The SignalOS loop
The SignalOS loop is four stages — measure, diagnose, act and prove — and every number in it traces back to a real AI answer. SignalOS asks your buyers' questions across ChatGPT, Claude, Perplexity and Gemini, five times per engine, stores every answer, then shows why you're missing, what to do, and what changed.
Last updated 2026-09-14
How does SignalOS measure what AI says about a brand?
SignalOS builds a question set from your category, your products and services, your competitors, your audiences and your markets. You review and approve it before anything runs. Each approved question is then asked on each engine five times, in a clean session with no memory or prior turns, and every answer is stored in full.
An approved, versioned question set
Approving a question set freezes a numbered version, so a trend never quietly compares two different lists. Candidate questions can also come from your Search Console data, and you accept or reject each one.
Five samples per question, per engine
A single AI answer is one draw from a distribution, so every question is sampled at the same depth: five times by default. Depth is reduced only when a budget or measurement volume runs low, never below three. If three samples cannot be afforded, that engine is not measured in that run, and the product tells you which engine and why.
The model version on every sample
Each stored answer records the engine, the model, the model version and the time it ran. Answers not in English are labelled and excluded from rates rather than guessed at.
Questions that stop telling you anything get measured less often
Questions whose results do not separate you from competitors are measured every three weeks instead of every cycle, and seasonal questions are measured less often out of season. When they are measured, they get the same number of samples as every other question. Refreshing the set produces a diff you approve, never a silent replacement.
How does SignalOS find out why a brand is missing?
Absence has different causes, and they need different fixes. Sometimes AI crawlers cannot reach your pages. Sometimes the sources an answer is built from never mention you. Sometimes the answer names you but positions you for the wrong buyer. SignalOS tells these apart with deterministic rules over stored evidence, not with an AI model's opinion.
AI crawler access
Checks your robots rules against named AI crawlers, including GPTBot, ClaudeBot, PerplexityBot and Google-Extended, and whether you publish an llms.txt file.
Retrievability
Scores a page against seven signals that decide whether content survives being extracted: a direct early answer, question-shaped headings, self-contained sections, entity definitions, dated figures, real tables and schema. A separate check renders pages in a real browser to find content that only appears with JavaScript.
Citation ownership
Every cited source is sorted into your own domain, a competitor's own site, a third party that materially shaped the answer, or a third party that merely appeared in it. That split is what shows whose content AI trusts instead of yours.
Positioning
Classifies which segment an answer places you in, such as enterprise, small teams or budget-conscious buyers, so you can see whether it is the position you want.
What does SignalOS tell you to do about a gap?
Knowing that a competitor's sources are winning is not an action. SignalOS turns each citation gap into a specific play with its own tracked stages, prepares the brief for it, and releases work at the pace your team can actually take on. Which of these a plan includes is shown on pricing.
Four kinds of play
Outreach to a third-party publisher the answers rely on. Your own profile on a review platform. A Wikidata correction, prepared with a notability assessment and a draft entity. Original research, when no good source exists for a question cluster at all.
Guesses are labelled as guesses
An outreach play suggests likely contacts, and they stay marked as guesses until someone confirms them. Whether a placement has started appearing in AI answers is recorded by a person, not claimed automatically.
Content briefs, not content
A brief lists the questions to target, the entities to cover, a structure that survives extraction, a schema type, internal links to your published pages, and third-party sources your answers already rely on.
A worklist paced to your capacity
You declare how many hours or items a team can take on per period. The worklist fills with the highest-priority startable work up to that line, holds anything blocked by an unfinished prerequisite, and orders the rest as backlog.
How does SignalOS show what changed and why?
A visibility rate that moved is only useful if you know why it moved. SignalOS breaks each change into named causes, always shows the part it cannot explain, and states how long a piece of work normally takes to reach an AI answer before judging whether it worked.
A score waterfall with its residual
Change is split into content published, new citations earned and competitors gaining ground, each with the questions behind it. The unexplained remainder is always its own bar, never absorbed to make the chart look tidy.
