How Persipica measures AI answers

Which prompts Persipica sends

You tell Persipica who you sell to. It drafts prompts the way each kind of person you sell to would ask them, in their own words: "which merino socks stop blisters on long walks?" rather than a keyword. You approve the audiences it asks as, and you can add prompts of your own.

27 prompts in each observation by default. You can set anywhere from 9 to 45.
Most prompts don't name you. That is how people find products they haven't heard of, so your mention rate is measured on these prompts only. Prompts that do name you, such as comparisons, are kept apart and never count towards it.
The audiences are simulated. They are profiles used to phrase prompts, not real people or customer data. Each attribute is labelled as your data, a public signal or an assumption, and results always say they come from simulated audiences.

How answers are collected

Persipica covers five assistants (ChatGPT, Claude, Gemini, Perplexity and Grok) and Google's AI answers (AI Overviews and AI Mode).

Chat assistants. Persipica asks them through their official APIs, with each assistant's own live web search turned on. These answers can differ from what a signed-in person sees in the app (see "What an observation doesn't tell you").
Google's AI answers. AI Overviews and AI Mode are read from Google search results, once per prompt. When Google shows no AI answer, that's recorded as "no AI answer shown" and left out of the mention rate. It isn't counted as a miss.
Kept apart. Chat assistants and Google's AI answers are scored separately and never pooled into one number.
Repeated asking. The same prompt can get a different answer each time, so each prompt goes to each chat assistant three times. Perplexity accepts a setting for how varied its answers are, so it is asked three times at each of two settings.
Follow-ups. Where a person would ask a follow-up, the chat assistant gets it in the same conversation, with the earlier turns included.
Country and language. You choose the country and language for each brand. Prompts are written for that market, and asked and read in that language.

What a result shows

Every result shows three things, kept apart so one can't hide another.

Coverage: how many answers arrived. For example, "60 of 81 answers usable · 21 timed out". It's shown whenever any answers are missing, with the reasons, largest first.
Identity: whether a mention is really you. Persipica doesn't yet check what each match refers to, so a match is counted as a mention and labelled unverified. Every score that includes one carries this note: "Brand identity has not been verified. Shared names may produce false matches."
Mention rate: how often answers name you. Persipica works out a rate for each prompt, then averages the prompts. The answer counts are shown next to it, for example "30 of 60 answers", but the percentage is the average across prompts. It can differ slightly from the counts when prompts have different numbers of answers.

Reading a mention rate. A mention rate is read next to the other two, and with its range. Before treating a drop as real, check how complete the observation was and which answers are missing, how many prompts the rate rests on and how wide its range is, and whether some matches could be another company with a similar name. A drop with full coverage and a narrow range is evidence that fewer answers name you. The same drop with a third of the answers missing may only be incomplete evidence.

Likely ranges. Every rate comes with a likely range, written as "Likely 29% to 65% (95% range across 24 prompts)". It is a Wilson interval, worked out across prompts rather than answers. Answers to the same prompt tend to look alike, so one prompt sent nine times still counts as one prompt, and the range doesn't look more certain than the evidence is. It covers the uncertainty from the prompts that were answered. It never covers answers that didn't arrive; coverage shows those. It doesn't cover matches that turn out not to be you either; identity shows how many are unverified.

One prompt isn't enough. A result that rests on a single prompt is shown as "Based on one prompt – insufficient data for a reliable summary.", with no range.

Comparing results. Persipica shows both numbers. In findings and reports it says one is higher or lower only when their 95% ranges don't overlap. Otherwise the change is within normal variation. That is not a finding that nothing changed: it means these two observations can't show it either way, and a real change smaller than the observation can detect looks the same as none (see what an observation can detect).

Where that rule isn't applied yet. Alerts, shared report links and the public proof score still show the raw difference between two observations without checking the ranges. Read those as raw differences, not as established improvements or declines.

Each assistant first. Every assistant gets its own result, and that's the main view. The overall result uses only the prompts every assistant answered, shown as "24 of 24 prompts answered by every assistant". Each assistant counts equally, set before the observation, so one that answered more often doesn't gain extra influence. If an assistant returned nothing, the overall result is marked incomplete. An average of the others may be shown, labelled as a partial average that leaves it out.

AI agents. A rate across Claude, GPT and Gemini works the same way: each agent's rate counts equally, and its likely range rests on the runs each agent completed (the fewest, where they differ), not on all the runs added together.

When an answer doesn't arrive

Assistants sometimes time out, hit rate limits or show a CAPTCHA. Before collecting, Persipica plans every answer an observation should get, and it records what happened to each one: it arrived, it came back blank, every attempt failed, or it wasn't sent, with the reason. "Not sent" is always reported apart from "sent and failed".

