How to stress-test AI competitor research before briefing your next ads

In the last article in this series, I showed you how to use AI to analyze hundreds of competitor ads. The output is a report with creative families, thumbnails of the ads in each, and an explanation of what the advertiser is testing.

But you still have hundreds of analyses to work through. You have gone from watching hundreds of ads to reading about them. That is progress, but it still demands time and attention you may not have. And some of the analysis may simply be wrong.

Watching every video yourself would reverse much of the time you saved. Instead, review the clusters, examine selected ads, and use your observations to ask AI to reconsider the entire categorization.

Here is how to do that, with three examples of what to challenge and a master prompt you can use to produce a revised HTML/PDF report.

Why do this manual step?

Yes, this part is manual. That is exactly why we need human review: AI can make things up, miss the main argument, or confidently group ads for the wrong reason. Asking it to verify its own work without giving it specific observations can leave the same mistakes intact.

You do not need to redo the whole analysis. Look at the cluster overview and selected screenshots, check the conclusions against the relevant videos, and point out what seems wrong or incomplete. 

Basically sanity-check the AI’s conclusions.

Three things to challenge

1. Does AI clustering reflect the primary theme or a secondary one?

Is this example ‘Secret progress’ or ‘Emotional connection’?

A woman secretly learns Arabic to meet her fiancé’s family. Their surprise gives way to a tutor demonstration. Should the cluster be ‘secret progress’ or some form of ’emotional connection’? 

The correction explains how she practices; emotional connection gives that practice a purpose. The existing cluster captures the reveal. 

Does the report also capture the promise behind it, or should that promise change how this ad is grouped?

2. Does the AI cluster reflect only the opening hook or the entire ad?

The opening asks whether you understand English but cannot speak it. The body explains the gap through a lack of conversation practice and presents an AI tutor as the remedy. “Level callout” captures the opening. A fuller interpretation might be turning passive understanding into spoken ability. Ask the AI which description best fits the complete ad.

3. Are there other competing candidates for potential clusters for specific ads? Which one is the most salient?

Current family: Secret progress. Candidates to examine: surprise reveal, rapid-fluency promise, and tutor demonstration.

A character surprises a shopkeeper by speaking Chinese. A later pitch promises fluency within a month and shows tutor practice. Surprise, recognition, and product demonstration are all possible ways to describe the ad. Which is most salient to its argument, and which should determine its cluster?

These are challenges to investigate, not three mandatory reassignments. Emotional connection, a surprise reveal, and a correction demo can coexist in one ad. The model should explain which dimension is the most salient – and what the most important groups are.

How to do the actual stress-test: ask AI to review the entire clustering

Give the AI your observations, the existing report, all analyzed ad records, and access to the source videos. The examples below illustrate the feedback format. Replace them with the actual labels and ad references from your report.

I reviewed the cluster overview and selected ads and noticed several interpretations worth reconsidering:

1. The family-surprise ad is categorized as Secret progress. Emotional connection appears to be its central promise: the woman learns Arabic to connect with her fiancé’s family. The surprise reveal is the story device, and the correction demo supports the promise. Should this change its classification or the way the family is described?

2. The comprehension-versus-speaking ad is categorized as Level callout. That describes its hook, but the body argues that conversation practice is the missing ingredient. Are other ads making this argument grouped elsewhere because their openings differ?

3. The shopkeeper-surprise ad is categorized as Secret progress. Its surprise reveal, rapid-fluency promise, and tutor demonstration are competing ways to describe it. Which is most salient to the complete argument, and which elements should be recorded as supporting details?

Based on these observations, review the entire clustering and the assignments of all analyzed videos. Does this warrant re-clustering, recategorization, or clearer family definitions? If yes, go ahead and make those changes throughout the collection.

Do not automatically agree with my interpretations. Where you still strongly support an existing classification, keep it and explain why using evidence from the ads. Apply a consistent approach across the collection. Distinguish the main argument from emotional motivation, opening hooks, and supporting demonstrations where necessary.

Check every video’s assignment against the source evidence, using the complete video where needed. Identify any video you could not verify. Keep advertiser claims separate from observed events and verified outcomes. Do not infer performance from the size or prominence of a cluster.

Return the revised family definitions and all member assignments. Explain what you changed, what you retained despite my challenge, and what remains uncertain. Regenerate the HTML and PDF from this revised analysis, preserving the previous versions so I can review the changes.

“Review the entire clustering” is the most important framing. You are using a few observations to question the model’s approach across the whole collection. It should reconsider both the definitions of the groups and the membership of every ad, rather than fixing only the examples you named.

The model must also be able to access the evidence. A transcript can omit an overlay, and a thumbnail cannot establish a sequence. If a source is unavailable, that ad should remain flagged for review rather than being silently counted as verified.

Review the revised families and report

Start with the change summary. Read the model’s reasons for renaming, splitting, merging, or retaining families. Check its response to your observations, including cases where it disagrees with you.

The result is a reviewed map of the competitor’s creative approaches, with source examples and reasoning behind the categorization. You can now choose an ad or family for deeper study without treating the first AI-generated labels as settled conclusions.

In Article 3, we will take that next step: compare selected ads beat by beat, extract a recurring structure, and use it to develop a new concept without copying the original script.