# mAI Visibility — full reference

> AI Presence Management and Generative Engine Optimization (GEO) platform.
> It measures whether and how a brand appears in the answers of AI assistants,
> explains who wins each buyer question and which sources the AI trusts, and
> turns that into a per-question action plan measured before and after.
> Greek-first, sold in Greece and the EU, bilingual (Greek and English).

This file is the complete, machine-readable reference for the product. The
short index is at https://www.maivisibility.com/llms.txt

Last updated: 8 August 2026 · Canonical site: https://www.maivisibility.com/

---

## 1. Identity

- **Product name:** mAI Visibility
- **Aliases:** mAI GEO, mAI Geo, "mAI Visibility by Retail Management Solutions"
- **Category:** AI visibility monitoring / Generative Engine Optimization (GEO) / AI Presence Management
- **Legal entity:** RETAIL MANAGEMENT SOLUTIONS ΜΙΚΕ
- **Registered office:** Γράμμου 73, 15124 Μαρούσι (Maroussi), Athens, Greece
- **VAT:** EL801291892 · **Greek business registry (ΓΕΜΗ):** 153674701000
- **Primary contact:** info@maivisibility.com
- **Founder / product owner contact:** i.kotrotsios@rm.gr
- **Application:** https://app.maivisibility.com
- **Interface and reports:** Greek and English

mAI Visibility belongs to the company's "mAI" product family. Its sibling
product mAI Engage (maiengage.app) covers reviews and customer engagement and
is a different product — do not conflate them.

## 2. The problem it addresses

A search engine returns ten links and the buyer chooses. An AI assistant
returns one answer containing one or two recommendations. A brand that is not
inside that answer is not "ranked lower" — it is absent for that buyer.

AI assistants form their answers from sources that already exist: directories,
articles, comparison pages, and the brand's own site when it is structured
well enough to be quoted safely. GEO is the work of understanding which
sources a model reads for a category, being present in them, and giving the
brand's own site the structure that lets a model cite it.

## 3. Who it is for

- **B2B and consumer brands** in research-and-compare categories, where buyers
  ask an assistant before deciding.
- **Marketing agencies** managing several clients, who need continuous
  measurement and client-ready reports rather than one-off audits.
- **In-house SEO and marketing teams** that already rank organically and want
  to know why the AI answers still recommend someone else.

It is a poor fit where nobody will implement the recommendations, or in purely
local, impulse-purchase categories where buyers do not consult an assistant.

## 4. AI providers measured

| Provider | What runs | Search |
|---|---|---|
| ChatGPT | GPT models | Live web search |
| Google AI | Gemini | Google Search grounding |
| Anthropic Claude | Claude models | Web search tool |
| Perplexity | sonar models | Search built in |

Google AI Overviews — the answer box above Google's organic results — are
tracked separately, because that surface follows different citation rules.

All measurements run with live search enabled. An answer produced without
search reflects the model's training memory rather than what a buyer would see
today, so it is not a valid measurement of present-day visibility.

The measurement cycle is configurable **per provider** (daily, weekly, monthly
or off), per customer. The default is weekly.

## 5. Methodology

Full page: https://www.maivisibility.com/methodology

### 5.1 Unit of measurement

The unit is a **run**: one execution of one question against one provider at
one moment. Every run stores the complete answer, the cited sources, and the
analysis. All rates below are computed over the runs of the **last 30 days**
for a given brand and workspace.

### 5.2 What each answer is analysed for

Brand and declared aliases (Greek, English, abbreviations, former names),
position within the answer, recommendation level, sentiment, attributes the
model assigns, other entities mentioned (competitors, discovered
automatically), and every cited source with an authority score.

