1001 SEO Media
All posts
By 1001 SEO Mediaai prompt trackingprompt governanceclient reportinggeo agency

How to do AI prompt tracking: running a client prompt set like a contract

How we govern AI prompt tracking inside a retainer: who owns the list, a frozen core set, a change log, when prompts are added or retired, and client sign-off.

Building a prompt set takes about two weeks. We described that process in how we build client prompt sets with volume and difficulty data, and the wider client setup in how to do AI rank tracking. This post is about the eleven and a half months after that, which is where most prompt tracking quietly goes wrong.

Nobody breaks a prompt set on purpose. It drifts. A stakeholder asks for "just a few" prompts before a board meeting. Someone rewords a question to sound more on-brand. A product is discontinued, its prompts keep running, and the score drops for reasons no one remembers. Six months later, the trend line in the monthly report is measuring a different set from the one it started with, and nobody can say when it changed.

So we treat a client's prompt set the way we treat a contract: named owners, a fixed core, written changes, and sign-off. Here is how that works in practice.

Step 1: Decide who owns what

Two people own every prompt set, one on each side.

On our side, the account lead owns the method: how prompts are written, how they are tagged, what goes in the core, and how the platform is configured. On the client's side, one named person owns the business truth: what the company sells this year, which markets matter, who the real competitors are, and what a buyer actually asks a salesperson. Not a committee. One person who can say yes.

Everyone else can suggest prompts. Only those two can approve a change.

In Promptwatch, the platform we run client programs on, this maps onto roles. There are owner, editor, and viewer roles with project-level access. We give client stakeholders viewer access to their own project, so the CMO can open the dashboard without being able to change the set. Editing rights go only to the people named in the governance document. If the client's in-house team also works in the platform, their editor gets access to that client's project and no other.

Step 2: Split the set into a frozen core and a watchlist

A trend line only means something if the questions behind it stay the same. A prompt set also has to respond to the business. Those two needs pull against each other, so we split the set in two.

The core set is the prompts the headline numbers are built on. It is frozen for the quarter. Nothing is added, reworded, or removed mid-quarter except for the narrow cases in step 5. The watchlist is where every new prompt starts. It changes monthly, it gets reported separately, and it never moves the headline.

In Promptwatch we run these as two monitors per market, for example "Core, Netherlands" and "Watchlist, Netherlands", with the same language and location. Dashboard PDFs and the citation views can be filtered by monitor, so the headline report reads the core monitor alone. When a watchlist prompt earns promotion at the quarterly review, Copy Prompts moves it into the core monitor. We write the promotion date in the change log, so nobody reads the core's earlier months as if they included it.

A prompt sits on the watchlist for at least one full month before we consider it for the core. One month of daily answers is enough to see whether the prompt is stable and whether it tells us anything.

Step 3: Keep a change log that the client can read

The change log is a plain document, shared with the client, with one row per change: date, prompt, monitor, what changed, why, who asked, and who approved. A reworded prompt is logged as one retirement plus one addition, because to an AI engine it is a new question.

The log also covers changes that are not prompts but still move the numbers:

  • Edits to the Brand Book's analysis instructions, which decide what counts as a mention. They apply only to new responses, so the date matters.
  • Competitors added or removed, including ones found in the auto-discovered competitor list.
  • Engines added to a monitor.
  • A new market. Because a monitor's language and location are fixed once created, a new market always means a new monitor, never an edited one.

On Business, Agency Scale, and Enterprise plans, Promptwatch's audit log sits behind our document. Owners can see every change and whether it came from the dashboard, the API, an MCP client, or Agent Chat. That last part matters once changes go through Agent Chat or the MCP server as well as the interface. Agent Chat asks for every missing value before it changes anything, which helps, but the audit log is how we confirm afterwards that a change made in a chat thread actually matches the row in our log. On Kick-off and Growth there is no audit log, so the shared document has to carry the full load.

