Everything the monitoring layer does — and the things it refuses to.
RhinocerosAI is the instrument, not the intervention. It measures how AI engines answer for your clients, stores the evidence, and hands you a report you can defend. It does not write content, edit sites, or submit anything anywhere.
Every feature below runs across every engine above, in every plan. The per-engine add-on is the category’s standard margin trap and we do not ship it.
Bundled in every tier. 1 of 6 adapters are sourcing answers today (as of 2026-08-27) — the rest are in build, and we publish which.
Prompt sets and prompt groups
A curated set of questions per brand, run on a schedule across every engine in the market.
It does not generate your prompt set for you. It ships a documented recipe instead — the denominator is the client's business, not a model's guess.
In practice
- Prompts organised into groups — two or three phrasings of one intent, rolled up as one tracked question with each phrasing independently reportable.
- Language is a dimension of the group, not a separate prompt, so an English and an Afrikaans phrasing report together and separately.
- Categories and tags cut every metric by product line, funnel stage or campaign.
- The full prompt list is visible in-product and printed with the client report. A hidden denominator is an unauditable one.
Mention detection and prominence
For every captured answer: was the brand named, where, and among whom.
It does not infer intent behind a mention. If you need to know why an engine described a client a certain way, read the answer — it is one click away.
In practice
- Alias-aware — trading names, abbreviations and the version with the suffix all count as the brand.
- Prominence is recorded, not averaged away: first-named in a three-brand answer is a different commercial event from a footnote in a list of twelve.
- Competitor presence in the same answer is captured at the same time, which is what makes share of voice computable rather than estimated.
- Every detection is attached to the stored answer that produced it.
Share of voice and mention rate
Two different numbers, reported separately because they answer different questions.
It does not produce a single 0–100 visibility score. That figure is easier to read and impossible to defend when a second tool disagrees.
In practice
- Mention rate — how often the brand appeared at all. No competitor denominator, so it can rise for a whole category at once.
- Share of voice — the brand's share of all mentions across the declared competitor set. Zero-sum, and meaningless without a published set.
- Both are reported per engine and as a market-weighted aggregate, not a flat average across engines nobody locally uses.
- Both are bands with run counts. Always.
Competitor benchmarking
The same prompt set, the same runs, side by side against the set the client's board would recognise.
It does not track competitors you have not declared. Discovering the set is an analyst's job with an analyst's judgement, and a silently expanding denominator is how a trendline becomes meaningless.
In practice
- Head-to-head visibility across every engine, with per-engine breakdowns that show where a rival is winning specifically.
- Overtaking moments are flagged when two bands stop overlapping — not when a wobbling line crosses another wobbling line.
- Competitor citation sources are surfaced, so you can see *what* is recommending them.
- Optional control brands per market: a fixed reference so platform-wide drift can be separated from your client's actual movement.
Citation and source analysis
Which domains shape the answers in a category, and which of them the client owns.
It does not tell you how to get cited. It tells you precisely which pages are doing the citing, which is the input to that decision.
In practice
- Every cited URL extracted, normalised to a domain, and classified as brand-owned, competitor-owned or third-party.
- Owned-domain sets, not a single domain — regional sites, legacy domains, docs subdomains. Getting this wrong understated one live brand's owned citations by 5.2×.
- Click-through from any source to the answers that cited it.
- Churn-aware: source overlap for identical prompts runs 34–42% day to day, so a domain cited once is noise and a domain cited across a fortnight is a source.
Trend, history and model events
Every metric over time, with the platform's own changes marked on the line.
It does not backfill history for a brand added today. The first trend point needs two completed cycles, and we will not fabricate the first one.
In practice
- Fortnight-over-fortnight as the default comparison window, because that is the shortest period in which movement can exceed sampling.
- Overlapping consecutive bands are reported as *no detectable movement* rather than drawn as an arrow.
- Model-version changes appear as annotations. One provider update redistributed roughly 47% of citations within 48 hours; an unexplained cliff on a client's chart is a meeting you do not want.
- Every historical point retains the run count and coverage it was computed from.
Six engines, every tier
ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and Google AI Mode — bundled, never per-engine.
It does not scrape consumer interfaces. That buys fidelity at the cost of terms-of-service exposure your client would inherit.
In practice
- Assistants acquired through grounded model APIs; Google surfaces through a SERP vendor. Both declared per figure.
