How to Measure AI Search Visibility: A Reliable AEO Framework

Measure AI search visibility with a fixed, versioned prompt set. Repeat observations by engine, model or mode, geography, and date. Retain evidence of mentions and citations, then calculate results over a stated period with an explicit denominator.

For example, a clinic software company could run the same appointment-reminder prompts in ChatGPT and Gemini, record each answer and its cited URLs, and compare the results with a later period. The comparison is meaningful only if the prompt set, engines, audience, geography, and calculation rules remain comparable.

Keep Google Search features separate from conversational answers. Google Search Console reports on Google Search AI features, including AI Overviews and AI Mode. A separate prompt-monitoring method or product is needed to observe selected conversational engines such as ChatGPT, Gemini, or Perplexity.

What AI search visibility measures

AI search visibility is a collection of observable appearances in AI-generated responses, not one universal ranking. Set the measurement boundary before collecting data: name the surface, audience, geography, prompt set, engines, model or mode variants, and observation period.

Answer-engine visibility and Google AI Search visibility should therefore be reported as separate measurements. A Search Console result describes Google Search features. A prompt run describes what a selected answer engine returned under particular conditions.

Record these outcomes separately:

• Brand mention: the answer names the brand, whether or not it links to a page.
• Owned-page citation: the answer cites a page on the brand’s own site or another owned asset.
• Third-party citation: the answer cites an external page that discusses the brand or supports a claim.
• Recommendation: the answer explicitly suggests or ranks the brand as an option. A mention alone is not a recommendation.

A citation is evidence that a URL appeared in an answer. It does not by itself show that the answer endorses the brand, represents the page accurately, or influenced a purchase.

Choose signals that answer a business question

Choose metrics to support a decision. Before reporting one, document its numerator, denominator, eligible answers, competitor set, engine mix, prompt-set version, and aggregation period. Keep the underlying observations separate from period-level calculations.

Signal	What to record	Decision it can inform
Mentions	Brand identity, context span, and position if observable	Investigate entity recognition or missing coverage when the brand is absent.
Citations	URL, source type, citation position, ownership, and quality status	Review relevant pages and cited sources when owned citations are missing or inaccurate.
Context and sentiment	Evidence span, proposed label, classifier version, confidence, and review status	Correct outdated facts or investigate misleading descriptions and external representations.
Recommendation and share of voice	Recommendation status, formula, competitor set, numerator, and denominator	Compare competitive visibility for a defined prompt set, not all AI answers.

One possible share-of-voice formula is brand mentions divided by all tracked brand and competitor mentions across eligible answers. An answer naming several brands contributes several mentions to that denominator. A different measure, the percentage of eligible answers that contain the brand, counts answers rather than mentions. Label the two measures distinctly because neither is self-defining.

Sentiment is an interpretation of the surrounding text, not a raw fact. Preserve the evidence span, classifier version, confidence, and review status. Ambiguous names, sarcasm, negative reputational findings, and disputed citations should be checked by a person before they appear in executive reporting or a CRM record.

Build a repeatable prompt-monitoring protocol

Choose prompts from customer questions, sales conversations, and real buying decisions. Include category discovery, use-case evaluation, comparisons, and brand-specific questions. Keep exploratory prompts separate from the stable set used for period comparisons.

Maintain a prompt register with the exact text, prompt-set version, language, geography, buyer stage, engine, and model or mode variant when available. Run the same prompts under consistent conditions on a planned cadence. Save the timestamp and answer or an approved excerpt, subject to retention and privacy rules.

Repeated runs are observations, not proof by themselves that a change is statistically significant. If prompts, competitors, engines, geography, or aggregation rules change, record the change and qualify the comparison.

As a small illustrative protocol, a team might collect several runs per prompt during a baseline period, then repeat on a consistent schedule. This is an example of repeatable collection, not a statistically validated sample-size standard. Inspect individual answers as well as aggregates so a changed prompt or unusual response is not hidden by one score.

Model the data at the right grain

Define each record before building a dashboard or pipeline. A prompt definition is the tracked question. A prompt run is one execution of that question on one engine or variant at a specific time. A mention or citation is an item found within one run. An aggregate is a calculated metric over a defined set of runs and a time window.

One answer can contain multiple mentions and citations. Store each as a child record linked to its run. Do not store share of voice as an attribute of one answer, and do not treat a visibility summary as a CRM contact or deal event.

{
  "run_id": "illustrative-unique-run-id",
  "prompt_id": "P-014",
  "prompt_set_version": "2026-10-v1",
  "engine": "example-answer-engine",
  "model_variant": "record-if-exposed",
  "run_at": "2026-10-11T10:00:00Z",
  "language": "en",
  "geography": "US",
  "answer_hash": "illustrative-hash",
  "collection_method": "manual-or-documented-tool"
}

This hypothetical, vendor-neutral record describes one prompt execution. Its field names are proposed, not a published HubSpot schema. A stable run identifier must distinguish repeated executions, engines, model variants, and timestamps.

