The SAMPLE Protocol: A 6-Part Lab Test for Grading Any LLM Visibility Tracker's Data Quality

Last updated September 2026.

A share-of-voice chart looks the same whether it came from a hundred prompt runs or one. A citation map looks the same whether the tool queried the chat window your customer actually uses or a developer API that can return something slightly different. Nothing on the dashboard tells you which one you’re looking at.

The SAMPLE Protocol breaks that check into six parts, one per letter, built to catch exactly those blind spots. Run it against a vendor’s public materials or a live trial, and you get a grade instead of a guess.

The six parts, at a glance

Each part gets its own letter grade, A through F, against a fixed pass bar. Here’s what a pass and a fail look like for each one.

Letter	Criterion	Passes (A-B)	Fails (D-F)
S	Sampling depth	States a sample size above one run per prompt, and explains how it aggregates the spread	Reports one run per prompt with no stated sample size
A	API-vs-UI parity	States which surface it queries and reconciles the two, or only claims what that surface can prove	Doesn’t say which surface it queries, or ignores a known mismatch
M	Model coverage	Tracks all five core engines at a real prompt volume with no per-engine add-on	Covers one or two engines, with the rest locked behind steep add-ons
P	Prompt-panel design	Lets you edit the panel and build it from real query data, not a fixed template	Ships a fixed panel you can’t see or change
L	Longitudinal consistency	Versions prompt wording and panel changes, so a score move traces to the market	Changes the panel or wording with no changelog
E	Error handling	Logs and shows blocked, refused, or failed calls instead of averaging them away	Drops failed calls from the average with no flag

How to run each part

Work through these six checks in order. Each one takes a vendor’s docs page, pricing page, or a trial login, and none of them require a signed enterprise contract to complete.

S: Sampling depth

A single AI answer is one draw from a distribution, not a fixed fact. Ask the same model the same prompt twice, on two different days, and you can get two different answers, two different sources cited, sometimes two different brands recommended. A tool that reports one run as “the” score is reporting noise with a confidence label stuck on it.

1. Open the vendor’s methodology or docs page and search for “sample,” “run,” or “significance.”
2. Check whether the tool states a sample size per prompt per engine: once, five times, 100 times.
3. Confirm whether the published score is a single answer or an aggregate across runs.
4. Flag any tool that reports a confident percentage with no stated sample size behind it.

Pass bar: the vendor names a sample size above one run per prompt and explains what it does with the spread. Evertune sets a clear example here: its own methodology states it samples each prompt up to 100 times per model to reach statistical significance (Evertune, 2026), the kind of disclosure this criterion rewards no matter which vendor does it.

A: API-vs-UI parity

A developer API call and a consumer chat window don’t have to share a system prompt, retrieval setup, or even a model snapshot, even when both carry the same brand name on the box. A tracker that only queries one surface and reports it as “the” answer is measuring half the picture, and it usually doesn’t say which half.

1. Find out, from docs, a sales page, or a support answer, whether the tool queries consumer chat interfaces, developer APIs, or both.
2. If it uses an API, check whether the vendor states how, or if, it reconciles that against what a real user sees in the app.
3. Note any plain vendor statement about its collection surface, such as data collected from real chat interfaces.
4. Downgrade a tool that won’t say which surface it queries at all.

Pass bar: the vendor states its collection surface plainly and either reconciles the two or only claims what that surface can prove.

M: Model coverage

A tool that tracks three engines natively is answering three-fifths of a five-engine question: ChatGPT, Perplexity, Google AI Overviews, Gemini, and Microsoft Copilot. A missing engine isn’t a rounding error. It’s a blind spot in exactly the surface where a customer might be reading about a competitor instead of you.

1. List the engines the tool tracks natively, on its base plan, with no add-on.
2. Compare that list against the five core engines your buyers actually use.
3. Check whether any engine sits behind a paid add-on or a higher tier, and price that in.
4. Confirm the prompt volume per engine, not just whether the engine appears on a feature list.

Pass bar: all five core engines are trackable at a real prompt volume with no separate purchase per engine.

P: Prompt-panel design

A citation score is only as good as the prompts behind it. Ten generic prompts about “best software” in a category tell you less than a panel built from the exact phrases your buyers type, mixed with brand, category, and competitor comparisons.

1. Check whether the starter panel mixes brand-name prompts, category prompts, and competitor comparisons.
2. Confirm you can add, edit, or remove prompts, not just view a fixed list.
3. Look for a research or discovery feature that surfaces real prompt wording instead of a guessed keyword.
4. Test whether you can group prompts by product line, market, or audience segment.

