Every SaaS product has a feedback mechanism. Most of them are useless. A feedback button that opens a text field. A survey that asks "How likely are you to recommend us?" A support email that gets triaged but never analyzed. The feedback comes in. It goes into a spreadsheet. Nobody reads the spreadsheet. The product team makes decisions based on their own intuition and the loudest customer complaints. The feedback loop is broken because the loop never closes.
We designed mani's feedback loops differently. The primary feedback mechanism is not a form. It is the swipe. Every time you approve or reject a piece of generated creative, you are giving the system actionable feedback. Approve a headline with a question format? The engine notes that you prefer questions. Reject three consecutive creative pieces with heavy text overlays? The engine notes that you prefer minimal text. This implicit feedback is more honest and more frequent than any explicit feedback mechanism.
The swipe feedback has three advantages over explicit feedback. First, volume. Users give 20-50 swipe signals per day. No explicit feedback mechanism generates that volume. The statistical significance of 50 daily data points is incomparably better than one NPS score per quarter. Second, honesty. When someone fills out a feedback form, they perform. They write what they think they should want, not what they actually want. Swipes are instinctive. You approve what feels right without conscious filtering. Third, specificity. A feedback form says "I want better creative." A swipe pattern says "I approve warm-toned images with short headlines and reject cool-toned images with long headlines." The specificity is actionable.
The implicit feedback feeds into the generation engine through a preference model. The model tracks patterns in your approvals and rejections across multiple dimensions: color warmth, text density, headline length, imagery style, copy tone, and compositional structure. Over time, the model builds a profile of your preferences that supplements your Brand DNA. The DNA defines what your brand looks like. The preference model defines what you want your brand to look like. They usually align, but when they diverge, the preference model refines the DNA.
We also have an explicit feedback mechanism, but it is designed for specific moments rather than general sentiment. After your first week, we ask one question: "What was your biggest surprise, positive or negative?" After your first month, we ask: "What is missing?" After each billing cycle, we ask: "Would you recommend mani? Why or why not?" Three questions, timed to capture different stages of the user journey. The responses are read by a human (me, currently) within 24 hours.
The performance feedback loop is the third layer. When an approved creative runs on a platform, the performance data flows back into the system. If a creative with a specific headline style gets 3x the CTR of average, the engine increases the weight of that style in future generations. If a creative with a specific visual composition gets half the average CTR, the engine decreases that weight. This is the deepest feedback loop because it measures actual audience response, not just founder preference.
The three loops work together. The swipe loop captures founder taste. The explicit loop captures strategic feedback. The performance loop captures audience response. A creative approach that the founder likes and the audience responds to gets heavily weighted. A creative approach that the founder likes but the audience ignores gets a flag for review. A creative approach that the audience responds to but the founder rejected gets surfaced as a suggestion: "This style performed well with your audience. Consider approving similar creative."
The compounding effect of three feedback loops running simultaneously is significant. After 30 days, the engine has 600-1,500 swipe signals, 1 explicit feedback response, and 2-4 weeks of performance data. After 90 days, it has 2,000-4,500 swipe signals, 3 explicit responses, and enough performance data to identify statistically significant patterns. The engine at 90 days is materially better than the engine at day 1, and the improvement is specific to each user's brand and audience.
This is the feedback loop that most AI ad tools miss. They treat generation as stateless: same input, same output, forever. Mani treats generation as stateful: the output improves over time because the engine learns from every interaction. This learning is the strongest retention mechanism in the product. After 90 days, the engine knows your brand well enough that switching to another tool means starting from zero. The feedback data is the moat.
Mani's feedback architecture is intentionally designed to close the loop. Every signal is captured, processed, and applied. No spreadsheets. No triaging. No decision bottleneck. The system learns continuously, and the learning shows up in better creative, every day, without any human intervention beyond the daily swipe session.
The engineering implication is that the generation engine cannot be stateless. Every generation must be informed by the history of previous generations, approvals, rejections, and performance data for this specific user. This requires a per-user state layer that grows richer over time. Most AI ad tools treat generation as a stateless function: same input, same output. Mani treats generation as a stateful function: same input, better output, because the engine knows more about you today than it did yesterday.