Store separate mention records with the run ID, normalized brand identity, position if observable, evidence span, and review status. Store separate citation records with the run ID, normalized URL, source type, citation position, ownership, and quality status. Store aggregate records with the period, engine, prompt-set version, brand, metric name, numerator, denominator, and calculation version.

For concurrent writers, enforce uniqueness in the database and use a transactional or atomic upsert. A read-then-insert check alone can race. A proposed run key can combine prompt ID, engine, model variant, normalized run time, and prompt-set version, with a generated run ID for the individual execution. Use similarly scoped keys for citation and mention children, including the run ID and a citation or mention ordinal or fingerprint.

Keep machine classifications and later human corrections with their provenance. If an answer changes or a reviewer corrects a sentiment label, preserve the earlier value and correction history instead of overwriting the evidence.

Turn observations into reviewed signals

Use deterministic rules for checks with clear answers: match citations against verified company domains, normalize URLs, identify known brand variants, resolve redirects, check canonical destinations, and detect duplicates. These tasks do not need an AI interpretation.

For nuanced classification, an AI system may propose whether a passage recommends, compares, criticizes, or merely mentions a brand. Require a proposed label, supporting evidence span, confidence, and classifier version. Validate allowed labels and required fields before saving. Route ambiguous brand matches, sarcasm, negative reputational findings, and questionable citations to a person before using them in reporting or a CRM record.

For citation quality, resolve the URL, check redirects and canonical destination, confirm that the page is relevant, and verify that it supports the claim attributed to it. Keep an irrelevant, unavailable, stale, or redirected citation as an observation, but flag its quality rather than counting it as a useful source without qualification.

Do not infer a lead, influence, or revenue impact from visibility alone. For downstream analysis, seek corroborating evidence such as analytics, campaign data, referral information, or self-reported source data. CRM system design is relevant when defining how those separate outcomes should be recorded and reported.

Use the right source and tool for each surface

Google Search Console: Google Search AI features

Google’s Search Central guidance covers AI Overviews and AI Mode as Google Search features. Google states that existing Search fundamentals remain relevant, pages must be indexed and eligible for a Search snippet, and there are no special AI schema or additional technical requirements for inclusion. Indexing and serving are not guaranteed. Google also documents query fan-out in these Search features; that should not be generalized as a mechanism used by every answer engine. See Google’s guidance on AI features in Search.

Google announced dedicated generative-AI performance reporting in Search Console for Search features, including AI Overviews and AI Mode, with worldwide rollout completed by August 31, 2026. Confirm availability and the fields shown in the verified property before using them in an internal report. This is Google Search reporting, not a measure of ChatGPT, Perplexity, or general chatbot answers. See Google’s Search Console reporting announcement.

HubSpot AEO: selected conversational engines

HubSpot describes AEO as a Beta product monitoring ChatGPT, Gemini, and Perplexity, with prompt setup, visibility and citation analysis, competitor analysis, and recommendations. Its documented setup includes brand name and domain, name variants, competitors, product, ideal customer profile information, and prompts. HubSpot’s guide says initial brand configuration can only be completed once per portal, so review those inputs carefully before setup.

HubSpot’s product catalog currently lists standalone AEO at $50 per month, with 25 tracked prompts, each run daily across three engines, and 2,500 answers per month. Additional limit packs are listed at $20 per month for 10 more prompts and 1,000 additional answers per month. These terms, limits, and prices can change, so verify the current catalog before procurement. See HubSpot AEO, its setup guide, and the product catalog.

Reviewed public HubSpot documentation does not establish a public AEO API, webhook, export schema, or direct CRM sync for observation-level data. Do not plan a data connection until the required access is confirmed. HubSpot’s AI Search Grader is a free, one-time diagnostic, not a recurring prompt log or historical trend series.

When evaluating a monitoring method, check its surface coverage, control over prompts and geography, access to answer and citation evidence, historical retention, documented export or API access, and review workflow. Map each desired metric to its actual source before procurement: Search Console for Google Search AI reporting, a separate method or product for selected conversational engines, and analytics or CRM evidence for downstream outcomes.

Convert measurement into a practical improvement loop

Start with a reviewed, observable gap: an incorrect brand description, an outdated product fact, a missing relevant citation, or absent visibility for a defined buyer-intent cluster. Inspect the answer and its cited sources before changing content. Where the evidence supports it, update factual details, clarify the page’s purpose, improve crawlability and internal linking, or review third-party representations.

For Google Search AI features, follow Google’s documented fundamentals: make pages crawlable and eligible for a Search snippet, use helpful and accurate content, and ensure structured data matches visible text. Google says special AI schema is not required. Do not infer that a specific format, update interval, word count, or content length guarantees citations in other answer engines.

Assign one owner to the prompt register and one to the content, technical, or reputation action. Record the action date, then rerun the same prompt-set version under comparable conditions. Report observations separately from business outcomes. A change after remediation is useful evidence for the next decision, not proof that the change caused it.

For help assessing HubSpot access, data ownership, and a proposed reporting workflow, see HubSpot implementation support.