Skip to content
SilktideHelp

Accuracy

Accuracy is where you tell Silktide what is true about a and find out which AI assistants get it wrong. You declare a fact, Silktide turns it into a question and an expected answer, asks every assistant you have enabled, and grades what comes back.

This is not the same as visibility prompts. Visibility asks discovery questions and measures who gets mentioned. Accuracy asks about claims you need stated correctly - founding dates, regulated facts, product capabilities, campus locations - and reports whether the assistants agree.

What a fact is

A fact is a checkable statement about a brand, for example:

  • "Northbridge University was founded in 1868."
  • "City of Meridian recycling collection runs weekly for single-family homes."
  • "Helios Energy's residential solar plans include a 25-year panel warranty."

Silktide derives a question and expected answer from each one (for example, "When was Northbridge University founded?" with "1868" expected). You can review the derived pair, but the fact stays the thing you track.

Facts belong to the brand rather than to a workspace. A founding date does not change between the UK and the US.

Adding facts

You can add facts one at a time. You can also paste a larger passage - a blog post, press release, or fact sheet - and Silktide will pull the individual claims out of it for you to review. That is the practical path when marketing has already written the truths and you want assistants held to them.

When assistants already disagree about the brand, Accuracy suggests facts from those disagreements. Review each suggestion, edit it if you need to, and accept it as a tracked fact, or dismiss it.

Every fact is checked as soon as it is created. You do not have to run anything.

Silktide Agent

Paste this About page into Accuracy for Apple and turn the hard facts into things we should monitor.

I'll extract checkable facts from the passage, add them to the Apple brand, and show you the derived questions and expected answers so you can drop anything that is marketing fluff rather than a fact.

Silktide Agent

Create a few suitable facts for Silktide and check them.

I'll add concrete facts about Silktide (derived questions and expected answers are automatic), then summarise which assistants agree or disagree once the checks come in.

Verdicts

Each assistant's answer is graded against the expected answer:

GradeMeaning
AgreeThe answer matches the expected fact
PartialSome of the fact is right, or only part of a compound claim is covered
AmbiguousThe answer is unclear or hedges too much to judge cleanly
DisagreeThe answer conflicts with the expected fact
RefusedThe assistant declined to answer

Those per-assistant grades roll up into one verdict for the fact, which is what the list shows:

VerdictMeaning
DisagreeNo assistant got it right
MixedSome assistants agree and some do not, or they only partly agree
AgreeEvery assistant got it right

The list opens on the facts that need attention, and the tabs above it filter by verdict. A fact that has never been checked has no verdict yet and appears only under All.

Separate from visibility

Why Accuracy is not a visibility prompt

Visibility measurement is optimised for sample size across many discovery prompts. An accuracy question is different by construction: it names your brand on purpose, so its answers are kept out of every visibility, share-of-voice, and sentiment figure - counting them would flatter every number they touched. The answers are still analysed in full, the same way as any other response, so an accuracy answer that mentions a competitor or cites a source is still readable that way on the Answers screen. Both share the same assistants, the same workspace cadence, and the same "ask models, store verbatim answers" foundation. Adding a new fact also checks it soon after, without waiting for the next scheduled measurement.

Last updated

Was this page helpful?