# Best practices: prompt sets and review routine

Source: https://support.seoai.de/docs/best-practices/

> How to build a prompt set worth tracking, what to check every week and month, and how to turn results into work.

Tracking only pays off when two things are right: the questions you track, and the habit of reading the results. This page collects what works in practice: how to build a prompt set, a weekly and a monthly review on the screens SEOAI already has, the order in which to act on what you find, and the mistakes that make numbers misleading. Each step links to the page that explains the screen in detail.

## Building a prompt set

### Start from real questions

A tracked prompt is a question a customer would type into an AI assistant, not a keyword. "best crm" tells you little; "which crm works for a two-person real-estate office without a sales team?" tells you who wins a real buying decision. Good sources for real phrasing:

- Sales calls, demo requests and support tickets: the questions people ask before they buy
- Reviews and forum threads in your market, including the objections
- Your search data: on Pro and Agency, the Search data tab in Suggest prompts turns the Search Console or Bing queries you already get impressions for into questions
- Keywords AI surfaced at the bottom of the Reports page: the terms assistants use when they talk about your market
- Demand you don't track on the Search tab in the Gaps area: search queries none of your prompts covers yet

Keep the qualifiers that change the answer: company size, budget, use case, industry, "for beginners", "near me". Write prompts in the topic's prompt language.

### Mix the intents

Every prompt gets one of six intents: Brand, Definition, Best-of, Comparison, How-to or Fact. Use them as the frame for the mix. A good starting point for a topic of twenty prompts:

- About half buying intent: Best-of ("best X for Y") and Comparison ("X vs Y", "alternatives to X"). This is where recommendations turn into customers.
- About a quarter research intent: How-to and Definition. These show whether assistants use your content to explain the category, and they produce most of the fan-out sub-queries worth covering.
- The rest Fact and Brand. Fact prompts check that assistants get specifics right, such as prices, opening hours or features.
- At most two or three Brand prompts. A prompt that names your brand will almost always mention you, so it lifts your visibility score without telling you whether you are recommended. Keep a few to watch what assistants say about you, their sentiment and possible name confusion.

Intents are classified automatically after the first runs; correct any that are wrong on the prompt list. The intent funnel then shows in which stage you drop out, which tells you where to add prompts and where to focus fixes.

### Size the set

Ten to thirty prompts per topic is the useful range. Below ten, one unusual answer moves the score noticeably; above thirty, a topic usually covers more than one question complex and reads better split in two. Plans cap prompts per topic at 10 on Free, 20 on Starter, 30 on Pro and 50 on Agency, so on Free each topic holds the ten questions that matter most.

Remember the cost rule: every prompt runs on every selected platform on every scheduled run. Twenty prompts on five platforms weekly is 100 executions a week, and 120 quota units because Gemini counts double. Reducing your costs explains how to keep the set affordable without losing signal.

### One topic, one market

A topic owns the prompt language, the market and the optional city. Keep one language and one market per topic; an English question asked from the UK and one asked from Germany are two measurements. For local businesses, set the city on the topic instead of writing it into every prompt: SEOAI applies the topic's city when the prompt runs, and the same prompts can then be reused in a topic per location.

### Keep wording stable

Prompt text is fixed once a prompt is added, because changing the words changes the answers and would break the history. To track a better wording, add it as a new prompt and switch the old one off with its Active toggle: a paused prompt keeps its results. Deleting a prompt deletes its results too.

Change the set in small steps. Replacing half the prompts at once changes your scores as much as any real market movement, and the previous period is no longer comparable.

### Pin target pages

When several pages on your site could answer a prompt, pin the one you want cited as the prompt's target page. The citability audit then analyzes exactly that page.

## The weekly check

Fifteen minutes once a week is enough for most brands. Look for what changed, not for absolute numbers.

1. Dashboard: the four metric cards show the change against the previous period of equal length. Note which metric moved and on which platform in Visibility by Platform.
2. Alerts: on Pro and Agency, open the alert feed. Each alert names the topic and what crossed the threshold, and opens the affected topic. The Monday Share of Voice digest email gives the same week-over-week picture without logging in.
3. Prompts rollup on the Reports page: find the prompts that flipped between cited, mentioned and absent, and open one execution to read the actual answer. Answer drift on the result page shows whether the answer really changed or was only rephrased.
4. Competitors: check the Share of Voice delta per competitor, and head-to-head for the names you compete with directly. A competitor rising on every platform usually shipped something; rising on one platform is usually a source change there.

