← The audit service · OPUSCRIPT
The Answer-Engine Audit Method
OPUSCRIPT makes accurate records of things that cannot otherwise be verified. This is one of two: the record of what machines say about a business. The other is the record of how a handmade object came to exist.
Your business is being described to prospective customers by machines, in your absence, in conversations you will never see. The Answer-Engine Audit Method is a repeatable way to find out what they are saying, count what they get wrong, and identify the specific corrections that change the answer.
Ten buying-intent questions · four engines · present, absent, error, stale — the divergence is the finding
A prospective customer no longer types your category into a search box and works down ten blue links. They ask a question in plain language and receive one answer, assembled from sources they never see, naming two or three businesses.
If you are not one of those two or three, you were not outranked. You were not in the room. There is no page two to be on, no impression to measure, and no notification that it happened. The customer got a satisfactory answer and never learned you existed.
The harder problem is the one underneath. When you are named, the machine states things about you as fact — your services, your prices, your locations, your credentials — assembled from whatever it could find, including material that is years out of date, material about a different business with a similar name, and occasionally material that does not exist anywhere. The customer has no reason to doubt any of it.
The method does not attempt to fix this by gaming the machines. It does the boring, verifiable thing first: it asks the questions a real buyer would ask, records what comes back verbatim, and counts the errors. You cannot correct what you have not read.
The most common mistake in this work is asking the machine about the brand by name. That is a vanity query. A customer who already knows your name is not the customer you are losing.
A valid battery is built from buying-intent questions — the ones asked by somebody who has a problem and does not yet know who solves it. A minimum battery is fifteen questions spanning all five types below, with at least three of type 1 and three of type 2.
The base case. Establishes whether the business is in the consideration set at all.
The buyer describes a symptom, not a category. Usually the largest source of missed presence, because it requires the machine to make a connection nobody has written down.
Direct competitive framing. Reveals how the machine characterises you relative to named rivals.
The due-diligence question, asked late in the buying process when the deal is nearly closed. The costliest place to have a wrong answer.
Tests factual accuracy directly. Pricing, hours, locations, credentials, staff, services offered.
Paraphrasing the machine's answer destroys the audit. The client's reaction — the thing that sells the corrective work — comes from reading the actual words in the actual tone, especially where a competitor is recommended and they are not.
Record the date, the engine, and whether search or browsing was active. Answers move. An undated capture is worthless six weeks later.
Each is scored 0–100 across every question-and-engine cell in the battery, then weighted into a single figure. Presence carries the most weight because absence cannot be compensated for by accuracy — being described perfectly in an answer nobody receives is worth nothing.
Does the business appear at all in answers to questions its ideal customer would ask? Counted per cell, with partial credit for a passing mention that carries no recommendation.
Of what is stated as fact, how much is true today? Driven by error classes E2, E3, E4 and E7 below.
Where the business is named, are the things that actually win the deal present — the differentiator, the specialism, the credential?
How is it characterised beside competitors? Recommended, listed, hedged, or qualified with a caveat the competitors do not receive.
Where are the machines drawing from? Sources the business controls or can influence are actionable; sources it cannot touch are not. This dimension decides what the corrections can realistically achieve.
Do the engines agree with each other? Wide divergence usually means thin, contradictory source material rather than an engine-specific problem — and it is often the cheapest thing to fix.
Audit Score = 0.30·Presence + 0.25·Accuracy + 0.15·Completeness + 0.10·Framing + 0.10·Sourcing + 0.10·Consistency
Every error found is tagged to exactly one class. The distribution across classes, not the total count, is what determines the corrective work — a business with twelve E3s has a completely different problem from one with twelve E1s.
| Class | Name | What it is, and what it usually means |
|---|---|---|
| E1 | Absence | Not mentioned in an answer where at least one direct competitor is. The default finding, and usually a sourcing problem rather than a reputation problem. |
| E2 | Misstatement | A specific claim asserted as fact that is untrue. Services not offered, wrong location, wrong credential, wrong price. |
| E3 | Staleness | Was true once. Old pricing, a closed premises, a departed staff member, a discontinued line. Generally the cheapest class to correct and the most common in established businesses. |
| E4 | Confusion | Merged with, or mistaken for, a different entity with a similar name. Damaging out of proportion to its frequency, and the hardest class to correct. |
| E5 | Omission | Named, but the thing that actually wins the deal is missing. Not an error of fact — an error of emphasis, and the class most responsive to correction. |
| E6 | Framing | Characterised in terms competitors are not: hedged, caveated, described as expensive or niche or unproven without a stated basis. |
| E7 | Fabrication | A detail with no traceable source anywhere. Rare, and worth flagging separately because it cannot be corrected by fixing source material — there is no source to fix. |
Stated at length and near the front, because this field is full of people promising things that are not deliverable, and the fastest way to be trusted in it is to be the one who says so.
Answer engines are non-deterministic. The same question asked twice in the same session can return materially different answers. Answers also vary by account, region, language, conversation history, whether browsing was active, and which model version was serving that day.
This means an audit describes what a set of engines said, to one auditor, on one date, under recorded conditions. It does not describe what every customer sees. Any practitioner who reports a stable percentage without stating capture conditions is reporting noise as signal.
This category is being built right now by people selling dashboards with confident-looking numbers behind them. The durable position is not a better dashboard — it is being the practitioner whose reports state their own uncertainty, and who therefore gets believed when they say something is wrong.
An audit that admits what it cannot see is worth more to a serious buyer than one that does not, and it is the only version that survives contact with a client who checks.
This method is published for open use. Any practitioner, agency, or business may adopt the method, cite it, or build tooling for it without permission or fee.
The dimension weights in § 3, the error classes in § 4, and the limits in § 5 must not be altered while the OPUSCRIPT name is used — particularly § 5. A version of this method with the limits removed is the exact thing it was written to displace.
Suggested licence: Creative Commons Attribution 4.0. Confirm preferred terms before publication.
The OPUSCRIPT name and marks are owned by CIONAOD Inc. The method is free to use; the name identifies who publishes and maintains it.