Check if AI assistants recommend your brand when customers ask for advice. See who gets mentioned instead of you, with sources — and what changed since your last check.
Sources now come with links you can open. Until now you got the domain (pcmag.com); now you get the full address of the page the assistant actually drew on, both in the data (sourceUrls) and as clickable links in a new "Sources behind these answers" chapter at the end of the report.
And you can see WHICH assistant cited each source. A site cited by one assistant will not help you with the other two — the report now knows that, and will no longer suggest working on a site to improve your visibility on an assistant that never cited it.
Clearer wording on sources. The three assistants expose different breadths: ChatGPT and Claude list what they cited in the answer, Perplexity lists the wider set its answer rests on. The column is now called what it is — the sources behind the answer — instead of implying they are all the same thing.
The week-over-week comparison ignores questions with partial coverage. If a run was cut short (time or budget) and a question reached only one assistant instead of three, that question is left out of the comparison: otherwise it would look like you gained visibility when all that changed was how many assistants we asked. When nothing was cut short, nothing changes.
You are not charged for the report when we could not measure your competitors. The report is still delivered, but if our extraction failed, the "who wins instead" chapter cannot be written — and you should not pay full price for it. If the extraction ran and genuinely found no competitors, that is a real result and it is charged.
A run that is restarted no longer charges you twice. If the platform moves your run to another machine, or you press "Resurrect", the Actor now recognises that the results were already delivered and stops instead of repeating (and re-charging) every check.
Scheduling, honestly explained. The FAQ used to suggest a setup that quietly compared every run against the same old dataset. It now explains the limitation and the habit that actually works.
0.1.40 — 18 August 2026
Fixed: the insight report could compare two different measures and reach the wrong conclusion. When you tracked named competitors, it could put your share of voice next to a competitor's mention rate — two numbers with different denominators — and conclude you were behind when you were ahead. Every name in the report, yours included, is now measured the same way and ranked for you on both measures. If you ran a report before this date with competitors set, re-run it: the numbers in the dataset were always correct, only the written interpretation could be wrong.
Fixed: being named is no longer counted as being recommended. If an assistant mentions your brand in passing — or to say it is not the right choice — that no longer counts towards your share of voice, your position, or the head-to-head comparison, where competitors could only ever be counted as actual recommendations. Your headline "appears in N of M answers" is unchanged, and the questions where you are named but not recommended are now listed for you.
New: a "Run summary" table. The run summary used to sit in the checks table as a nearly empty row. It now has its own view, with totals, gaps, competitors and the comparison laid out properly.
We now tell you when we could not read your website. If your site blocks us or its robots.txt asks us not to read it, the questions are generated from your brand name alone — more generic, and the language can be wrong. That is now written in the log and in the summary (siteContext) instead of happening silently.
Fixed: sources for Claude. The "Sources cited" column listed domains Claude had looked at as well as those it actually cited. Now it lists only the cited ones.
Rows come out in a stable order — question by question, models always in the same order — so two runs can be compared side by side.
Answer excerpts are cut at a word boundary and end with "…" when there is more, instead of stopping mid-word with no sign.
The comparison with a previous run now says when that run was, and is refused outright if that run could not measure competitors — instead of reporting drops and "new competitors" that never happened.
Hardened the safety check on the website address you provide, so it cannot be used to reach addresses that are not on the public internet.
Clearer pricing: "Number of questions" now spells out what a run costs at both ends of the dial.
0.1.35 — 17 August 2026
Faster, and no more rate-limit failures. Each AI provider is now queried at its own safe pace. Selecting a single model no longer causes most of its checks to fail.
Your results are always delivered. If a run approaches its time limit, the Actor stops taking new checks, delivers everything collected so far and tells you in the log exactly how many were skipped. Skipped checks are never charged.
A full default run (10 questions × 3 models) completes in about 90 seconds.
0.1.27 — 17 August 2026
Claude replaces Google Gemini as the third assistant checked. Claude is queried with live web search, so what you see is what it recommends today.
Comparisons across a model change are refused, not faked. If your previous run used a different set of models, the headline percentages are reported as not comparable instead of being presented as a trend.
0.1.10 – 0.1.26 — 15 to 17 August 2026
First public release and the corrections that followed, in short:
Your questions checked on ChatGPT, Perplexity and Google Gemini, with mentions, competitors, cited sources and an interpretive report.
Week-over-week comparison made to work: when you point a run at a previous dataset, the same questions are reused, so the two runs are comparable. Before that, auto-generated questions differed every time and the comparison was always empty.
Never a number we did not measure: if a check fails, or the competitors could not be extracted, the affected figures are reported as "not measured" instead of as zero. A zero would read as "you are invisible" when it really means "we could not tell".
