The anatomy of a report
Every report is one real run of the same pipeline. Here is what goes into it, module by module, and the guarantees that keep it honest.
Reviews in
Real Google reviews for the subject and its nearby same-type competitors, pulled from the Places API.
Extract
One Claude call reads the whole corpus and produces the analytical substrate: dimensions, scores, themes, segments, price reads, the punch-list, and the competitor weaknesses.
Select
The single sharpest weakness is chosen in code, deterministically, by strength then mention count. Not left to the model.
Two outputs
Two more Claude calls turn that one weakness into the marketer's campaign and the PM's ticket, in parallel, both pinned to the same weakness.
What each section tells you
Reputation gap map
A matrix of each business against the four to eight dimensions the reviews actually raise. A cell is scored only when the reviews speak to it; when they are silent it stays blank instead of guessing. The leader on each dimension is marked.
Verbatim quote mining
Recurring themes per competitor, each backed by quotes copied verbatim from real reviews. Every quote is checked to be an exact substring of a stored review before a report ships; a mismatch fails the build.
Customer-segment breakdown
Families, couples, solo and remote workers, tourists. Inferred from who shows up in the reviews and how they talk, with a wins or loses read per business so you can see who is up for grabs.
Price-perception index
One means customers feel overpriced, five means a steal. When a business's reviews never touch price, the index stays blank rather than inventing a number.
Operational punch-list
Three to six concrete moves drawn from the gap map and the rivals' weaknesses, each tagged by priority and effort, so vague sentiment becomes a list you can actually work through this week.
The marketer and the PM
The centerpiece. The single sharpest weakness becomes a marketing angle with ad copy and, side by side, a product brief with a ready-to-file ticket. Both halves are pinned to the same weakness in code, so they never drift apart.
Review reply drafts
Warm, specific owner replies to your own notable reviews, with a copy button. Nothing is ever posted automatically; these are drafts you read, tweak, and send.
Why you can trust the numbers
The whole point of ryvl is that it is real. These are built in, not promised.
Verbatim quotes or it fails
Every quote shown is checked to be an exact substring of a real review. A single mismatch fails the build, so a report can never show a quote nobody wrote.
Honest blanks
A business is scored on a dimension only when its reviews speak to it. Silent reviews leave a blank cell, never a guessed number.
One insight, mechanically
The marketer and PM outputs are coupled to the same selected weakness in code. They cannot drift onto different problems.
Enough data, or nothing
A run needs a real floor of reviews before any AI spend. Thin data stops the run instead of producing a confident-looking but empty report.