToPromptsBulk and attachTagsToPromptsBulk create any missing topic or tag as they go. That last part is a reason to keep the closed topic list in the request, or you'll end up with "pricing" and "prices" as two topics.
When a prompt set proves itself in one market, copy it. Promptwatch can copy prompts to a monitor in another country if the language matches, and it can create the target monitor as part of the copy.
Step 6: turn weak prompts into content briefs
A tracked prompt you keep losing is a content request. Define "weak" so the agent can find them: tracked for at least two weeks, and either your brand is missing or your Position is late (lower is better) while competitors get cited.
For each weak prompt, read the actual responses and the fan-outs before writing anything. The fan-outs tell you what the model searched for, and the citations tell you which pages it trusted. Often the gap is a page you don't have, such as a comparison, a pricing explainer, or an alternatives page.
In Promptwatch, Content Creation starts directly from a tracked prompt or a query fan-out, with a brief that pulls in your branding, Google results, internal links, and up to five competitor URLs. That needs AEO articles, which start on Professional (5 a month) and rise to 10 on Business. On Business, the Content Agent can take the whole queue, planning and writing against visibility gaps on a schedule and publishing to WordPress, Webflow, or Framer through a review inbox where you approve, edit, or decline. Keep review on. The agent knows your prompts, not your legal team's opinions.
Step 7: check results at the prompt level
Topic averages hide what happened. When you check whether a new page helped, open the prompt it was written for, read the responses per model, and look at who is named and in what order. The Responses export lists each brand mention with its position, like Acme (1), Globex (2), which makes before-and-after comparisons easy in a spreadsheet.
Answers vary between runs even when nothing changed, so compare windows, not single days. Promptwatch's monitor comparisons measure a selected range against a previous period of the same length. Use Page Tracker on the new URL to see whether it gets cited, and crawler logs (from Professional) to see whether AI crawlers fetched it at all. For the visibility side of this check, how to do AI rank tracking sets thresholds for when a change counts.
Step 8: retire prompts without breaking trends
Prompts outlive their usefulness: a retired product, a competitor that left the market, a launch that's over. Deactivate them rather than deleting. Deactivated prompts stop running but keep their response history. Never edit a core prompt's wording. If it needs to change, add the new version and deactivate the old one, so the trend line has a clean break you can see.
If you ever move to a smaller plan, Promptwatch lets owners pick which prompts keep running. Having a core tag makes that a two-minute decision.
The weekly rhythm
| When | Agent does | Person does |
|---|---|---|
| Monday | Proposes up to 10 prompts with reasons | Approves, dismisses, applies |
| Weekly | Flags weak prompts | Picks which become briefs |
| Monthly | Bulk relabels topics and tags | Reviews the Manage tab, retires prompts |
| Per brief | Drafts in Content Creation or the Content Agent | Edits and approves before publishing |
Where to run it
You can patch this routine together from a tracker, a spreadsheet, and a writing tool, but the handoffs are where it breaks. Our AI prompt tracker comparison shows which tools cover which steps, and the best AI prompt tracker for 2026 ranks them on how much an agent can take over. Our broader list of the best AI rank trackers covers the measurement side.
Promptwatch runs every step in one project: Prompt Explorer and fan-outs for candidates, Scheduled Chats for the weekly proposal, the Agent Chat panel and MCP for adding and filing, Content Creation and the Content Agent for weak prompts, and Page Tracker, crawler logs, and visitor analytics to check the result. Start the 7-day free trial on Professional to run this routine with on-demand articles, or on Business if you want the Content Agent working the weak-prompt queue.