Retries. Each request gets a limited number of attempts, each with a time limit, with spaced, randomised waits. Persipica follows an assistant's own instructions about when to try again. Errors that will never succeed, such as a rejected key, are not retried.
Pauses. If an assistant keeps failing, Persipica pauses it. Its remaining prompts are recorded as not sent, rather than as failures, and coverage shows how many.
Blank answers, refusals and "none of these". A blank response is left out of every rate and counted. A refusal, an answer that declines and offers nothing, is left out of the mention rate and counted as refused. An answer that hedges but still offers options is a real answer. So is one that recommends nothing; it counts as not naming you.
Conversations that break off. If a follow-up fails, the earlier answers are kept and the conversation is marked incomplete. Its opening still counts towards what the person was first told.
What an observation costs. Each answer that arrives uses credits, including a refusal or a Google “no AI answer shown”. How many depends on the assistant, because they cost different amounts to ask: see the table. Before it starts, Persipica shows the most it can use, and the charge never exceeds that. Blank answers, failed requests and prompts not sent are never charged, and retries never cost extra. If no answers come back, the observation isn't charged.
If you stop an observation. You pay only for the answers already received, and the partial findings show exactly those answers. An observation that we stop costs you nothing.
If an observation goes wrong because of Persipica or an AI provider. Tell us within 30 days and the credits go back into your current month's allowance.

What an observation can detect

More prompts give clearer results than observing more often. How big a change has to be before two observations can be expected to show it depends mostly on how many prompts each observation sends.

Prompts in each observation	Smallest change two observations will usually show	From a mention rate of 25%, that is a rise to
9	About 73 points	About 98%
27, the default	About 49 points	About 74%
45	About 37 points	About 62%

Smaller moves are within normal variation for that number of prompts. These figures rest on the assumptions below, and they change when the assumptions do.

Starting rate. A mention rate of 25% in the first observation. The calculator below takes any rate.
What "show" means. Persipica's own rule: the two 95% ranges (Wilson intervals across prompts) don't overlap. That is stricter than a two-sample significance test, which would give smaller figures than Persipica will ever state.
How often. 80% of the time. Each figure is the smallest true change for which two observations' ranges separate in 80 pairs of observations out of 100.
This is not the decision rule. The decision is made on the two observations in front of you: do their ranges overlap or not. The figures here say how large a true change must be before that rule will usually catch it. A smaller true change is sometimes shown, and usually isn't.
Prompts are what count. Asking each prompt three times makes that prompt's own result steadier. It doesn't add prompts: 27 prompts asked three times is still 27.
Two observations of the same size, treated as separate samples. The calculation takes no account of whether the two observations reuse the same prompts.
How results vary. Each observation's result is taken to vary as much as a rate across its prompts can. Real results vary no more than that, so the true figure is the same or smaller.
What it applies to. The mention rate on prompts that don't name you, for one result: one assistant, or the overall result. It counts the prompts that got usable answers. An audience with 8 prompts has the limits of 8 prompts, not of the whole observation.
Direction. Towards 50%: a rise from a rate at or below 50%, a fall from one above.

How Persipica knows an answer means you

Matching. Persipica looks for your brand name, any other names you've given it and your website's domain, as whole words. A product name counts only when your brand appears in the same sentence, product card or cited page.
Not verified yet. A match isn't yet checked against what it refers to, so another company or a common word that shares your name can be counted as you. That's why scores that include matches carry the "not verified" note. Open the answers to read each match in context.
Name mix-ups. A separate check reads your website and judges how likely your name is to be confused with others. If the check can't run, the report says "not assessed", never "low".

What an observation doesn't tell you

It's a sample. Rates describe the prompts in your observation, not every question people ask. Different prompts can give different results.
It's a snapshot. Assistants change their answers from week to week. Regular observations show the trend; one observation shows a moment.
Real people's answers can differ. Assistants can tailor answers to a person's account, location and past chats, which no outside tool can see.
Mentions aren't recommendations or sales. A mention rate shows how often AI names you. A mention can be neutral or even negative, and it doesn't show what people buy. Visits and conversions from AI assistants come from your own analytics, which Persipica can bring in.

Repeats on reports

Each prompt shows one answer, with an "All 3 answers" control that opens every repeat and its own result. When the repeats disagree, a note says so, for example "Mentioned in 2 of 3 answers".

Check it yourself

Every answer behind every number is kept. Open any result to read the exact answers, which assistant gave them, when, and the sources it cited. The sample report shows how this looks for a made-up brand. If anything here is unclear, or you want to see it on your own brand, get in touch.