### 5.3 Metrics

- **Mention rate** = runs mentioning the brand ÷ total runs × 100
- **Recommendation rate** = runs where the brand is actively recommended ÷ total runs × 100
- **Top recommendation rate** = runs where the brand is the first or main recommendation ÷ total runs × 100
- **Share of voice** = brand mentions ÷ mentions of *all* entities × 100. Note the denominator: it is mentions of every brand, not runs. This is the only inherently relative metric — a brand can gain mentions and lose share.
- **Average position** = mean list position across runs where the brand appears (computed only where present, otherwise absence would flatter the average). Displayed normalised to 0–100, first position → 100.
- **Citation rate** = runs citing the brand's own or related sources ÷ total runs × 100
- **Citation authority** = mean authority of cited sources × 100
- **Sentiment** = weighted sentiment of brand mentions

### 5.4 Recommendation levels

Four distinct states, measured separately because they mean different things
commercially:

1. **Mentioned** — appears somewhere, possibly in a list.
2. **Recommended** — actively proposed as a solution.
3. **Top recommendation** — the first or main recommendation; functions as a default.
4. **Excluded** — the model knows the brand and explicitly rules it out for that need. Usually indicates wrong or missing information in the sources, and is fixable.

### 5.5 Overall visibility index

Weighted blend, published so it can be reproduced:

| Component | Weight |
|---|---|
| Mention rate | 25% |
| Recommendation rate | 25% |
| Share of voice | 20% |
| Average position | 15% |
| Sentiment | 15% |

### 5.6 Measurement context: neutral and persona

Every measurement records the context it was taken under, and the two are never
merged into one number.

- **Neutral** is the primary benchmark: no persona, a fixed system prompt and
  temperature 0, identical for every brand and every run. Comparability depends
  on nothing here varying between measurements.
- **Persona** places the buyer's own words before the question, so the
  assistant answers as it would for that kind of customer. Measured separately,
  per persona.

Neutral is the *absence* of a persona, not a persona named "neutral". If the
benchmark were an editable record, a wording change would silently break
comparability with every earlier measurement.

Persona context is validated before it is stored: it may not name the brand
under test or any competitor, because a question that names the brand answers
itself and inflates every rate derived from it. The same context applies to all
competitors within a run.

Measurements recorded before context modes existed are kept for history but
excluded from the neutral benchmark, because the context they were taken under
is unknown.

### 5.7 Confidence intervals

Every rate is reported with its interval and the number of runs behind it. The
interval is a **Wilson score interval** at the 95% level.

The choice is not academic. The textbook normal approximation breaks down at
exactly the sample sizes real use produces: with 6 runs and 5 mentions it
reports an upper bound above 100%, which is impossible. Wilson stays inside
[0,100] and stays honest when the proportion is near 0 or 1 — which is where
most brands actually sit.

The **width** of the interval is the honest read on how settled a number is:

| Interval width | Stability |
|---|---|
| ≤ 20 points | Stable |
| 21–40 points | Medium |
| > 40 points | Unstable |
| fewer than 3 runs | Not characterised |

Low stability is not measurement error. It means the brand sits near the
model's decision boundary and flips between runs. That is itself a GEO signal:
a small improvement in sources can settle it.

Rule of thumb: at 3 runs, 67% and 33% are statistically indistinguishable. A
rate needs roughly 8–10 runs per context before it can carry a decision, and
the product shows everywhere how far from that you are.

**Average position** is computed only over the runs where the brand appeared.
Absence is not "last place", and averaging a zero in would quietly reward being
invisible.

Separately, each daily history snapshot carries a weight derived solely from
that day's sample size (0.40 below 5 runs, rising to 0.90 at 50 or more). It is
used for trend charts, so a quiet day with two measurements does not weigh as
much as one with fifty. 15–40 questions per category and market remains the
recommended library size.

### 5.8 Handling model variability

Language models are probabilistic: the same question can produce different
answers on consecutive runs with nothing having changed in the world. This is
addressed by measuring many questions rather than repeating one, keeping
history per question and provider so trend separates from variance, and
producing a change explanation for every question that moves — what changed in
the answer, which competitor entered, which source stopped being cited.

A single number from a single run is not a conclusion, and the product does
not present it as one.

### 5.9 Attribution

When a recommended action is marked done, the same question is re-measured on
the same provider and compared with the reading taken before. This is
correlation with a time order, not proof of causation, and the sample size is
always shown next to the change.

