Answer
How do you review work you did not write?
Check the claims against something outside it, then check what it does not cover. Reading it carefully is the least productive part.
Verify the specifics against something outside the document, then check what it fails to cover. Reading for quality finds least, because the qualities that signal care in human work are produced for free here.
Reviewing a colleague's work relies on signals that are informative because they were expensive to produce: care in the phrasing, hedging where they were unsure, the structure of an argument that was actually built. Those signals are produced at no cost here, so they carry no information, and a reviewer applying the usual method is reading the least informative part of the artefact with their full attention.
So the order changes. Verification of specifics comes first: every figure, name, date, citation, reference and quotation checked against something outside the document. This is mechanical, it is fast because the checks are independent, and it is where a disproportionate share of errors sit. Marking the specifics and checking only those finds more in ten minutes than an hour of careful reading.
Coverage comes second and is the check nobody performs. What is not addressed, which case is not handled, which stakeholder is not considered, which question the brief asked and the document does not answer. This requires comparison against the requirement rather than against the document, and it is the failure that reading cannot surface because there is nothing on the page to notice.
Structure and argument come third, and this is where reading is genuinely useful. Does the conclusion follow from what precedes it, are the steps present, does anything load-bearing rest on an assertion. Generated work is frequently strong here and occasionally contains a gap that is invisible at sentence level because each sentence is well formed. Reading for the joints rather than for the prose is the productive form.
There is one specific thing to look for that has no analogue in human work: the confident bridge. Where the material did not contain something, a plausible value or claim will have been supplied in the same register as everything around it. These are identifiable by asking, for each specific, where it came from, and they are the errors most likely to survive to publication because nothing about them looks different.
Finally, be honest about volume. If more work arrives than can be reviewed at this depth, the answer is to reduce the arrival rather than the depth, because a lightly reviewed document carries the same risk as an unreviewed one and additionally creates a record implying it was checked. This is the uncomfortable part and it is the one that decides whether the review means anything.
You are not reading for whether it is good; you are reading for whether it is true and whether it is complete, and those are different activities.
Siddharth Sharma, Context Theory
Related questions
Should you review it differently if you would have written the same thing?
That agreement is worth less than it feels, because you are comparing against what you expected rather than against what is correct. The specific hazard is that a document matching your prior view gets a lighter check exactly when your prior view is the thing that should have been tested.
Can the review be delegated to another system?
The mechanical layer, usefully — missing coverage against a stated requirement, unsupported claims, inconsistencies — particularly where the reviewer sees the artefact and the requirement without the reasoning that produced it. The judgement about whether the work should exist and whether it is right for this situation is not that layer.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | The signals a reviewer relies on in human work — careful phrasing, hedging where uncertain, visible construction of an argument — are informative because they were costly, and they are produced at no cost here, so they carry no information. | Comparing defect detection from reading alone across human-written and generated work of similar length. |
| Response | Verification of specifics is mechanical, independent and fast, and finds a disproportionate share of errors, so marking and checking only the specifics outperforms extended careful reading. | Timing a specifics-only check against a full careful read of the same document and comparing errors found. |
| Constraint | Coverage requires comparison against the requirement rather than against the document, which is why reading cannot surface it: there is nothing on the page corresponding to what was omitted. | Listing the requirements the work was meant to address and checking each against the document. |
| Software | The confident bridge — a plausible value supplied where the material lacked one, in the same register as everything else — has no analogue in human work and is identifiable only by asking where each specific came from. | Tracing each specific in a document back to the source material it should have come from. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one