SignalOS Insights
SignalOS Insights is where we will publish research on how AI answers get made: how often they change, which sources shape them, and which AI crawlers sites block. Nothing is published yet. Every piece will state its method, use real numbers from the measurement the product runs, and say what the data cannot tell you.
What is SignalOS Insights?
SignalOS Insights is the research section of SignalOS, an AI search visibility tool that measures how ChatGPT, Claude, Perplexity and Gemini describe a brand. The research asks questions about AI answers themselves, rather than about any one customer: how stable an answer is, what it draws on, and what stops a site from being read at all.
Most writing about AI search visibility is published by the tools that sell it, and most "best tools" roundups in this category are written by a vendor on the list. We would rather be a source that gets cited than a name in someone else's roundup, and the only way to earn that is research a reader can check. How SignalOS works explains the measurement this research is built on.
Has SignalOS published any research yet?
No. Nothing is published yet, and this index is empty on purpose. The four pieces below are planned, not written, and none of them has a finding to report. We have not attached dates to them, because a date we might miss is a claim we can't stand behind.
There is no mailing list for research yet. If you'd like to hear when the first piece is out, get in touch and say so — a person reads every message.
What standard will every SignalOS research piece meet?
A piece that can't meet all four of these doesn't get published. They are the same rules the product holds its own numbers to.
A stated method
Every piece says exactly how its data was collected: which engines, how many questions, how many samples of each, and over what window. The method is the one on our methodology page, so a reader can compare the two.
Real numbers, with their sample size
Every figure comes from stored AI answers, and every rate is shown with the number of answers behind it and a confidence interval. A figure that rests on too few answers to defend is left out, not rounded into something that looks firmer.
A "what we can't tell you" section
Every piece ends by naming the questions its data cannot answer, and the conclusions a reader might want to draw that the data doesn't support. That section is what makes research credible, and it is written before the headline, not after.
At least five organisations in any aggregate figure
Where a piece reports a figure across organisations, such as an industry average, that figure includes at least five of them. A figure below that floor is not published, so no single organisation's results can be read out of an aggregate.
What research is SignalOS planning first?
Four pieces, in priority order. Each describes what it will investigate and why it matters. None of them states a result, because none has one yet.
1. How often does an AI answer actually change?Planned
This piece will ask the same questions of the same engines repeatedly and measure how often a brand named in one answer is named in the next, and whether that differs by engine. It matters because if answers move, a single reading records chance, and a tool that asks once is reporting a coin flip as a fact.
2. Which sources do AI assistants actually trust?Planned
This piece will look at the difference between a source an answer cites and a source that shapes what the answer says. It matters because work aimed at a listed link that the answer doesn't rely on is wasted. SignalOS records cited domains today; separating the two needs a method we have not finalised, and the piece will state it before any figure.
3. Is your site blocking the AI crawlers you're trying to reach?Planned
This piece will examine how sites' robots.txt files treat named AI crawlers such as GPTBot, ClaudeBot, PerplexityBot and Google-Extended, and whether a block looks deliberate. It matters because a single rule can quietly undo visibility work, and checking your own takes a few minutes.
4. The AI Visibility IndexPlanned
A planned quarterly index of AI visibility by industry, aggregated and anonymised, with at least five organisations behind every published figure. It matters because a visibility rate means little without knowing what is normal in a category. An industry below the five-organisation floor will not appear.
How will SignalOS research use the product's own measurement method?
SignalOS research will not use a separate, research-only process. The data comes from the same measurement the product runs for every brand it tracks, so the method behind a published figure is the method a customer can already inspect on the methodology page.
That method asks each question five times per engine, because one AI answer is one sample from a distribution, not a fact. It never samples a question fewer than three times: if fewer samples are affordable in a cycle, that engine is not sampled at all, rather than reported from one or two answers that can't show a pattern.
Each sample runs in a clean session, with no memory, no prior turns and no personalisation, and the sampling refuses to use a connection to an engine that can't guarantee this. Every answer is kept in full, with the engine, the model version the provider reported and a timestamp, so any figure in a piece can be traced back to the answers behind it.
When a provider changes its model version, SignalOS records the change. A comparison that spans one is flagged rather than quietly treated as like for like. Rates are reported with Wilson score intervals, which stay honest at the small sample sizes a single question produces.
The crawler piece draws on the product's AI crawler access audit, which reads a site's robots.txt and checks each named AI crawler against it separately, so a rule that blocks one crawler and allows another is reported as exactly that.
What will a SignalOS research piece never contain?
A SignalOS research piece will never contain a finding written before the data exists, a percentage without the number of answers it came from, or an aggregate figure small enough to identify an organisation.
It will not claim that AI visibility work drives revenue. Nobody has shown that yet, including us, and research that assumed it would be building on a claim it can't support.
It will not quietly change its method partway through. If a method changes between editions of a piece, the piece says what changed and why, the same way the product records its own methodology changes.
Questions about SignalOS research
- When will SignalOS publish its first research piece?
- There is no date yet, and we won't publish one we might miss. The first piece planned is on how often an AI answer changes. To hear when it is out, get in touch and ask.
- Why does SignalOS research include a section on what it can't tell you?
- Because every dataset has limits, and research that hides them invites a reader to draw conclusions the data doesn't support. Naming the limits is what lets someone quote a figure accurately.
- Where will the data in SignalOS research come from?
- From the same measurement the SignalOS product runs: each question asked five times per engine, in clean sessions, with every answer stored alongside its model version. The methodology page describes it in full.
See the method behind the research.
Every figure we publish will come from the same measurement the product runs, and that method is documented in full today.