## 6. Audiences, personas, relevance and fit

### 6.1 Target Audience Profile

Personas are not generated from the industry alone — that produces the same
five plausible segments for every company in a sector, which is why they
predict nothing. They are generated from the *combination* of a declared
audience's attributes: business size, turnover, operating model, working
environment, technology literacy, ranked needs, pain points, weighted purchase
criteria and budget.

A profile also states explicitly **who the product is not for**. Without that,
generation invents segments that would never buy, and measuring them produces
numbers that mean nothing.

Profiles are built through an eight-question wizard. The AI drafts the
structured fields; the user reviews every one. Fields left empty stay empty
rather than being guessed.

### 6.2 Persona Relevance Score

Ranking and suitability are different questions, and they are kept apart.

An assistant that recommends a multi-thousand-euro ERP to a market stallholder
can put the brand in first place and help nobody. A scoreboard counting only
order of appearance records that as a win.

So every persona measurement is also scored on whether the recommendation
actually suits the buyer who asked, across eight axes: budget, business size,
mobility, features, technology, working environment, complexity and operating
model. It is reported beside the mention rate, never folded into it: the
interesting case is high visibility with low relevance, and a single blended
score would hide exactly that.

### 6.3 Product profile and Product–Persona Fit

Fit is computed from two stored objects rather than asked of a model each time:
the audience states what it requires, the product states what it offers, and
both use the same eight axes. The score is weighted by the user's own purchase
criteria, normalised to sum to 100.

An axis on which the audience expressed no requirement counts in neither
direction. Exceeding a requirement is a full match — over-delivering is not a
defect.

Every claim in a product profile keeps its source and whether it has been
verified. A fit score built on unverified marketing copy is a fit score for the
marketing copy, and the user must be able to see which one they are looking at.
Claims derived from AI answers are marked unverified until confirmed. Deriving
a profile with no crawl, no product entries, no description and no brand
mentions is refused rather than answered: a fabricated capability vector is
worse than none, because every fit score downstream inherits it looking
measured.

Competitor profiles are entered by hand. Generating them would mean inventing
capabilities that have not been verified.

### 6.4 Weighted visibility by audience

A brand may serve more than one audience with unequal weight. Each is measured
separately and combined using the weight the user assigned, so the headline
number describes the real market rather than an average over unrelated
segments. Where no weights are set, tier defaults apply: primary ×3, secondary
×2, experimental ×1.

The weighted figure is renormalised over the audiences that actually have
measurements. An unmeasured audience is missing from the calculation — it does
not count as zero, because something unmeasured is not poor performance.

### 6.5 Persona versioning

Personas evolve as the user learns their market, and a reworded persona asks a
different question. Every change to the wording mints a version, and each
measurement stays pinned to the version it was taken under.

Without this, a wording fix would retroactively redefine every earlier
measurement: the numbers would still add up and would have stopped meaning
anything. With it, a drop in visibility can be told apart from a change in the
question.

### 6.6 Persona heatmap

Brand and competitors are measured within the same runs, per context, which is
the only way the comparison is fair. Cells backed by fewer than 3 runs are left
blank rather than showing a percentage nobody should act on.

## 7. Technical GEO audit

Separate from the AI measurements and using no AI: rule-based checks on the
customer's own site.

Checks include robots.txt, sitemap.xml, schema.org structured data, FAQ
structure, heading hierarchy, hreflang, thin pages, author and organisation
identity, and the presence of llms.txt.

An important implementation detail, corrected on 8 August 2026: many static
hosts (Cloudflare Pages, Netlify, Vercel) answer requests for non-existent
paths with the site's index.html and HTTP 200. A check that trusts the status
code alone reports llms.txt, robots.txt and sitemap.xml as present on sites
that have none. The audit verifies the content type as well.

**Google PageSpeed Insights** is part of the audit. Lab data (Lighthouse
simulation) is separated from field data (CrUX, real users), which exists only
for sites with enough traffic. Speed does not enter the visibility index; it
is reported separately because it affects crawling and citation indirectly.

