When we set out to build Mavrtr's pipeline, we benchmarked against the audience-research checklists every marketing blog posts. Half of them turned out to be theater — work that looks like research but doesn't change which ads you'd write.
Here's the ranking we landed on, based on what we keep seeing actually move ad performance.
Layer 1 — Review mining (highest signal per minute)
Reviews are the most underused research source in ecommerce, and the easiest to mine. The reason they win: customers writing reviews are not anchored on marketing language. They write the way they actually talk about the problem.
The pattern we keep seeing: the exact phrasing in a top-voted review is almost always more persuasive than anything a copywriter would have written. We've watched teams paste verbatim quotes into ad copy and see CPM-adjusted CTR jump in the first 48 hours. The data doesn't argue.
What to read first: 1-star and 2-star reviews of your direct competitors. They're a free list of unmet expectations you can address in your ad creative.
Layer 2 — Competitor Meta Ad Library (high signal, fast)
The Meta Ad Library is the single best free competitive intelligence tool in ecommerce. The signal that matters most is time: ads that have been running for more than 60 days are working. Otherwise the advertiser would have stopped paying for them.
Two observations from looking at thousands of ads in the library:
- The hook patterns that recur across multiple competitors are the ones the market has already validated. You can borrow them without copying.
- The formats that scale (static vs UGC vs script video) shift by category. Pinterest-style flat lay scales in some categories; talking-head UGC scales in others. The library tells you which one your category is currently in.
What most teams miss: the first frame of a long-running video ad is the part you should study. It's the hook that survived ad fatigue.
Layer 3 — Reddit (high signal, easy to misread)
Reddit is the only place where you get unprompted, narrative-format customer discussion at scale. The vocabulary you extract from Reddit threads is usually the strongest hook material in any brief.
But Reddit is also where we see teams over-index. Some failure modes we've watched:
- Treating the top-voted comment as representative. It's usually the funniest comment, which isn't the same as the most common opinion.
- Assuming the relevant subreddit's demographics map to the buyer base. They often don't — subreddits skew younger, more tech-savvy, and more opinionated than the median buyer.
- Mistaking a vocal minority for a market signal.
Reddit is great for language. It's unreliable for demographics. We weight it accordingly inside the pipeline.
Layer 4 — Catalog and product analysis (low signal solo, multiplier when combined)
Looking at a store's catalog, top sellers, and price points won't tell you who the customer is — but it tells you what the brand has been monetizing successfully. Combined with reviews and competitor ads, it becomes a third corner of the triangulation. Solo, it's mostly inventory data.
The mistake we see: teams running a "store analysis" that's 80% catalog observations and 20% audience. That ratio gets the priority backwards. The catalog is context for the customer research, not a substitute for it.
Layer 5 — Surveys and direct outreach (lowest signal per hour)
This one is going to be controversial. Survey data is high-cost to collect and almost always tells you what people say they want rather than what they buy. We've seen brand-survey personas that look nothing like the actual buyers a few months later.
There's a place for surveys — measuring NPS, validating a specific hypothesis, post-purchase friction questions. But "do a customer survey to understand your audience before launching ads" is a recipe for collecting low-signal data slowly.
If you have a budget choice between "use a tool that reads public market data automatically" and "run a 200-person survey," we'd take the tool every time. The buying signal in public reviews is stronger than the self-reported signal in a survey.
How we weight this in Mavrtr
The pipeline inside Mavrtr roughly mirrors this ranking. Reviews and competitor ad library get the most weight. Reddit contributes language and angle hypotheses but doesn't drive segment demographics. Catalog data is context. We don't ingest surveys.
What this means in practice: every Mavrtr brief is biased toward the layers that empirically move ad performance, and against the layers that look productive but don't.