By Ayan Chakraborty, Founder, SBA Innovation Pvt Ltd, Kolkata. Last updated . Reviewed by Ayan Chakraborty.
How many prompts should a lender track for AI visibility?
Start with about 100 prompts and grow only if you can keep running them consistently. In our consumer lending engagement we used 300 prompts across rates, eligibility, documents, safety and RBI registration and head-to-head comparisons, which gave enough coverage to see movement by prompt group. A smaller set is fine if it covers each intent cluster and you run it the same way every time.
Group prompts by intent so you can see where you win and lose: rules and compliance, calculators and metrics, product questions, trust and registration, comparisons and brand queries. Use real borrower wording, mix English and Hindi where your audience does, and include your own brand questions. The sample below is illustrative; replace the bracketed items with your own lender and products.
| Cluster | Example prompt | What it tests |
|---|---|---|
| Trust and registration | Is [Lender] RBI registered and safe? | Entity facts and registration |
| Eligibility | Who is eligible for a [Lender] personal loan? | Extractable eligibility answer |
| Documents | What documents do I need for a [Lender] loan? | Plain-HTML document list |
| Fees and rates | What is the processing fee on a [Lender] loan? | Accuracy of fees and rates |
| Comparison | Which personal loan lender should I choose in India? | Recommendation and shortlist |
| Brand | What is [Lender]? | Entity clarity |
From our client work: the method behind the figures
The consumer lending engagement ran 300 borrower prompts across five assistants each month, three times per prompt, and recorded the median result. The prompt set and scoring rubric were shared in full with the client because they contain only the questions borrowers ask.
Method: Engagement 031: neutral India-located session with no chat history; model updates were not controlled for, which is a real limit on precision, so the prompt set, competitor list and rubric were frozen at the start.
What did not work: Model updates inside the period are not controlled for and limit precision; we cannot separate their effect from our own changes.
Read case study 031 Prompt set, baseline, assumptions and footnotes.