Which prompts will AI actually recommend you for?
Not whether AI mentions you, or whether it gets your facts right — whether it puts you forward as the answer when the user is asking for a suggestion. And for which kinds of prompts. Entidex projects your entity onto a canonical set of prompt classes so you can see, per class, how likely each engine is to still recommend you at the end of the conversation.
Ask ChatGPT, Claude, Gemini, Perplexity and Grokthe same question three different ways and you get three different lists. “Best CRM for a small team” is not the same prompt as “we tried a few CRMs and they were all confusing — what would you actually recommend?” The first is Discovery; the second is Constrained final choice. Any entity can be strong on one and invisible on the other. AI Recommendation Potential shows exactly where you sit on each.
A canonical taxonomy of prompt classes
A prompt class is a user-intent shape that AI engines answer differently. The framework covers eight canonical classes — three we measure today from stored Decision Survival probes, five more described here that light up as the probe families expand. Planned classes always render as null for an entity; nothing is fabricated to fill a gap.
Discovery
MeasuredOpen-ended recommendation prompts — the user asks the engine to suggest something and the entity either surfaces or does not.
Evaluation under pushback
MeasuredFollow-up prompts that challenge an early recommendation with an objection — the entity has to survive the pushback to still be recommended.
Constrained final choice
MeasuredDeep multi-turn prompts that force a single recommendation under real-world constraints — the hardest test of AI recommendation potential.
Hiring and vendor selection
PlannedPrompts framed as choosing a supplier, agency, professional or partner — a purchasing decision with a shortlist step.
Learning and informational
PlannedPrompts that ask AI to explain a topic or entity — the entity either features as a canonical reference or is never cited.
Category leaders
PlannedPrompts asking who leads or defines a category — the highest-authority signal an entity can earn from AI engines.
Risk and safety
PlannedPrompts about safety, trustworthiness or downsides — the entity can be praised, avoided, or (worst case) surface only in warnings.
Adjacent recommendation
PlannedPrompts asking for something like a reference entity — the entity is surfaced when it is seen as a peer of a comparator that is already loved.
One score per class, per engine, against real conversations
For each measured class, Entidex replays 3-5 turn scenarios against every AI engine, then pools the outcomes into a per-engine recommendation rate on 0-100. The number answers a specific question: at the final turn of this class of conversation, is the entity still recommended?
Strongest classes
The prompt classes where AI engines are most likely to still recommend the entity at the final turn. Where the story lands.
Weakest classes
The classes where the entity gets displaced by a competitor or dropped entirely as the conversation deepens. Where the work is.
Per-engine per class
Not just an overall figure — a per-engine breakdown per class, so “Perplexity recommends us for discovery but Claude drops us under pushback” becomes a specific, fixable read.
Every class carries a 95% Wilson interval so a thin sample never reads as certainty; the overall figure is the pooled proportion across measured classes. The score is a coded reducer over stored probes — zero LLM cost, no model grading itself.
Strong on discovery, weak under pressure — a fixable read
A per-class breakdown turns “AI recommends us” from a mood into a matrix. A brand can be recommended cleanly on open-ended Discovery prompts and then quietly displaced as soon as the user pushes back on cost, timeline, or fit. That is a very different story from a brand that never surfaces at all — and it points to a very different fix. Strengthen the sources engines cite when they answer that class of prompt (analyst grids, comparison pages, honest trade-off write-ups), then re-probe: the class-level rate moves as engines adopt the new sources.
Provenance under every reading
Every class-level rate is anchored to the probe runs that produced it — which engine, when, and which scenario. Nothing is asserted without evidence, so the score stands up to scrutiny from a board, a customer, or your own team.
Go deeper with Entidex
The full entity intelligence platform — beyond Explore Entidex
- Continuous multi-source entity observation
- Alerts when sources diverge or drift
- Cross-surface consensus + visibility over time
- Evidence-anchored intelligence reports
Frequently asked questions
What is AI Recommendation Potential?
It is a per-prompt-class breakdown of how likely AI engines are to recommend your entity. Different user intents get different answers — a prompt like "best CRM for a small team" (Discovery) and a prompt like "given a tight budget, pick one" (Constrained final choice) draw the entity in and out of the recommendation in very different ways. Entidex projects stored Decision Survival probes onto a canonical prompt-class taxonomy so an owner can see which classes the engines lean toward for their entity — and which ones they never surface it for. The overall figure carries a 95% Wilson interval so a thin sample never reads as certainty.
How is it different from AI Visibility or LLM Knowledge Accuracy?
AI Visibility asks whether the engines surface you at all. LLM Knowledge Accuracy asks whether what they say about you is true. AI Recommendation Potential asks a colder question: when the user is asking for a suggestion — not just information — will the engines put you forward, and for which kinds of prompts? A brand can be highly visible and factually well-known while never being recommended for the queries that matter to the business.
Which prompt classes do you measure today, and which are planned?
Three are measured directly from stored Decision Survival probes: Discovery ("what are the best options in {category}?"), Evaluation under pushback ("you mentioned X, but I heard Y — who else should I consider?"), and Constrained final choice ("given these constraints, pick one and tell me why"). Five more are described publicly as the framework — Hiring and vendor selection, Learning and informational, Category leaders, Risk and safety, and Adjacent recommendation ("if I like X, what else?") — but they are not yet probed. Planned classes are returned as `measured: false` with a null reading; nothing is fabricated.
How is the score computed, and what does it cost to run?
The score is a coded reducer over stored Decision Survival probes — a deterministic aggregation that runs at zero LLM cost. For each measured class, per-engine survival counts across replayed 3-5 turn conversations are pooled into a recommendation rate on 0-100, carried with a 95% Wilson interval. The overall figure is the pooled proportion across measured classes. It is a confidence-backed metric, not a §3 Score — no model grades recommendation likelihood.
What do I do once I see a weak prompt class?
Two things. Strengthen the sources AI engines cite when they answer that class of prompt — publish the verified record, ensure the entity ranks and is described accurately on the authority sites those engines read, and add the retrieval anchors that carry your recommendation-fit story. Then re-probe. Decision Survival runs on a cadence; the class-level rate moves as engines adopt the new sources.