- Adapters are isolated behind one interface, so a vendor is a swappable decision rather than an architectural one.
- Partial coverage is reported honestly — a cycle that reached five of six engines says so and widens its bands, rather than presenting a five-engine aggregate as a six-engine one.
- Copilot is deliberately not included, and is available if a market asks for it.
Markets and languages
A market is a configuration object — engine mix, geo, languages, surface availability, intent rules.
It does not offer eighty countries today. It offers one country configured properly and an abstraction that makes the second one a configuration exercise.
In practice
- Geography is a request parameter recorded on the capture, not a phrase injected into the prompt.
- English and Afrikaans at launch, isiZulu following. Language explains roughly a third of answer variance.
- AI Overview presence as a tri-state — present, absent, or not detected — because vendor detection tops out near 68%.
- Local intent rules stop a map-pack query being scored as an AI Overview miss.
Alerts
Told when something moved further than sampling explains — and not otherwise.
It does not alert on week-to-week wobble inside overlapping bands. An alerting system that fires on noise trains people to ignore it.
In practice
- Share-of-voice movement, fired on non-overlapping bands rather than on any change.
- Stability-class transitions: a prompt moving between always, contested and never.
- New competitor citation on a prompt you hold; a reliable citing domain going quiet.
- Coverage failure — a cycle that finished without an engine, so a quiet dashboard is never mistaken for a stable one.
AI-readiness diagnostic
A read-only check of whether AI crawlers can reach and parse what you want cited.
It does not fix anything, and it never will. A tool that both scores you and sells you the fix has an incentive problem no disclosure repairs.
In practice
- Per-bot access for GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, PerplexityBot and Google-Extended — including CDN and WAF blocking, which is the common silent failure.
- Content in the initial HTML. No major AI crawler executes JavaScript.
- Answer-first chunk structure, then schema as a supporting signal.
llms.txtreported as informational only, with the evidence stated plainly.
God-view and white-label reports
Every client on one screen; every client report carrying your identity, not ours.
It does not let you rebrand the engine palette or the semantic colours. An engine's hue is how a reader identifies it across every chart; it is information, not decoration.
In practice
- God-view — all brands, their movement, their coverage and their cycle health in one place.
- Agency logo, accent colour and domain across the client-facing report.
- Export by printing the same page the client reads, so screen and PDF cannot disagree.
- The methodology footer is not removable — it is what makes the figures above it defensible.
Read API and export
Tenant-scoped, token-authenticated access to everything the interface shows.
It does not offer a write API. Configuration changes that silently alter a denominator should be made by a person who can be asked why.
In practice
API live · connectors on roadmap- Brands, prompts, results, scores, sources and individual captures.
- CSV from any results table; branded PDF from the report.
- Every API figure carries its band and run count. There is no simplified numeric endpoint that drops the sample context.
- Looker Studio connector, alert webhooks and an MCP server are on the roadmap.
What is next, and what we are not promising
The roadmap is published because an agency signing a twelve-month retainer on top of a platform deserves to know what it is betting on.
| Next | What it is | Why it is not shipped yet |
|---|---|---|
| Prompt intelligence | Suggested prompts scored by intent class and AI Overview likelihood | The honest version needs real demand data, not a model guessing at questions. Shipping the guess would make the denominator problem worse. |
| GA4 and Search Console connectors | Correlating AI visibility with actual referral traffic and search context | Every report already ships a “verify it in your own analytics” recipe; the full ingestion is the fast-follow. |
| AI crawler analytics | Log and CDN ingestion showing which AI bots actually visited | The only signal in this whole discipline free of sampling noise, and the one requiring the most access to a client's infrastructure. |
| Consumer-interface calibration | Periodic sampling of consumer surfaces to calibrate the API instrument | It is the honest path to a fidelity claim. It is also a vendor dependency with terms-of-service exposure, so it is a decision rather than a sprint. |
| Adaptive sampling | More samples spent on contested cells, fewer on deterministic ones | Needs enough history per market to classify reliably. The stability class shipped first because it is the useful half. |
| Additional markets | The second country configuration | The abstraction is built and validated on one market. The second is a customer-led decision, not a speculative build. |
See it against your own client list
A working demo runs your prompts, in your market, on live engines — not a sandbox with seeded data. Bring one client brand and three competitors.