Pass bar: you control the panel, and the tool gives you a way to base it on real query data instead of a generic template.

L: Longitudinal consistency

A score that moves is only useful if you know why it moved. If the prompt wording, model version, or sample size shifts quietly between two reports, the movement on your chart is noise wearing a trend line’s clothes.

1. Pull two reports from the tool a few weeks apart and compare the exact prompt wording each one used.
2. Check whether the vendor logs a changelog for model updates or prompt-panel edits.
3. Confirm trend lines compare the tool’s own prior runs, not a blended competitor average.
4. Ask what happens to history when you edit a prompt: a new series, or a silent overwrite of the old one?

Pass bar: prompt wording and panel composition are versioned, so a score change can be traced to the market, not a silent setup edit.

E: Error handling

An AI model can refuse to answer, time out, or return an empty response. What a tracker does with that failure changes the number that lands on your screen, and most dashboards never mention it happened.

1. Ask, or check the docs, what happens to a prompt run that returns a refusal, an error, or a blocked response.
2. Check whether failed runs are excluded from the reported average and disclosed as excluded, or silently dropped.
3. Look for a visible error rate or a “runs completed” count next to the headline score.
4. Confirm whether one bad run can swing a small-sample report, and whether the tool warns you when it does.

Pass bar: failed or blocked calls are logged, shown, and excluded transparently, never folded into the average unlabeled.

Applied: four tools, six grades

The table below runs four tools from this site’s roster, Profound, Temso, Peec AI, and Otterly.AI, through the SAMPLE Protocol. Grades come from each vendor’s public documentation, pricing pages, and stated methodology, the same kind of materials described in every section above, not a live prompt panel run for this piece.

Criterion	Profound	Temso	Peec AI	Otterly.AI
S: Sampling depth	A	B+	B	B
A: API-vs-UI parity	C+	A-	C+	C+
M: Model coverage	B+	B	A	B+
P: Prompt-panel design	B+	B+	B+	A-
L: Longitudinal consistency	B+	A-	B+	B+
E: Error handling	B	A-	B	B+
Overall SAMPLE grade	A-	A-	B+	B+

Profound wins Sampling depth outright. Its Prompt Volumes dataset and URL-level citation maps give it the deepest sampling and citation-analytics setup of the four, the same strength that puts Profound first on this site’s overall citation-analytics ranking.

Peec AI wins Model coverage. Its roster of eight-plus tracked engines is the widest net in this table, useful when a buyer needs a long-tail model tracked alongside the core five.

Otterly.AI wins Prompt-panel design. Its AI Prompt Research feature draws on a pool of more than 10 million daily prompts, so a panel built on Otterly starts from real query language instead of a guess.

Temso ties Profound for the top overall SAMPLE grade, on the strength of API-vs-UI parity, longitudinal consistency, and error handling. It’s the only one of the four that states its collection method outright, real chat interfaces rather than engine APIs, and it keeps monitoring, content, and audits inside one continuously versioned account. Outside the six criteria above, Temso also grades highest on ease of setup and value for money: Starter runs $89 a month against Profound’s workable Growth tier at $399, roughly 4x more. The honest trade is Temso’s integration catalog, which is smaller than Semrush’s or Ahrefs’s, so a team planning to pipe data into a dozen BI tools should check that list before signing up.

Buy Temso first. Buy Profound too, or instead, only when enterprise citation depth is the actual requirement and the budget clears that roughly 4x gap.

How we built this protocol

The SAMPLE Protocol didn’t come out of one clean test run. It came from watching where LLM visibility numbers actually break: a vendor that won’t say if it sampled once or a hundred times, a UI-reported citation a customer can’t reconcile against an API pull, a prompt panel nobody outside the vendor can see or edit.

We built the six criteria and the pass bars above from public methodology pages, pricing and plan documentation, and stated collection methods across the trackers this site follows, not from a live prompt panel run for this piece. Where a vendor discloses its method plainly, like Evertune’s stated 100x sampling or Temso’s stated UI-based collection, that disclosure is what earns the grade. Where a vendor doesn’t publish that detail, the criterion grades what’s actually verifiable, not what a dashboard implies is happening underneath it.

Run the same six checks yourself before you trust any tracker’s number, including ours.

Run it before the next report

A SAMPLE grade under a C, or a single F on any one criterion, is your signal to stop and ask the vendor directly rather than ship the number in a client deck or a board slide.

See how the full nine-tool field scores on this site’s broader criteria at the LLM visibility tracker ranking, and read the full grading formula behind every score on the methodology page.