Starter Prompts for AI Citation Tracking: 12 Prompts, 144 Checks

The first version of our starter-prompt generator produced 18 prompts. Run once across four AI engines, three samples each, that is 216 checks. Our Starter plan included 25 checks a month and Basic included 100. The default set could not run once on either. We fixed both ends before this shipped: the default set is now 12 prompts, which is 144 checks a run, and Starter now includes 150 checks a month.

This post covers what the generator produces, why it is built the way it is, and how to size a tracker so it fits your plan.

The blank page problem

Citation tracking asks a simple question: when someone asks ChatGPT, Gemini, Perplexity or Claude about what you sell, do you come up? To answer it you need a list of the questions buyers ask. Until now you had to write that list yourself, and most people stall there.

GET /citations/prompts/starter writes a first draft from four things you already know: what you sell, your brand name, your competitors and who buys from you.

curl -G "https://seoscoreapi.com/citations/prompts/starter" \
  -H "X-API-Key: YOUR_KEY" \
  --data-urlencode "topic=SEO audit API" \
  --data-urlencode "brand=SEO Score API" \
  --data-urlencode "competitors=Ahrefs,Semrush,Moz" \
  --data-urlencode "audience=marketing agencies" \
  --data-urlencode "features=a free tier" \
  --data-urlencode "use_cases=CI/CD checks"

What it returns

This is the real output for our own product, grouped by category:

Two per category, 12 in all. Send per_category=3 for the 18-prompt set, which adds a third question to each category, among them "Is SEO Score API a good SEO audit API?" and "How much does SEO Score API cost?". The response also has a flat prompts list, ready to send to POST /citations/trackers, and an items list that flags which prompts name your brand or a competitor.

The flat list is interleaved: the first prompt from each category, then the second from each. If you cut it to five, you still get five different kinds of question. For our product, limit=5 returns the first prompt from five of the six categories.

Why six categories

These are the kinds of question a buyer asks on the way to a decision: who is best, how two options compare, what else is out there, who has a specific feature, who fits a specific job, and what it costs.

There is a mechanical reason to cover the spread as well. Google says its AI features "may use a 'query fan-out' technique", "issuing multiple related searches across subtopics and data sources" to build one answer. A buyer's single question gets broken into comparison, pricing and feature searches behind the scenes. A tracker that only watches "best X" misses most of where an answer is assembled from. We can't see how the other engines work internally, and we don't claim they do the same.

Our position: most of your prompts should not name you

Of the 12 prompts above, 2 name our brand, both of them comparisons. The 18-prompt set names us in 6. The smaller default is the better ratio, and that is deliberate.

A prompt like "Is SEO Score API a good SEO audit API?" will mention you almost every time, because the question does. It tells you what an assistant says about you, which is worth knowing. It tells you nothing about whether you come up when a buyer hasn't heard of you. If half your tracker is brand-named prompts, your mention rate will look healthy and mean very little.

The prompts worth the money are the ones where you might be absent: "What is the best SEO audit API for marketing agencies?" That is where a competitor is taking the recommendation.

It is templates, not a model

The generator is a table of sentence patterns. No AI model writes the prompts, nothing is fetched, and no check is spent. The same inputs always return the same list.

We chose that on purpose. A model would phrase them more naturally, and our other endpoint, GET /citations/prompts/suggest?domain=, does use one to read prompts off your website. But for a starter set built from inputs you typed, a template has two properties we care about more: it cannot invent a competitor you didn't name, and a pattern that needs something you didn't send is left out, not filled with a guess.

You can see that second rule in the output for a local business with only a topic. With topic=emergency plumber in Grand Rapids and nothing else, the comparison category has one prompt, not three:

What separates a good emergency plumber in Grand Rapids from a bad one?

There is no brand and no competitor to compare, so the generator says so in notes and doesn't pad the list.

What a starter set costs to run

A check is one prompt, on one engine, sampled once. Answers vary from run to run, so each prompt is sampled three times by default and everything is reported as a rate. The default tracker uses up to four engines.

Against what each plan includes:

These allowances went up in October 2026. They were 25, 100, 400 and 1,500 checks, and 5, 15, 50 and 200 tracked prompts. Prices did not change.

Two limits remain and we would rather state them than have you find them on the first run. First, the dedicated AI visibility tools run every prompt daily, and at $5 we don't: Starter is a monthly reading and Basic is roughly every ten days. Daily tracking of even a small set needs Ultra. Second, free keys can generate prompts but cannot run checks, because every check costs us money at an AI provider.

If you want more runs from the same allowance, track on fewer engines. (If you would rather add checks, top-up packs are 500 for $10 or 2,000 for $36 on any paid plan.) One engine at three samples makes the default set 36 checks, which is four runs a month on Starter. Pick the engine your buyers use most.

Creating a tracker in one call

If you'd rather skip the copy and paste, send starter in place of prompts:

curl -X POST "https://seoscoreapi.com/citations/trackers" \
  -H "X-API-Key: YOUR_KEY" -H "Content-Type: application/json" \
  -d '{
    "brand": "Acme CRM",
    "domains": ["acme.com"],
    "competitors": ["Rival"],
    "engines": ["perplexity"],
    "cadence": "monthly",
    "starter": {"topic": "CRM", "audience": "agencies"}
  }'

The set is cut to the tracked prompts your plan has left, and to what a single run can spend of your monthly checks at the engines and samples you chose, so the call doesn't fail on size and the tracker can always run. If you leave cadence out, a starter tracker runs weekly when your allowance covers a month of weekly runs and monthly when it doesn't. The response includes starter.generated, starter.kept, starter.checks_per_run and starter.cadence so you can see what was dropped and why. If you send your own prompts, those are used and starter is ignored.

Limits

• The phrasing is stiff. Templates produce "What SEO audit API do marketing agencies use?", which a person might word differently. Rewrite the list in your buyers' words before you track it. The response says this in notes every time.
• topic has to be singular. "standing desk" works. "standing desks" produces "What is the best standing desks?" and nothing corrects it.
• We don't know what people type into assistants. Nobody outside the AI companies has prompt volume data, and we don't pretend to. These are reasonable buyer questions, not measured ones.
• It is a starting point, not research. It won't find the odd, specific question that your best customers asked before they bought. Your sales inbox will.

The endpoint works on every plan and is not metered. Running the prompts needs a paid plan. If you want to see what a tracker reports before committing, the citation tracking overview has the details.