Read the heatmap column by column. Each row is a signup period, and users get older as you move right. The drop in the first column is usually the steepest, so check it first in your own matrix; the real question is whether the curve flattens after it. Every cell carries its number, and colour only guides the eye.

The pooled curve combines cohorts of different sizes into one line. Because each period only counts cohorts that have observed it, new and still-short rows do not distort the curve. A plateau says the loss slowed within this window; it does not say a lasting loyal base exists. Only a longer observation can show that.

Pattern hints are not diagnoses. Loss concentrated in the first period may point at onboarding and the first moment of value; mid-period erosion with no plateau makes you ask whether the product builds a habit; fluctuation can also signal a definition or data problem. The cohort-over-cohort comparison assumes rows run oldest to newest. A segment difference does not show in a single matrix: paste a second matrix by channel or plan and compare them on the same chart. The hint thresholds are this tool's rules, not benchmarks (for example, the newest three cohorts' M1 average differing from earlier ones by at least 5 points).

Read the unobserved share of LTV separately. If the matrix shows only a few periods and the horizon reaches well past them, a large part of the result may be an assumption about periods you have not seen. The share depends on how fast loss continues: small if loss keeps running fast, large if the curve has flattened. Above 50% (the tool's threshold) the box switches to a warning tone; observing the curve a while longer before committing budget to that number is reasonable.

The scenario block shows what lifting M1 retention by a few points does to LTV. The result can be smaller than intuition suggests: cohort size does not change, only the share remaining after the first period grows. This is a sensitivity calculation; only an experiment shows whether the improvement will happen.

The start month option starts retention and CAC from a later month instead of M0. a16z's "Retention Is All You Need" (2025) recommends starting this calculation at M3 for AI products; it writes that the "tourists" who sign up and leave within a couple of months have mostly gone by M3, and that curves most often begin to flatten around M3. The piece rests on a dataset of dozens of AI companies with more than $1M in ARR that the authors have seen. The authors also say that, as long as the curve eventually flattens, the exact timing matters less for long-term dollar retention. M3 is the month a16z recommends; this tool does not treat it as a fixed threshold: you can take the month from your matrix's plateau point or choose it manually. The tool does not filter out or estimate who is a tourist; it only computes from the chosen month onward. CAC per remaining customer, CAC ÷ S[k], is a rebasing and assumes all acquisition spend is attributed to those still active in that month; it is not a measured CAC. Conditional LTV/CAC leaves out revenue before k while keeping the full CAC, so for k > 0 it is lower than the M0-based ratio; read the two as answers to different questions.