Your AI says one number. Your dashboard says another. Which one is lying?
On every other analytics tool, those are two different computations. The dashboard runs a fixed report. The AI layer writes a fresh query each time you ask it something. When they disagree, nothing tells you.
Here there is one query path. The dashboard, the HTTP API and the MCP tools your editor calls all read the same computed report. Do not take that on faith. Press the button and we will run the comparison against the live demo instance right now, and hash both answers in front of you.
What this does not prove
It proves the surfaces agree with each other about what your stored events say. It does not prove your tracking is right. If an event never fires, or fires with the wrong properties, every surface will agree on the same wrong answer, confidently. That is a different problem and it is the reason the tracking plan and the CI gate exist.
It also does not prove the reports are the ones you would have written. It proves you get the same answer wherever you ask, so that when a number is wrong it is debuggable instead of unfalsifiable.
Questions
- Why does my AI assistant give a different number than my dashboard?
- Because on most analytics tools they are two different computations. The dashboard runs a fixed report; the AI layer writes a fresh query (HogQL, SQL, or similar) each time you ask. Two query paths can disagree, and typically nothing tells you when they do. smolanalytics has one query path: the dashboard, the HTTP API and the MCP tools your editor calls all read the same computed report, so there is no second path available to disagree.
- What is actually being compared on this page?
- For each report we call the public HTTP endpoint (for example GET /v1/trends?range=7d) and the MCP tool your editor would call (trends), then take a SHA-256 of each response. Both requests hit the live demo instance when you press the button. Nothing on this page is precomputed or cached.
- Does this prove my tracking is correct?
- No, and that is an important limit. It proves the three surfaces agree with each other about what the stored events say. If an event never fires, or fires with the wrong properties, every surface will agree on the same wrong answer. Instrumentation correctness is a separate problem, which is why the tracking plan and the CI gate exist.
- What happens if a check fails?
- The page shows it in red and says so. A proof you can only pass is not a proof, so failures are reported rather than hidden. The same comparison runs in CI on every build via the agreement test, which fails the build if any surface disagrees.