How Mavzena measures

A prompt is the question put to the engines. Mavzena generates prompts from your brand's category, in the brand's language, in six patterns: recommendation, use case, comparison, alternative, reputation (is the brand reliable, what do reviews say) and brand facts (price, locations, ownership). Each prompt is about one of the products you confirmed or your category, and a local business's prompts name its city. Most prompts do not name your brand, so they show whether engines recommend you without being asked about you. You can also write your own or import them from a CSV file.

Each prompt keeps its source permanently, so a generated prompt is never presented as real demand. Prompts can be tied to customer types, which are labelled by origin too: generated, added by you, or derived from listening signals. Groups such as branded, competitor and unaided are worked out again each time from the prompt text and your current competitor list, so they stay correct when that list changes.

Commercial intent (buying, comparing options, going to a specific site, looking for information) is assigned by fixed keyword rules. It is a classification, not a measurement, and a value you set by hand is never overwritten.

A measurement sends the same prompts to ChatGPT, Gemini, Claude and DeepSeek. Answers are collected in two ways: some engines run a live web search before answering, others answer only from what the model already knows. In a measurement with web search, ChatGPT, Gemini and Claude search; DeepSeek always answers from what it already knows. A brand's first measurement uses web search on every plan; after that, web search comes with the paid plans.

The free plan asks 3 prompts on ChatGPT, Gemini and Claude. Paid plans ask all four engines, 30 prompts per brand. Once you switch automatic measurement on, it repeats by itself at the frequency you choose. The model behind each engine is recorded with the run.

A second model reads each raw answer and returns a structured verdict: is your brand named, at what position, which competitors appear, and is the tone positive, neutral or negative.

If that verdict cannot be parsed, the answer is not recorded as “not mentioned”. It is left out of every metric and tried again at the next measurement. Counting an unreadable answer as absent would quietly lower your score.

We regularly test the step that reads the answers against fixed examples whose result we already know.

Three measurements make up the score. Mention rate is the share of scored answers that name your brand. Average rank is your position in the answers where you appear; lower is better. Sentiment is the tone of those same answers.

The same formula is applied to each engine separately, and to unaided prompts: the ones that name no brand at all. That shows whether engines recommend you without being asked about you.

A brand that is never named scores 0, because that was measured. A brand that was never measured shows no score.

Engines can answer the same prompt differently from one moment to the next, which is why repeats matter. Every prompt is asked once on each engine. In older measurements, prompts that contain your brand name were asked three times on one engine; in the score, those three answers count as one. Choosing more repeats when you start a measurement is not available today.

This is why most results read Low. A measurement that asks each prompt once is a single reading, and the label says so instead of implying certainty.

Each measurement opens a numbered measurement run. It records the prompts, engines, models, answer mode, sample count, start and finish time, and how many requests succeeded or failed. Every answer is stored against that run. A new measurement adds new rows and never overwrites earlier ones, so a trend is drawn only from measurements that really happened.

• If an engine fails, the run is marked partial and reports state how many requests failed. The figures then cover only the answers that came back.
• A brand can have only one measurement in progress. A double click or a retried job cannot measure the same thing twice under two run numbers.
• Not measured is never 0. A value with no measurement behind it is shown as not measured, and a report that cannot be linked to a run says so.
• False-information verdicts are append-only at database level: edits and deletions are rejected. The workspace history is only ever added to.

Mavzena records the pages an answer cites. For ChatGPT it also records the pages the search read but did not cite. Links are stripped of tracking parameters, grouped by site, and sorted by type (review, forum, social, encyclopedia, app store, news, other) using a fixed list of domains rather than a model.

Whether a source is your own site or a third party is worked out when you view it, from your current website address, so changing your site leaves no stale labels. No source is attributed to a competitor by guessing from names. An answer that cites no sources has nothing to record.

Some numbers are measured from answers. Others are estimates built from written rules. Every score on the Overview carries one of three labels: measurement, estimate, or includes an estimate.

Organic visibilitymeasurement

The visibility score described above.

Room to growestimate

The average organic priority of the measured prompts. High means a lot is still on the table.

AI growthmeasurement

70% × organic visibility + 30% × coverage (the share of prompt × AI engine pairs in which at least one answer mentions the brand). Hidden while organic visibility is unmeasured.

The weights are rough and written down, to be calibrated with real data. Opportunity priority (impact × ease × commercial intent) is a ranking rule, not a forecast.

Public Voice follows what people write about you and your competitors on Reddit, in Google Maps ratings and in the RSS feeds you add. For brands in Turkey it also reads Turkish sources (Şikayetvar, Ekşi Sözlük, DonanımHaber, Technopat). It is a different measurement from AI visibility and is never blended into it.

• Each source reports its own status. “Nothing found” and “could not check” are different results and are shown differently.
• Share of conversation is computed only over sources that worked for every brand being compared, so a competitor without a Şikayetvar page cannot make you look louder.
• Tone and topics are assigned by rules. An AI classifier exists, but stays off until it has been measured against hand-labelled posts.
• Ekşi Sözlük text is never sent to an AI model, as the site's robots.txt asks. Collected posts are kept with their authors for 90 days, then deleted.

Every measurement has edges. This one cannot tell you:

• How an engine will answer tomorrow. A few repeats show today's spread, not a guarantee.
• What other model versions would say. When a provider updates a model, the same prompt can get a different answer.
• Anything about prompts that were not asked. A score covers the prompts and engines in its run, nothing more.
• How often real people ask a prompt. Generated prompts describe your category; they are not search volume.
• Exactly what a person sees in the consumer app. Engines are called through their APIs; the apps can add personalisation, location and ads, and ads shown there are not visible to this measurement.
• Revenue. Being more visible in AI answers is not proof of more sales, and Mavzena does not convert visibility into money.
• An accuracy rate. The judge that reads answers and the listening rules have not been measured against hand-labelled data, so no accuracy figure is claimed.