How long is a measurement window?

Rolling 14 days, updated after each weekly panel run; single runs are never reported. Panel versions are never pooled.

What will Primato not do?

Never fake reviews. Never astroturfing or sockpuppets. Never undisclosed placements. Never a rank claim, or a number from a single run. Always disclosure, word for word: "Disclosure: I work at Primato / [client] is a client of ours."

Method · Measure → Move → Prove

How Primato measures AI visibility

Primato runs a frozen panel of real buyer questions through ChatGPT and Gemini, signed out, seven times per question per engine, and reports how often a brand appeared: "appeared in X% of runs over the window." Primato never reports a rank.

The standing protocol
Parameter	Standard	Why
Panel	A frozen, versioned list of buyer-intent questions	Any edit means a new logged version; versions are never pooled.
Runs	7 per question, per engine	A single AI answer varies from run to run.
Session	Signed out	No personal history shaping the answer.
Engines	ChatGPT + Gemini	Two engines, stated. Perplexity is a monthly manual calibration, labelled separately.
Window	Rolling 14 days	Updated after each weekly panel run; single runs are never reported.
Log	Append-only	Raw logs are never edited.
Report	"[Brand] appeared in X% of runs over the [date range] window."

Why run every question more than once?

Because a single AI answer varies from run to run and proves nothing on its own. A claim built on one run is a coin flip presented as fact.

When SparkToro and Gumshoe had about 600 volunteers run 12 identical prompts nearly 3,000 times, the odds of getting the same recommendation list twice were under 1 in 100, and the same list in the same order closer to 1 in 1,000. (Search Engine Land, 28 Jan 2026)

That is why a rank is the wrong unit. The question we answer is how often you were named, across enough runs to trust the number.

What is in a panel?

A frozen, versioned list of the questions buyers in a category actually ask an assistant, from early problem-aware questions to "[you] vs. [competitor]" and "[competitor] alternatives".

Our own standing panel has 150 questions in four tiers: 60 problem-aware, 45 qualified, 30 comparison and 15 head queries. For an audit, you see and approve your competitor set before we run anything.

Which AI engines does Primato measure?

ChatGPT and Gemini, on every panel run. Perplexity is checked in a separate monthly manual calibration, and any Perplexity figure is labelled tool-only and uncalibrated.

We name the engines with every number. A caveat we volunteer costs nothing.

Do you use AI-visibility tracking tools?

Yes, as radar. The manual panel is the truth: where the tool and the panel disagree, the panel wins, and no tool number is published without a saved snapshot behind it.

“A dashboard is enough if you can act on it. Here is the harder part, which nobody selling you one will lead with. […] So before you make a decision off a number, ask how many times the question was asked before you were shown it. If you need a direction, buy the dashboard. If you need a number you can put in front of a board, you need the panel behind it.”

— Prakhar Gupta, founder

The mechanism Retrieval → Citation → Recommendation

How a brand gets named.

A brand wins in three sequential stages: retrieval (the machine must fetch you) → citation (it must quote you) → recommendation (it must name you in the answer). Every piece of work serves exactly one stage, prioritised by two artifacts: the prompt panel (what buyers ask) and the citation map (where models look).

Owned

Your own sites

Rebuilt so machines can read, quote and recommend from them: server-rendered pages, direct answers, tables, structured data and dated facts.

Earned

Where models look

Review profiles, communities, roundups and trade press, chosen from the citation map. Every placement disclosed.

Paid

AI ad surfaces

Tested in small, measured experiments where the platforms allow it, and reported separately from organic visibility.

The evidence What raises visibility

What the research says works.

The GEO study (Aggarwal et al., KDD 2024) tested nine ways of rewriting a source against 10,000 queries. Adding quotations, statistics and citations helped most; keyword stuffing gave little or no improvement. (Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024)

Position-adjusted word count, GEO-bench (higher = more visible)
Method	Score	vs. no optimisation
Quotation addition	27.2	+41%
Statistics addition	25.2	+31%
Fluency optimisation	24.7	+28%
Cite sources	24.6	+27%
No optimisation (baseline)	19.3	—
Keyword stuffing	17.7	−8% on position-adjusted word count (Table 1)

Percentages are calculated from the paper's Table 1 against the 19.3 baseline. The paper's own summary: "GEO can boost visibility by up to 40% in generative engine responses."

Clean hands What we will not do

The moat is the record.

Platforms publish rules against manipulating AI answers, and Primato follows them.

NeverFake reviews.
NeverAstroturfing or sockpuppets.
NeverUndisclosed placements.
NeverA rank claim, or a number from a single run.
AlwaysDisclosure, word for word: "Disclosure: I work at Primato / [client] is a client of ours."

Proof of work This website

Built the way we tell clients to build.

We publish structured data and llms.txt: other engines and agents may read them, and they cost nothing.

HTMLEvery page is plain, server-rendered HTML, so nothing needs JavaScript to read.
Crawlersrobots.txt names and welcomes the search and assistant crawlers from OpenAI, Anthropic, Google, Perplexity, Apple, Microsoft and Meta.
SchemaOne schema.org graph per page: Organization, WebSite, WebPage, Service with its price, and FAQPage for question-led pages.
EntityOne description, used verbatim everywhere: "Primato — the AI visibility (GEO) agency for B2B."
AnswersQuestion headings, with the direct answer in the first sentence beneath them.
Mapssitemap.xml and llms.txt, updated with every change.

Find out what the machines tell your buyers.

Book the audit Pricing