Skip to content
SilktideHelp

Measuring visibility

On a schedule, Silktide asks every active across the AI assistants you have enabled, stores each answer verbatim, and analyses it: every brand named, the order they appear, how favourably each mention describes the brand, and every source URL used as evidence.

Because AI answers vary from run to run, a single answer means little and trends mean everything. Repeated measurement is what makes mention rate, position, and sentiment trustworthy. The Presence sidebar shows when this workspace was last tested. Test now (or Retest, after the first pass) starts a measurement without waiting for the next scheduled run.

How often we measure

Most plans measure every prompt daily. Some plans measure on a different rhythm - anywhere from twice a day to weekly - and plans with frequency control can give individual prompts their own schedule, so a brand-critical question can be asked daily while a long-tail one is asked weekly.

Measuring less often trades sample size for coverage: each day's reading rests on fewer answers, so single days are noisier, while the trend over weeks stays just as dependable. The metrics are built so that this trade is safe to make:

  • Every prompt counts once. Visibility averages each prompt's own mention rate, so a prompt measured fourteen times in a window carries exactly the same weight as one measured twice. Changing how often something is measured does not move your scores - only how quickly they settle.
  • Points are averages, not single days. By default each point on a daily trend averages the trailing 7 days (you can change this on the chart, from no averaging up to 28 days, or switch to weekly points). A slower measurement schedule still produces a complete, comparable reading at every point.

Confidence, shown honestly

Hover any line on a visibility trend and a shaded band appears behind it: the 95% confidence range for that reading, given how many answers it rests on. The band is wide when the evidence is thin - a workspace measured for the first time yesterday, or a short averaging window - and narrows as answers accumulate. It never disappears entirely, because even a full window is finite evidence.

The band is how to tell movement from noise: a wobble that stays inside it is what repeated sampling of an unchanged reality looks like, while a shift that walks outside it is a real change worth investigating.

Assistants we measure

Silktide queries the consumer surfaces people actually use. Each result records which underlying model produced it:

  • ChatGPT (OpenAI)
  • Google AI Overviews - the AI answer block inside Google Search (different from the Gemini app)
  • Google AI Mode - Google's chat-style AI answer inside Search (different from AI Overviews and from Gemini)
  • Gemini (Google)
  • Microsoft Copilot
  • Perplexity
  • Claude (Anthropic), where your plan includes it

Google's own search results are always measured alongside - Google prints its AI answer on its results page, so measuring both costs no more than measuring one - and some plans add Bing. Which assistants a workspace measures is its own choice: see Settings → Platforms.

See AI assistants we measure for per-assistant detail, and AI assistant market share for why these surfaces matter.

Model changes

AI companies change consumer models regularly. When that happens, your visibility can shift overnight through no action of yours. Silktide records the exact model behind every measurement and marks model changes on trend charts: a vertical marker appears on the day an assistant's model moved, and hovering it names the assistant and both models - an unexplained cliff becomes an explained event.

How answers are collected

Answers are measured through each assistant's API with live web search enabled where available. That closely approximates - but is not identical to - the consumer apps, which add personalisation and session context. No tool can perfectly reproduce a personalised chat; what matters for optimisation is consistent, comparable measurement over time.

Google AI Overviews and Google AI Mode are special cases: they appear inside Search rather than a chat API. Overview text Silktide receives can occasionally omit a word or break mid-sentence. Frequent measurement still makes trends reliable - see those pages for detail.

When a response fails

A response can end in one of three states, and two of them are normal:

  • Answered. The assistant replied and Silktide analysed the answer.
  • No answer shown. The assistant deliberately showed nothing. This is a real result, most commonly on Google AI Overviews and AI Mode, and it is measured as an absence rather than an error.
  • Failed. Silktide could not get an answer at all.

A failure means the measurement did not happen: the assistant was unreachable, refused the request, or took too long. Nothing is recorded, so a failed response is left out of your visibility figures rather than counted as an absence. It does not lower your scores.

Occasional failures are expected. Some assistants are slow enough that a small share of requests time out, and every one of them has outages. Silktide retries on the next scheduled measurement, so an isolated failure needs no action from you.

What is worth reporting is a pattern: one assistant failing on most prompts, day after day. That suggests a problem on our side rather than a bad moment on theirs. Silktide records the underlying technical reason for every failed response, so tell support which prompt and which assistant and we can say exactly what happened without re-running anything.

What 'cost' means in customer terms

Roughly, how much you measure is driven by active prompts × enabled assistants × how often you ask. Brands and deleted prompts do not multiply that the same way. Workspaces do: each topic belongs to one workspace, so two workspaces covering the same subject means two topics and two sets of questions asked. Deleting weak prompts and focusing assistants is how you keep measurement sharp without drowning in noise.

Last updated

Was this page helpful?