ROCKMILK
ROCKMILK GEO Audit

See when your brand shows up in the questions that matter.

The audit runs real questions on AI engines, keeps every answer as evidence and shows only what it observed: rates with numerator and denominator, and a 0 to 100 score with the formula in plain sight.

What do you want to run?
Competitors (optional, up to 4)A domain helps tell homonym companies apart and recognise their citations.
About the audit

About the ROCKMILK GEO Audit

What the ROCKMILK GEO Audit is

A tool to look at how a brand shows up in answers written by AI assistants and search engines. GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) are two names for the work of shaping that presence.

It has two modes and they never share a number. The pre-diagnosis measures nothing. The observed audit measures only what it actually observed.

Two modes, kept apart

  • Pre-diagnosis

    No question is sent to any AI engine. You get what you typed, facts read from your own site, hypotheses marked as hypotheses, the questions and sources worth investigating, and a verification plan. There is no visibility score, no ranking and no “invisible” status: missing information is not missing presence.

  • Observed audit

    Runs the questions you approved through the official APIs of the providers enabled in the deployment, stores every answer, and computes the metrics from those stored answers. It is available only where at least one provider is configured; if none is, the tool says so and nothing is simulated.

What the observed audit measures

Every number is shown with its numerator and denominator and can be rebuilt from the stored observations (a CSV export is included). Only questions that do not name your brand count toward discovery; questions that name it are reported apart.

  • Mention rate: valid answers that mention the brand, divided by valid answers.
  • Recommendation rate: valid answers that recommend the brand, divided by valid answers where that could be decided.
  • Own-domain citation rate: answers with a citation attached by the provider to your domain, divided by answers where the provider had search or grounding on.
  • Share of mentions: your presences divided by the presences of all brands in the set you chose, each brand counted at most once per answer. It depends entirely on that set. It is ROCKMILK's operational definition, not an industry standard.
  • Coverage and failures: timeouts, errors, blocks, refusals and empty answers are reported on their own and are never counted as your brand being absent.
  • ROCKMILK Visibility Score: one number from 0 to 100, computed from the four rates above as 35% mention + 35% recommendation + 20% own-domain citation + 10% share of mentions. It is ROCKMILK's operational definition, not an industry standard; the weights and the 60 and 30 band cuts are a project choice, not calibrated against data. It appears only when all four rates can be computed (a missing one is never counted as zero), always beside the sample size and the four rates, and never in the pre-diagnosis.

Repetitions of the same question on the same model are not independent samples, so no confidence intervals are shown.

Mention, recommendation and citation are different things

  • Mention

    The brand appears in the text of the answer. Short names that could mean another company are flagged for review unless the answer also carries your domain.

  • Recommendation

    The answer suggests the brand as an option for the need. A model helps judge this, but only by quoting a passage that appears verbatim in the stored answer; otherwise the case stays undecided.

  • Citation

    The provider itself attached a reference to a page. A link the model wrote in its own text is kept separately and is not treated as proof that it retrieved the page.

  • Position

    Reported only when the answer contains an ordered list. Otherwise it is not applicable.

What it does not do

  • It does not reproduce the consumer apps. It queries APIs, and an API answer is not an exact copy of what an app shows.
  • A call to the Gemini API with grounding is not a test of Google AI Overviews or AI Mode, and is never presented as one.
  • It does not claim to know how any engine ranks or chooses sources, and it does not promise that any engine will recommend a brand.
  • It does not predict anything. A result is one dated collection with a stated configuration.
  • It does not attribute a later change in a metric to an action just because the change came after it.

The technical read of your site

Shown separately from visibility. The audit's server reads your home page, robots.txt and a sitemap and reports status, the rule robots.txt states for the site root for each crawler token, canonical, title and headings, structured-data types and the amount of text available without JavaScript.

Being able to open a page from the audit server does not prove that an answer engine can crawl it, and a sitemap entry is not proof of indexation.

Evidence, privacy and retention

For each observation the tool stores the question and its set version, provider and environment, the model requested and the one the provider reports, the search configuration, the market and language requested (not a guaranteed location), the date, the repetition, the original answer, the citations the provider attached, and the classification with its version.

An audit is private to the browser that created it: it is reachable only with a secret token that is stored only in that browser. It can be deleted at any time from the report and is erased automatically after 90 days.

After the audit

A one-off audit gives you a dated baseline you can repeat. Monitoring means repeating the same collection on a schedule and comparing it against the baseline; it is not part of this tool. Executing the improvements is a separate engagement, scoped after reading the evidence. Deliverables, limits and exclusions are agreed per engagement, and this page publishes no prices.

GEO Audit: questions and answers

Is it free?

The pre-diagnosis and the report are free to read. Leaving contact details is optional and separate from the report.

Does it test my brand on ChatGPT, Perplexity, Gemini and Claude?

Only the observed audit queries engines, only through the official API of each provider enabled in the deployment, and the report names exactly which ones and which models. The pre-diagnosis queries none.

Is there a single score?

The observed audit shows a 0 to 100 ROCKMILK Visibility Score, but only as a summary of the four rates, which are always shown with their numerators and denominators. The formula and weights are public on this page, they are ROCKMILK's choice and were not calibrated against data, and the score is withheld when any of the four rates cannot be computed. The pre-diagnosis has none.

What if a provider fails or times out?

The failure is recorded and lowers coverage. It is never counted as your brand being absent, and a partial result is labeled as partial.

Will the result be the same if I run it again?

Not necessarily. Answers vary between runs, which is why the questions are fixed and versioned, and why a comparison is only shown when the question set, providers, models and brand set are the same.

Can ROCKMILK guarantee that AI will recommend my brand?

No. Nobody can. The report points to what to verify and where the observed gaps are; the outcome is not promised.

Want to talk it through?

Run the audit above. At the end of the report you can optionally leave your contact details, or talk to ROCKMILK directly.

Start a Fresh Problem Diagnostic→

Or write to hello@rockmilk.com.br