Step 4: Add prompts by rule, not by request

New prompts come from a short list of triggers. The client launches a product or enters a market. Sales hears the same question three weeks running. A competitor we do not track starts appearing in answers. A topic is emerging before it has measurable demand. Or Prompt Explorer surfaces a suggestion with real volume from the client's Search Console data, ranking keywords, People Also Ask, a manual search, or one of the client's own pages.

Every candidate goes onto the watchlist with tags that record where it came from and when: src-sales, src-gsc, added-2026-12, and so on. Tags and bulk editing in Promptwatch make this a two-minute job for a batch of twenty. The tags are what let us answer, a year later, why a prompt is there.

Before a batch goes live we check quota. In Promptwatch, newly activated prompts start using responses immediately, usage alerts go out at 85% of a quota, and the Usage Overview shows each project's share. On an agency plan where several clients share one pool, one client's enthusiastic batch can eat into everyone else's month. We size the batch first.

Requests that do not meet a trigger still get an answer. A stakeholder's favorite question goes onto the watchlist with a src-stakeholder tag, and the data decides whether it moves up. That is a much easier conversation than a flat no.

Step 5: Retire prompts without losing their history

We prune on a fixed rhythm, quarterly, using the rules in when to drop a low-volume prompt. The governance part is how we retire, not what.

Retiring is not deleting. Promptwatch's MCP server describes deactivating a prompt as stopping future runs while keeping its response history. We still export the prompt's responses to CSV before retiring it, because a file in the client's folder survives any change of tool or agency.

There are only three reasons to touch the core mid-quarter: the product or service in the prompt no longer exists, the prompt names a competitor that has shut down or merged, or the prompt contains a factual error. Each one is logged, and the chart in the next report carries a dated note.

Step 6: Give the prompt set ten minutes in every monthly review

Our monthly review agenda already has a slot for the plan. Ten minutes of it belong to the prompt set. We bring three lists: watchlist prompts showing a signal, proposed additions with their trigger, and proposed retirements with their reason.

Two Promptwatch views do most of the work here. The Topics page shows Visibility, Citation Share, Average Position, Share of Voice, and Sentiment per topic, so a topic with too few prompts to read stands out. The Topics Manage tab lets us create a topic, pick its color, and check or uncheck its prompts in one screen, with a Suggested tab that proposes prompts sharing words with the topic. Nothing saves until we press Save changes, so we can rearrange topics live with the client and discard if they disagree.

For evidence, we use the Responses export, which lists each brand mention with its position in the answer. It is easier for a client to approve a retirement when they can see six weeks of answers that never named anyone relevant.

Every third month, the review becomes the quarterly reset: watchlist promotions, core retirements, and a fresh sign-off.

Step 7: Make sign-off explicit

The client signs off the prompt set three times: at the start of the engagement, at every quarterly reset, and whenever a mid-quarter core change from step 5 happens. Sign-off is an email reply to a single document, the current prompt set as a table with prompt, monitor, topic, type, intent, tags, date added, reason, and status (core, watchlist, or retired).

We keep it lightweight on purpose. Nobody should need a meeting to approve a list. But nothing moves into or out of the core without that reply, and we file it next to the change log. When a new CMO arrives in month eight and asks why the numbers look the way they do, the answer is two documents and a filter, not an afternoon of archaeology.

Our recommendation

Prompt tracking stays useful for a year only if someone governs it: two owners, a frozen core, a watchlist, a change log, and written sign-off. The tool cannot do that for you, but it can make each step cheap. Copy Prompts, tags, the Topics Manage tab, bulk editing, project-level roles, and the audit log in Promptwatch are the pieces we use, and they turn each step into a few clicks instead of a spreadsheet exercise.

If you are choosing a tool first, read the best AI prompt tracker for 2026 and AI prompt tracker compared, or our earlier AI rank tracker roundup. To try the structure yourself, start the 7-day free trial on monthly billing and set up a core and a watchlist monitor before your first run. If you would rather we run the governance for you, get in touch.