Lag stated on every impact reading
Each kind of work has a waiting window before a reading means anything. Once your own history holds enough completed examples, the window is learned from them instead of a default. A reading taken too early says so.
Confounders flagged, not hidden
If a model version changed, other work overlapped, or you recorded an outside event inside the window, the impact reading is flagged and its causal confidence is lowered.
Model updates marked on trends
A change in a provider's model version is recorded, raises an alert, and appears as a marker on trend charts, so a model change is not mistaken for a change in your position.
Can you see the actual AI answer behind a number?
Yes. Click a data point on the Visibility Matrix, the Citation Flow chart or the Score Waterfall, and the evidence drawer opens beside it. It shows the raw text of the AI answer, with every mention of your brand highlighted, next to the engine, the model version and the time the sample ran.
The drawer also shows the run distribution: each of the samples taken for that question in that run, marked as included in the number, excluded because the answer was not in English, or not completed. A rate that rests on fewer answers than it appears to is visible at a glance.
Every chart resolves its evidence through one shared lookup, so the drawer means the same thing wherever you open it. This is the check we would want as a buyer: not a number to trust, but the answer to read. The full method behind it is on our methodology page.
Raw AI answer · engine · model version · run timestamp · run distribution
What does a week with SignalOS actually look like?
There is no fixed measurement day. A scheduler checks every client daily, and a new cycle starts when your last completed measurement is old enough for your plan's cadence, which is every two weeks by default. Your cycle day is set by when the previous one finished, and every available engine is measured in the same cycle.
Other checks run daily in the background. Each question's differentiation tier and measurement frequency are recomputed, question clusters are checked for unusual drops, and any alert rules you have set run on the schedule you chose. Nothing waits for someone to log in.
Findings come from audits. Running the answer audit against a completed measurement, or the crawler and retrievability audit against your site, produces findings with a diagnosis and an opportunity attached. Re-detecting the same issue updates the existing finding rather than creating a duplicate.
Your team turns opportunities into recommendations and approves them. The paced worklist then releases only what fits the capacity you declared, highest priority first. If no capacity has been declared, it releases nothing rather than dumping every recommendation on a team at once.
When work ships, you record it with its date and the measurement it started from. Impact is read later, not the next morning. By default a technical fix waits a week before a reading counts, while earning third-party citations waits three. Every reading states its window and whether it was taken too early.
Agencies run the same loop across many clients. What that adds is on for agencies.
Where can you read more about each part?
The full inventory of what is built, and an honest list of what is not, is on features. How sampling, confidence intervals and exclusions work is on methodology. The REST API, the MCP server and webhooks are documented on developers.
Questions about the product
- Where do the questions come from?
- SignalOS generates them from your category, products and services, competitors, audiences and markets, and you approve the set before anything runs. You can add questions yourself, accept candidates from your Search Console data, and refresh the set later, which produces a diff you approve rather than a silent replacement.
- Can I see the actual AI answers?
- Yes. Every answer is stored in full. Clicking a data point on the main charts opens the evidence drawer, which shows the raw answer text with your brand highlighted, the engine, the model version, the time it ran, and which samples were included in the number.
- What if a model updates and my score moves?
- Every sample records the model version that produced it. When a provider's version changes, SignalOS raises an alert and marks the change on trend charts. Any impact reading whose window contains that change is flagged as confounded, with its causal confidence lowered.
- Do you write content for me?
- No. SignalOS generates the brief: the questions to target, the entities to cover, a structure that survives extraction, a schema type, internal links and sources to cite. Writing the piece is yours, or your agency's.
- Is there an API?
- Yes: a REST API, an MCP server and webhooks, on plans that include API access, which pricing lists. Responses carry a provenance block with the sample size, a confidence figure where a calibrated one exists, whether the figure is measured or modelled, and a link back to the evidence. Details are on developers.
See the loop run against your own brand.
A walkthrough runs on a brand you actually care about, yours or a client's, so you are reading real answers rather than a demo.