How do I build a way to track branded and non-branded prompts?

To track branded vs. non-branded prompts, you want a system that can:

1. Identify the brand terms in each prompt
2. Classify the prompt as branded or non-branded
3. Store the result so you can report on it later
4. Measure trends over time

Here’s a practical way to build it.

---

1) Define what “branded” means

First, create a clear rule set.

A prompt is branded if it contains any of:

• Your company name
• Product names
• Trademarks
• Campaign names
• Common misspellings or abbreviations of those terms

A prompt is non-branded if it does not contain any of those terms.

Example:

• “Best shoes for running” → non-branded
• “Best Nike shoes for running” → branded
• “Compare Acme Cloud vs AWS” → branded if Acme Cloud is your brand/product

---

2) Maintain a brand dictionary

Create a list of brand terms and variants.

Example structure:

{
  "brand_terms": [
    "Acme",
    "Acme Cloud",
    "AcmeAI",
    "Acmé",
    "ACME"
  ]
}

You can store this in:

• A database table
• A config file
• A feature flag / CMS
• A lookup service

Helpful additions:

• Lowercased versions
• Aliases
• Misspellings
• Internationalized versions

---

3) Classify prompts with a simple rules engine

Start with keyword matching.

Logic

• Normalize the prompt: lowercase, remove punctuation if needed
• Check whether any brand term appears
• Mark as branded if yes
• Otherwise mark as non-branded

Example pseudocode

def classify_prompt(prompt, brand_terms):
    normalized = prompt.lower()
    for term in brand_terms:
        if term.lower() in normalized:
            return "branded"
    return "non-branded"

This is the easiest and fastest MVP.

---

4) Store tracking data

For each prompt, store:

• Prompt text
• Timestamp
• User/session ID
• Classification: branded/non-branded
• Matched brand term
• Source channel
• Optional: language, geography, campaign

Example schema:

CREATE TABLE prompt_tracking (
  id SERIAL PRIMARY KEY,
  prompt ტექst TEXT NOT NULL,
  created_at TIMESTAMP NOT NULL DEFAULT NOW(),
  user_id TEXT,
  session_id TEXT,
  channel TEXT,
  classification TEXT NOT NULL,
  matched_brand_term TEXT
);

---

5) Add reporting

Once you have data, build metrics like:

• % branded prompts
• % non-branded prompts
• branded prompt volume by day/week/month
• most common brand terms
• branded vs non-branded by channel or campaign
• conversion or engagement by prompt type

Example questions:

• Are branded prompts increasing?
• Which product names are most searched?
• Do branded prompts convert better?

---

6) Improve accuracy

Keyword matching is a good start, but it can produce false positives/negatives.

Problems

• “apple” can mean fruit or brand
• “acme” might appear in unrelated contexts
• Misspellings and plural forms may be missed

Ways to improve

• Use exact phrase matching with boundaries
• Add context rules
• Use NER / entity recognition
• Train a classifier if you have enough labeled data
• Use a hybrid approach: rules first, model second

Example hybrid strategy:

1. Exact brand match → branded
2. Ambiguous term → model or manual review
3. No match → non-branded

---

7) Add confidence and review for edge cases

For ambiguous prompts, store:

• confidence score
• reason for classification
• human review status

Example:

• “apple watch repair” → branded with high confidence
• “apple pie recipe” → non-branded or ambiguous depending on your rules

---

8) Build an event pipeline if volume is high

If prompts come in at scale:

• Send prompt events to Kafka / Kinesis / PubSub
• Classify in a stream processor or worker
• Write results to analytics DB / warehouse

This lets you track in near real time.

---

9) Consider compliance and privacy

If prompts may contain sensitive data:

• redact PII
• hash user IDs
• set retention policies
• log only what you need

---

10) Suggested MVP architecture

Frontend/app → sends prompt event
API → normalizes and classifies
DB → stores classified prompt
Dashboard → aggregates branded vs non-branded metrics

---

Example end-to-end flow

Prompt:
“Best Acme Cloud pricing for startups”

1. Receive prompt
2. Normalize text
3. Match “Acme Cloud”
4. Classify as branded
5. Store:
   • prompt
   • branded
   • matched term = “Acme Cloud”
   • timestamp
6. Show in dashboard/report

---

Recommended starting point

If you want the fastest path:

1. Build a brand-term dictionary
2. Use keyword/phrase matching
3. Store classifications in a table
4. Add dashboards
5. Later, improve with NLP or ML

---

If you want, I can also give you:

• a database schema
• a Python/JavaScript implementation
• a SQL dashboard query
• or a more advanced ML approach for branded prompt detection