Do not act on a single run. Assistants vary between runs, so treat a change as real when it holds across two or three runs or shows in the trend.

## The monthly review

Once a month, set the range to 30 days and work through the deeper views. This is also the natural rhythm for client reports.

1. Reports: the topic by source heatmap shows where you are strong and where you are blind; Avg citation position shows whether your pages are cited earlier or later in the answers. A rising average position often comes before a lost citation. On best-of prompts, check the mention rank too: it shows whether the AI names you first or fifth.
2. Citations and sources: the citation hit-list names the domains cited instead of you, durability shows whether your citations stick (watch the one-day share and the churn warning), source types show where credibility lives in your market, and blind sources show platforms where competitors appear and you do not.
3. Consulted, not cited: in Gap diagnosis, the grey note on a prompt shows for Claude and ChatGPT whether the AI fetched a page of yours and then cited someone else. Those pages are found but rejected, which makes them the cheapest fixes you have.
4. Sentiment: open the negative segment in the sentiment view and read the mentions. Look for the recurring phrase and the source it comes from: an outdated review, an old comparison article, a forum thread.
5. Fan-out coverage: the not covered sub-queries per topic are questions the AI researches where you do not exist yet. Add the frequent ones to your content backlog.
6. Name confusion: check the card for same-name companies the AI cites in your topics, and mark each as This is us or Not relevant.
7. Prompt set hygiene: pause prompts that never cite anyone, including you, add prompts from Search data and Keywords AI surfaced, and check the intent mix in the intent funnel.

## From results to work

Work through findings in this order. Each step is cheaper than the next, and each one tells you whether the next is needed.

1. Classify: Gap diagnosis sorts weak prompts into Absent, Mentioned, not cited and Inconsistently cited, and the Not visible anywhere filter shows the prompts with no appearance at all. With search data connected, sort by Search demand to start with the prompts that have the most searches behind them. Prefer buying-intent prompts over definitional ones.
2. Audit: run the citability audit on mentioned-not-cited and consulted-not-cited prompts first. It compares your page with the winning page and returns the missing concepts, the answer shape issues and a suggested direct-answer sentence.
3. Earn presence: when the audit finds your passage on par with the competitor's and you still lose, the gap is off your site. The earned media view lists the pages the AI relies on for the prompt; those that name only the competitor are your outreach targets, and the citation hit-list shows the same at domain level.
4. Cover the research: for absent prompts and not covered fan-out sub-queries, the content does not exist yet. Add a section or a page that answers the sub-query under a heading that states it.
5. Make citations stick: inconsistently cited prompts and churning citations need anchor content competitors cannot copy quickly, such as original data, real project results and specific numbers.
6. Measure: after a change, watch the prompt in the prompts rollup over the next runs and repeat the audit. Treat every fix as a hypothesis until the citation moves.

Why AI is not citing you walks through the same path for a single prompt.

## Common pitfalls

- Tracking keywords instead of questions. Assistants answer questions, and the qualifiers in a question decide which brands win.
- Too many Brand prompts. They mention you by definition and make your visibility look better than it is.
- Reading one day. A single run is a sample; use the trend and the comparison with the previous period.
- Changing many prompts at once and reading the result as a market movement.
- Deleting prompts to tidy up. Pause them instead; deleting removes their history.
- Mixing languages or markets in one topic, so the score averages answers that were never comparable.
- Tracking every platform when your customers use two. It costs quota and dilutes attention.
- Looking only at scores. The answer text, the cited pages and the sentiment phrases tell you what to change.
- Ignoring name confusion. Citations of a same-name company distort every other number until they are sorted out.
- Expecting the next run to reflect a page change. Assistants need to find and re-read the page first, so judge a fix over several runs.

**[Why AI is not citing you](https://support.seoai.de/docs/why-isnt-my-brand-cited)**

**[Prompts](https://support.seoai.de/docs/prompts)**

**[The intent funnel](https://support.seoai.de/docs/intent-funnel)**

**[Citations and sources](https://support.seoai.de/docs/citations-and-sources)**

**[Query fan-out](https://support.seoai.de/docs/query-fan-out)**

**[Name confusion](https://support.seoai.de/docs/name-confusion)**

**[Alert rules](https://support.seoai.de/docs/alert-rules)**

**[Reducing your costs](https://support.seoai.de/docs/reducing-costs)**