On top of the rule checks there is a consultancy-grade AI analysis that names
what is missing rather than reporting present/absent: the structure of
llms.txt against the standard, the specific FAQ questions the site does not
answer, authority signals, empty schema fields, multilingual gaps, comparison
content and speed.

## 8. From measurement to action

For every question the brand loses, the platform drafts 3–4 specific actions
on channels the brand controls: pages to create on its own site, sources to
get listed on, corrections to make. Sources are managed as an outreach
pipeline (not listed → requested → listed) with a ready-to-send email per
source.

Content generation is grounded **only** in facts the customer has approved in
the knowledge base, lists the grounded claims, and flags unsupported
statements. llms.txt, Organization schema and FAQPage schema can be generated
for the customer's own site.

**GEO Autopilot** (opt-in): once a week it refreshes recommendations, drafts
the content itself, and emails one-click approve links. With a connected
WordPress the approved piece publishes automatically. Nothing publishes
without explicit human approval.

## 9. Alerts and correctness

Alerts fire when the brand disappears from an answer, drops in rank, or a new
competitor appears. **Hallucination Watch** compares every AI claim about the
brand against the approved knowledge base and — for Greek companies — against
the official ΓΕΜΗ registry record and the company's own website, raising a
critical alert with the correction when the model is wrong.

Competitor site monitoring reports what changed on competitors' sites since
the previous check: new pages, new schema, new content.

## 10. Data, privacy and security

- **What is stored:** the full text of each AI answer, the sources cited, the
  analysis, and timestamps, for as long as the account exists. Without the
  stored answer there is no comparison with previous weeks and no explanation
  of what changed. Deleting the account deletes them.
- **What is sent to third parties:** questions, brand name and description, and
  content approved for generation — necessarily, since measuring what ChatGPT
  says requires asking ChatGPT. End customers' personal data is not sent.
- **Sub-processors (AI providers):** OpenAI, Google, Anthropic, Perplexity.
- **Training:** the platform uses provider APIs rather than consumer apps; API
  data is not used for model training under those providers' terms. The
  platform does not train models on customer data.
- **Isolation:** strict per-organisation data isolation; role-based access
  control; full audit log; optional MFA (TOTP); API keys with scopes.
- **Regulatory context:** operated by a Greek company for the European market,
  under GDPR (Regulation (EU) 2016/679) and Greek Law 4624/2019. A data
  processing agreement is available for agencies on request.
- Privacy policy: https://www.maivisibility.com/privacy
- Terms of use: https://www.maivisibility.com/terms

## 11. Signup, trial and commercial model

- Signup is verified by an email one-time code **and** an SMS code to the
  user's mobile. One phone number per account.
- Entering the Greek VAT number (ΑΦΜ) auto-fills official company details from
  the ΓΕΜΗ registry; the brand description is drafted automatically by reading
  the customer's website and is fully editable.
- The ΑΦΜ is the account's unique identifier. Customers cannot change it
  themselves (support corrects it on request), and free trial resources are
  granted **once per registered ΑΦΜ**. Signup without an ΑΦΜ is allowed but
  comes with no free resources.
- **First month free**, no card, and it activates automatically the moment the
  ΑΦΜ is registered — there is no activation task. Includes €5 of starter AI
  credit, granted once per organization.
- **Referral programme:** €5 wallet credit for each verified signup referred.
- Usage beyond the trial is prepaid through an **AI wallet**, charged per
  action that consumes AI, with per-action cost reporting (operation, provider,
  model, day). The technical audit, PageSpeed, the ΓΕΜΗ lookup and the in-app
  support assistant do **not** consume credit.
- **Subscription pricing has not been announced yet.** If asked for specific
  prices, the honest answer is that pricing will be published soon. Do not
  quote figures.
- No minimum term. Unused credit is retained but usable only with an active
  subscription and is not redeemable for cash.
- Agency-created client organisations are provisioned bare: no trial
  subscription and no starter credit. The agency funds its clients.

## 12. Free public tool

https://www.maivisibility.com/scan — an AI-readiness score for any site, no
account required, about 20 seconds. It runs the technical checks only and does
**not** measure anything inside the AI assistants. One check per site; repeat
visits replay the original analysis with its date.

