How we measure AI search visibility

Why repeated querying?

“Single-run visibility metrics provide a misleadingly precise picture.”

LLMs are non-deterministic - the same prompt can produce different outputs on every call. A single snapshot confuses random variation with real signal.

We run each query N=5 times per AI engine and report bootstrap 95% confidence intervals - so you see not just the point estimate, but the uncertainty range. See also: arXiv:2604.07585 - “Don't Measure Once”, which establishes N=5 as the practical minimum for stable LLM visibility estimates.

The three measurements behind every score

Metric definitions from arXiv:2604.25707 (per-engine citation divergence + selection rate). Taken across the four prompt families and every engine we run, these three produce the eight scored KPIs the composite combines:

Vertical-specific scoring

The 8 KPIs combine into the composite under a vertical-specific weighting calibrated to how buyers actually discover that kind of business. A SaaS company is evaluated against named alternatives, so comparison and share of voice carry more weight; a publisher monetises referral traffic, so being cited matters more than being named. The weighting is normalised so scores stay directly comparable on the 0 to 100 scale.

Current score model: v2. Scores rendered under earlier models are tagged with the model they were computed under and are not silently re-priced. See your report header for the model badge.

The discovery gap

Research (arXiv:2601.00912) found a stark asymmetry in how AI assistants handle branded vs. open queries:

AI assistants are reliable at answering questions about your company, but rarely surface you unprompted in category discovery. The Agentic Commerce Readiness Score focuses on closing this discovery gap.

Agent Readiness

On top of measuring what AI engines say about your company, we check whether your domain is set up to be read by AI agents in the first place. Each scan fetches five public-web signals from your site. Every signal is weighted equally (0.20) and the composite is rounded to a 0–100 integer. Each fetch has its own 3-second timeout; a timeout, network error, or non-200 response counts as “absent” rather than failing the report.

Booking engine host

Agent Readiness reads your own domain. Your booking engine almost always runs somewhere else — reservation.one.dirs21.de, app.mews.com, login.smoobu.com, buchen.your-hotel.de — and that host has its own bot rules, set by your engine vendor rather than by you. A site can pass every crawlability check while every AI agent that tries to reach the page where the booking actually happens is turned away. So we check that second host too. The free report says so in one line on the Agentic Booking card when that host stops agents; the full verdict, the host and a ready-to-send request to your vendor are on the paid Agentic Booking page. It never changes the Agentic Booking rung.

The check is three probes at most — six requests at the outside, because each probe follows at most one redirect — once per engine host per day, shared across every hotel on that host: the host’s robots.txt, one request with an ordinary browser user-agent, and one with a representative AI-agent user-agent. If robots.txt already disallows AI agents we stop there and send nothing further. No AI model is involved and no booking step is ever reached.

Four verdicts, and what each one means. Agents allowed: an agent reaches your booking page and finds a booking form. Agents blocked: the host’s robots.txt disallows AI agents. Agents challenged: the host answers our ordinary request but intercepts the agent one (a 403, a rate-limit, or a Cloudflare/Turnstile challenge page). Could not check: anything we cannot classify — including the case where our own request was intercepted before it reached your vendor, which we never report as a verdict about that vendor.

One vantage point. Every verdict is measured from Bluerails servers, and the paid Agentic Booking page prints that beside it. A guest browsing from a phone in Germany may be treated differently from a datacentre IP, and a vendor can change these rules at any time. We show the date of the check for that reason.

Why we measure all 4 engines

Different AI engines exhibit meaningfully different citation behaviour. ChatGPT tends to cite fewer sources per response but exercises deeper influence on each cited page. Perplexity surfaces more sources per response but the per-source influence is shallower. Gemini and Claude follow distinct retrieval patterns again. No single engine represents your overall AI visibility - you need all four.

Our consistency score penalises high variance across engines, rewarding companies that are visible everywhere rather than just dominant on one platform.

How we compare

Confidence intervals are the uncertainty range you get from asking the same question repeatedly - competitors' single-run snapshots give you one point estimate with no idea how much it would move on the next run.

Primary research cited

• arXiv:2603.08924Repeated sampling and statistical confidence in LLM visibility measurement
• arXiv:2604.07585"Don't Measure Once" - N=5 minimum for stable LLM visibility estimates
• arXiv:2604.25707Selection rate definition and per-engine citation divergence
• arXiv:2601.0091299.4% named-query recognition vs 3.32% discovery surfacing in AI responses