# Which questions do you test, and why?

> How the questions are chosen, which engines answer them, how a verdict is read, and why the same questions get asked again and again.

Updated September 24, 2026 · AI Syndicate team · https://www.aisyndicate.com/platform/prompt-tracking/

**Quick answer:** We test the questions a real buyer would ask an AI assistant about your category and area, not generic filler. The Prompt Simulator suggests questions generated from your own site and industry, you add your own, and each run asks 12 AI engines at once. Saved questions keep every run, so trends come from dated history, not one-off screenshots.

## What is a prompt, and what is prompt tracking?

A **prompt** is the question typed into an AI assistant, such as "best med spa in Scottsdale". **Prompt tracking** means asking the same prompts on a schedule and recording what each engine answers, so you can see change over time.

The words our platform uses:

| Term | What it means |
| --- | --- |
| Prompt run | One question sent to all 12 engines at the same moment, with twelve real answers back, each checked for your brand. |
| Tracked prompt | A saved prompt whose every run is kept. It is the raw material for trends, share of voice and War Room battles. |
| Suggested prompts | Ready-to-run buyer questions generated from your real site and industry. |
| Citation and rank | Whether an answer mentions or links your brand, and where: first pick, further down, or not at all. |
| Grounded vs training | Whether an answer was built from a live web search or recalled from the model's memory. |

## How are the questions chosen?

From four places, all tied to your actual business:

1. **Suggested prompts from your site.** The Prompt Simulator generates category-level questions your buyers ask, from your real site and industry. Its own rule: no generic "best CRM" filler for a med spa.
2. **Your own questions.** Type any question a buyer would ask. If a customer told you what they searched, test those exact words.
3. **Gap prompts.** The Competitors feature lists the buyer questions where a rival gets cited and you don't. For each one you can watch the answer that left you out.
4. **Checkable fact questions.** Hallucination Watch asks real, checkable questions about your prices, hours, address and credentials, and judges each answer against your verified facts.

A good set mixes kinds of question. The examples below are made up:

| Kind | Example | Why we test it |
| --- | --- | --- |
| Category and place | "best med spa in Scottsdale" | The shortlist question most buyers start with |
| Problem | "who does laser hair removal for dark skin near Scottsdale" | Buyers describe the need, not the category |
| Comparison | "Example Med Spa vs Glow Studio" | Asked late, close to a decision |
| Your name | "is Example Med Spa any good" | What AI says when someone already knows you |
| Fact | "is Example Med Spa open on Sundays" | Wrong answers here cost real visits |

In a full engagement, the audit maps which questions matter before anything is built. We don't publish the exact prompts we build for clients; our methodology page explains that those are the part competitors would copy. You always see every prompt used for your own account.

## Which engines answer each prompt, and how?

Every run asks 12 engines at the same moment: ChatGPT, Claude, Google AI Overviews, Gemini, Perplexity, Copilot, Siri, Alexa, Grok, Meta AI, DeepSeek and Mistral. Our [methodology](/methodology/) groups them by how a buyer reaches them, and says how each is read:

| Group | Engines | How we read them |
| --- | --- | --- |
| Assistants | ChatGPT, Claude, Gemini, Perplexity, Grok, Mistral, DeepSeek, Meta AI | Probed directly |
| Search surfaces | Google AI Overviews, Copilot | Read from the live result |
| Voice | Siri, Alexa | Measured by reconstruction |

Siri and Alexa have no public answer APIs, so both run as reconstructions on the engines each assistant actually uses, and they are labelled that way on every run. Where Copilot's answer can't be read from a live result, the reading is labelled a reconstruction too.

Runs are live. Nothing is a stored screenshot or a demo script, which is also why answers can change between runs. OpenAI says ChatGPT "may search the web automatically when your question would benefit from current information", so the same question can be answered from a search one day and from memory the next. [OpenAI Help Center](https://help.openai.com/en/articles/9237897-chatgpt-search) That is why every answer carries its source label.

## How is each answer judged?

- **By name and by link.** Each answer is checked for your brand both ways.
- **With the proof shown.** The exact sentence that mentions you is highlighted in place, so you can check the verdict yourself.
- **With your rank.** Where you sit in the recommendation order: first, fourth, or absent.
- **Honestly when you're missing.** If your name never comes up, the verdict says "not cited", plainly.
- **With its source labelled.** Live web search, the engine's own search surface, a voice assistant's pick, or model memory. These fail for different reasons and get fixed differently.

Doubt a verdict? Run the prompt again and watch the answer come back live.

## Why ask the same questions again and again?

Because one answer is an anecdote. A saved prompt keeps every run, so "we're slipping on this question" becomes a dated fact instead of a feeling. That history is what share of voice (how often you are named versus rivals) and trends are built from.

It is also how we judge our own work. In an engagement, the same surfaces are asked the same questions again in weeks 6 to 12, against the baseline captured in the audit, and that comparison decides what gets built next.

When a tracked prompt keeps losing, the **War Room** turns it into a battle: the pages beating you, the engines that could flip, and a step-by-step plan.

> Google says there is no special trick for its AI features: pages need to be indexed and eligible to show with a snippet. [Google Search Central](https://developers.google.com/search/docs/appearance/ai-features) Tracked prompts tell you whether the ordinary work (clear pages, correct facts, trusted sources) is paying off.

## Which plans include prompt tracking?

Every plan. The Prompt Simulator has no per-feature cap; each run draws on your plan's monthly tokens. Plans start at $99 a month for AI Pulse. See [pricing](/pricing/), or try the [Prompt Simulator page](/prompt-simulator/) to see a run.

## FAQ

### Can I pick my own questions?

Yes. Type any question a buyer would ask, or start from the suggested prompts generated from your own site and industry.

### Why do answers change between runs?

Every run is live, and engines change what they search and say. Tracking the same prompts over time is how you separate a trend from noise.

### How do you test Siri and Alexa without an API?

Both run as reconstructions on the engines each assistant actually uses, and they are labelled as reconstructions on every run.

### Why don't you publish the prompts you use for clients?

The specific prompts we build are the part competitors would copy. Clients see every prompt used for their own account.

## Sources

1. [ChatGPT search](https://help.openai.com/en/articles/9237897-chatgpt-search) — OpenAI Help Center. Read Sep 24, 2026.
2. [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features) — Google Search Central. Read Sep 24, 2026.