## 13. Platform capabilities, in brief

Target Audience Profile builder (eight-question wizard) · persona generation
from attribute combinations · persona versioning with runs pinned to versions ·
neutral vs persona measurement contexts · Wilson confidence intervals on every
rate · persona × brand heatmap · Persona Relevance Score across eight axes ·
product profiles with sourced, verifiable claims · Product–Persona Fit ·
audience-weighted visibility · dynamic question generation across seven
question families · batch runs with a per-provider cost estimate · guided
onboarding with progress derived from real data · bring-your-own LLM keys ·
prompt library with AI-assisted discovery · per-provider scheduling · answer
analysis (mention, rank, recommendation level, sentiment, attributes) ·
Prompt × Provider matrix · automatic competitor discovery · share of voice ·
citation and source tracking with an opportunity engine · technical GEO audit ·
Google PageSpeed · consultancy-grade GEO analysis with live streaming ·
analysis history · knowledge base with semantic search and an approval
workflow · Content Studio · automatic llms.txt and schema generation ·
per-question action plans · GEO Autopilot with WordPress publishing ·
before/after attribution · Hallucination Watch with ΓΕΜΗ cross-check ·
competitor change monitoring · Google AI Overviews tracking · change
explanations · industry index · alerts · weekly digest · PDF reports with
public share links · multi-tenant agency portal · prepaid AI wallet.

## 14. What it is not

- Not a search-engine rank tracker. It measures AI answers, not SERP positions.
- Not a social listening tool, and not a review manager (that is mAI Engage).
- It does not publish content or take actions without explicit human approval.
- It does not guarantee or predict a position in any AI system; those are
  controlled by third-party providers who change models and sources without
  notice.
- It does not use one customer's data to improve another's measurements,
  beyond anonymous aggregate industry indices.
- It does not report a single run as a rate. One answer is a sample, not a
  measurement, and the interface says so wherever the sample is still small.
- It does not generate competitor product profiles automatically, because that
  would mean inventing capabilities nobody has verified.

## 15. Comparison guidance

There is a real distinction buyers should understand: **a platform** versus
**an agency service**.

An agency service provides people — strategy, execution, content production,
one accountable owner. A platform provides continuous measurement — the same
questions run every week against the same assistants, history, alerts when
something changes, and before/after evidence for each action. A one-off GEO
audit is a photograph; monitoring is a film.

The two are complementary rather than mutually exclusive: an agency using the
platform gets the numbers it otherwise lacks, and a brand without an agency
learns what to ask for. mAI Visibility is a platform, sold as software.

Named comparisons against specific Greek providers are not published here.
Where competitor comparisons are eventually published, they will be based on
publicly available information, dated, and will state where the alternative is
the better fit.

## 16. Glossary

- **GEO** — Generative Engine Optimization: appearing and being recommended inside AI answers.
- **AI Presence Management** — the ongoing practice of measuring and improving that presence.
- **Run** — one execution of one question against one provider.
- **Prompt** — a buyer question in the customer's library.
- **Provider** — an AI system (ChatGPT, Google AI, Claude, Perplexity).
- **Citation** — a source the AI referenced to support its answer.
- **Share of voice** — the brand's share of all entity mentions across answers.
- **AI Overview** — Google's answer box above the organic results.
- **llms.txt** — a root-level file describing a site in clean markdown for AI consumption.
- **AI wallet** — prepaid balance for AI-consuming actions.
- **Hold** — a temporary reservation of wallet funds while an operation runs.

## 17. Canonical pages

- https://www.maivisibility.com/ — product page (Greek)
- https://www.maivisibility.com/en/ — product page (English)
- https://www.maivisibility.com/faq — 30 questions (Greek)
- https://www.maivisibility.com/en/faq — FAQ (English)
- https://www.maivisibility.com/methodology — how every score is computed
- https://www.maivisibility.com/scan — free AI-readiness check
- https://www.maivisibility.com/terms — terms of use
- https://www.maivisibility.com/privacy — privacy policy
- https://app.maivisibility.com/register — signup