LLM Brand Tracking Loop
Use a stable prompt to monitor brand direction faster and more often
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 96%
The loop begins with a highly contextual prompt describing the company, product, target market, buyer roles, competitive position, and desired metrics. It asks the model for bounded estimates of measures such as aided awareness or net promoter score, plus the reasons customers choose the product. The prompt also requires an explanation of how each conclusion was reached. Iteration narrows unhelpfully broad ranges and improves consistency. Running the same prompt through multiple models reveals convergence and disagreement, while comparison with known survey results provides calibration. Once stabilized, the prompt runs at a regular cadence. The resulting numbers remain directional estimates, but their trend can show whether brand health is improving or deteriorating and indicate when an outlier warrants a conventional survey.
Origin
Kipp Bodnar demonstrated an iteratively refined HubSpot brand-tracking prompt on Marketing Against The Grain after discussing research comparing LLM estimates with traditional surveys.
Core principles
- 01Detailed strategic context produces more useful estimates.
- 02Constrain requested ranges to prevent vague answers.
- 03Require reasoning and sources behind every conclusion.
- 04Cross-model agreement is evidence, not definitive validation.
- 05Trend direction can be more actionable than an exact absolute score.
- 06Low-confidence brands still require surveys or additional data.
How to run it
- 1
Define the brand context
Describe the company, offering, category, customer size, target audience, and relevant decision makers. Include known market position without supplying the answer being estimated.
Pro tip Use the same definitions employed in existing brand research so comparisons remain meaningful.
Watch out Ambiguous audience definitions can produce estimates for the wrong population.
- 2
Select directional metrics
Ask for specific measures such as aided awareness, unaided awareness, customer-selection attributes, sentiment, or NPS. State the expected output format.
Pro tip Begin with metrics for which at least one historical benchmark is available.
Watch out An LLM estimate is not a statistically sampled measurement.
- 3
Constrain the response
Require numerical estimates within a narrow range and request concise rationales. This prevents broad, non-actionable answers.
Pro tip Set the maximum range width explicitly, such as ten percentage points.
Watch out Artificial precision can make a weak estimate appear more certain than it is.
- 4
Demand reasoning and sources
Ask the model to summarize how it reached each conclusion and identify the evidence it relied upon. Examine whether the reasoning reflects relevant market signals.
Pro tip Reject answers that rely only on category averages when brand-specific evidence should exist.
Watch out A model may produce plausible but unverifiable source descriptions.
- 5
Cross-check models
Submit the same prompt to multiple capable models and compare the estimates, attributes, and rationales. Investigate material disagreement.
Pro tip Keep model-specific follow-up prompts separate from the shared baseline prompt.
Watch out Agreement between models may result from shared training data rather than independent confirmation.
- 6
Calibrate against reality
Compare estimates with existing survey or internal data and refine the prompt without leaking target answers into it. Establish a confidence threshold for continued use.
Pro tip Test brands with known high and low scores to detect regression toward category averages.
Watch out Do not tune so aggressively to one company that the prompt stops generalizing.
- 7
Track the trend line
Run the stabilized prompt at a consistent frequency and preserve its outputs. Focus on directional movement and use unexpected changes to trigger deeper research or a survey.
Pro tip Record the model name and version with every result.
Watch out Changing prompts or models without documenting the change can create a false trend.
In the wild
The hosts described HubSpot, its market, company-size target, competitive position, and buyers, then requested bounded awareness and NPS estimates. ChatGPT produced results in the general vicinity of known values, while Claude returned an awareness range within a few points and identified similar purchase attributes.
→ The experiment suggested that LLMs could provide useful directional brand intelligence at far lower cost and latency than frequent surveys.
A B2B marketer applies one calibrated prompt to their company and three competitors every Monday. The report stores estimated awareness, sentiment, selection attributes, reasoning, and confidence. A sudden decline in one metric triggers manual review of recent coverage and customer feedback rather than an automatic strategic change.
→ The team gains an inexpensive early-warning system while preserving surveys for validation.
Common mistakes
Using a shallow prompt
A generic request without audience, market, and competitive context encourages generic category-level answers.
Treating estimates as survey facts
Directional model outputs should not be represented as statistically valid measurements or used alone for high-stakes decisions.
Breaking trend comparability
Frequent undocumented changes to the prompt, audience definition, or model can make movement in the reported numbers meaningless.
Is it for you?
Best for
Established brands with enough public discussion to support recurring awareness, positioning, sentiment, and NPS estimates.
Not ideal for
New or obscure brands with little public data or decisions that require statistically defensible survey measurements.
From the transcript
“I've iterated on this prompt five different times to get it to a place where I like”
“please Give me a summary of how you came to the conclusion for these questions”
“what you actually care about in brand tracking is not is less about what the absolute number is and more the trend line of is…”
From the episode
How This GPT-4 Prompt Is Breaking a $257 Billion Industry