The short answer
Define the decision, matched inputs and mandatory gates before testing candidates. Record actual editions and every planned result, including failures and unobserved checks. Publish a scenario-specific conclusion supported by your evidence, rather than a ranking that hides missing work.
What you’ll take away
- Preregister the useful decision and test gates.
- Separate documentation, actual observation and missing evidence.
- A failed mandatory requirement cannot be averaged away.
Make the comparison answer a real decision
A useful comparison begins with a task and acceptance conditions, not a list of brands. Suppose a fictional creator needs an editable export that preserves field meaning and does not include a deliberately excluded private note. Those two requirements are gates. A handsome preview or a fast export cannot compensate for failing either one.
The analytics-tool and recap-tool guides help choose an acquisition model and shortlist documented capabilities. This guide instead authors and conducts the bounded evaluation itself: what was exercised, what was observed and which conclusion survives the missing cases. Google’s review guidance asks for experience-based support and relevant decision factors; it does not certify a review or guarantee search rankings.
Freeze inputs, conditions and gates before results
Write the protocol before opening the candidates. Record the intended task, candidate edition/plan, environment, date, exact authorized input, expected outputs, gates, measured quantities and stopping conditions. Use the same relevant starting state and comparable observation method. If permissions or editions differ, report that difference instead of silently treating the test as matched.
Our original fixture names hypothetical candidates A and B and three checks: preview readable, field meanings retained and excluded note absent. The latter two are mandatory. Each candidate receives the same declared practice inputs. These are constructed outcomes for learning the method, not actual product tests, private accounts, paid trials or scores for named providers.
Read all six results before choosing a winner
Constructed table: A passes preview, fails field meaning and has no observation for note exclusion. B passes preview and field meaning but fails note exclusion. There are six planned candidate/check cells: three pass, two fail and one unobserved. A has one pass, one fail and one unobserved; B has two passes and one fail.
The decision is hold both for this task: A fails a mandatory check and leaves another unknown; B fails the note-exclusion gate. Reporting B as “best” from two passes would discard the requirement that matters. Reporting A as a50% success rate among observed checks hides its unobserved third check if the planned denominator is omitted. Show the table and gate decision rather than a flattering composite score.
| Check | Required gate | Candidate A | Candidate B |
|---|---|---|---|
| Preview readable | No | Pass | Pass |
| Field meaning retained | Yes | Fail | Pass |
| Excluded note absent | Yes | Unobserved | Fail |
Resolve three common review mistakes
R1 — Constructed changed weights: the author increases preview weight after seeing that a favorite performs well. Keep the original task gates and disclose any new scenario separately. A later question can justify a new comparison, not a rewritten explanation of why the first test supposedly succeeded.
R2 — Constructed missing result: A’s third check was not performed. Keep it unobserved and explain why; a missing trial is neither pass nor evidence that the provider cannot perform it. R3 — Constructed relationship disclosure: the publication also makes a competing tool. State that relationship and distinguish publisher documentation, own observations and unresolved claims. A disclosure alone does not make the evaluation independent.
Write the conclusion at the scope actually tested
Complete constructed verdict: “For the declared editable-export task, neither hypothetical candidate is ready: A fails meaning preservation and has an unobserved exclusion check; B fails exclusion. The preview results do not repair those gates. This fixture establishes no real provider ranking. A genuine review would need its actual editions, inputs, observations and limitations attached.”
In a real authorized comparison, link your own minimal shareable evidence near each important claim. Distinguish documented availability from observed performance and a local result from general reliability. Report dates and relevant plan constraints; check a current actual quote before a spending decision. Do not invent firsthand access, remove failed attempts or call a short test representative of all users.
Keep the protocol and the unresolved cells
The bilingual packet contains the full test contract, six-cell table, constructed verdict and three review cases. The blank record separates decision, candidates, exact inputs, environment/version, gates, observations, failure/missing reason, relationship disclosure and supported conclusion. Its reproduction script checks declared fixture counts and gate outcomes only.
Google’s actual review guidance was read on 1 November 2026. No provider account, trial, payment, product benchmark or independent review was conducted for these examples. This is not another grade in our fixed MagicPost comparison set. Real reader preference and real customer value still require their separate observed evidence.
- Fix the decision and mandatory requirements.
- Record comparable inputs, editions and observation method.
- Keep every planned cell, including failure and missing results.
- Apply the original gates before considering trade-offs.
- Publish only the conclusion supported by actual scoped evidence.
Sources & product references
Product details reflect the publication date. External features and interfaces may change.
