The Prompt Recommender

The Prompt Recommender generates a curated set of measurement prompts for your company. These prompts are what GEOforge runs against AI answer engines each week to measure your Share of Voice (SoV) — how often your company shows up versus competitors when a real user asks a question in your category.

You don't need to write these prompts yourself, but you could if you want. The agent can assemble them from a mix of your brand profile, competitor positioning, and live keyword data, and gives you a final shortlist to approve.

What it produces

Each run produces a shortlist of prompts (target: 30) designed to be distinct but on-topic — covering different angles, intents, and audience framings but making sure these do not drift off-topic.

Every recommended prompt comes with:

• The prompt text — the actual query that will be sent to AI engines.
• A category — the angle or intent bucket the prompt falls into (e.g. comparison, how-to, definitional).
• A rationale — why this prompt was chosen and what it's meant to measure.

You review the list, approve the prompts you want tracked, and discard the rest. Approved prompts then feed the weekly SoV measurement pipeline.

What the agent uses as input

To generate prompts, the Recommender pulls from four sources:

1. Your brand profile — name, website, description, and strategic goals from Settings.
2. Your competitors — name and website for each, also from Settings.
3. Sitemap crawls — page titles and meta descriptions from your site (up to 50 pages) and each competitor site (up to 20 pages each). This gives the agent a real-world view of what topics your category is currently writing about.
4. Keyword research — search-volume and competition data seeded from your brand name and description.

Prerequisites

The agent will not run until your brand profile is complete. Required fields:

• Brand name
• Website
• At least one competitor (with website)
• Strategic goals

If any of these are missing, you'll see an error telling you which field needs to be filled in. Fill them in under Settings, then re-run.

Generating prompts

You trigger a run from SignalForge → Prompts using the Generate Prompts button. If you already have prompts in the library, the button becomes Regenerate.

When you trigger a run, the agent:

1. Crawls your sitemap and your competitors' sitemaps.
2. Runs keyword research against seed terms from your brand profile.
3. Sends everything to the recommender model along with your strategic goals.
4. Returns the curated shortlist for review.

A typical run takes long enough that you should expect to wait — the sitemap crawls and keyword research happen live, not from a cache.

Approving and re-running

You don't have to approve all 30 prompts at once. If you already have some approved prompts tracking regenerating will only replace the unapproved prompts.

Scope discipline

The Recommender is tuned to stay on your category. Prompts that drift into adjacent or unrelated topics — questions a real prospect in your audience wouldn't realistically ask — are filtered out before you see them. If you find a recommended prompt that feels off-topic, discard it during review; the agent will not re-suggest discarded prompts on the next run.

What "good" looks like

A healthy approved prompt set:

• Covers several distinct categories (comparison, how-to, definitional, problem-led, etc.) rather than concentrating in one.
• Reads like questions a real prospect would type into ChatGPT or Perplexity, not like SEO keywords.
• Does not mention your brand by name. SoV is measured on neutral, unbranded prompts so the result reflects how AI engines surface you organically.
• Maps back to topics your company has genuine authority on — otherwise the measurement will reliably show a low SoV with no clear path to improving it.

If a recommended prompt fails any of those tests, discard it during review.

Related

• SignalForge Prompts — where you approve, edit, or discard the recommendations the agent produces.
• Brand profile settings — the source of the brand and competitor data the agent uses.