AEO metrics show how a brand appears in answers to a defined set of prompts across specified answer engines. They measure observed visibility, not whether that visibility caused a visit, lead, or sale. If a completed answer to a tracked comparison prompt names your brand and cites your site, that is evidence of exposure in the observation. It is not proof that a prospect saw the answer or later converted.
Use AEO measurement alongside SEO reporting, not in place of it. Google Search Console reports Google Search activity such as impressions, clicks, queries, and pages. AEO tracking records generated answers, brand mentions, and citations. Keep those exposure observations separate from downstream search and business outcomes.
This article presents a practical measurement contract, an illustrative observation ledger, and a readiness-focused approach to tools and attribution. Where a workflow is proposed rather than vendor-documented, it is labeled as such.
What AEO metrics answer, and what they cannot prove
An answer observation is one captured response to one versioned prompt, from one answer engine, at one execution time. Record the model or version when the source exposes it. A new engine run is a new observation, even if the prompt has not changed.
SEO reporting describes how people find and use a site through Google Search. AEO observation describes what an answer engine returned for a selected prompt, including whether it mentioned a brand or cited a source. Google’s guide to Search Console and Google Analytics explains the distinct roles of those products. Neither turns an external AI answer into a confirmed visit or referral.
An answer mention is evidence of observed visibility. On its own, it is not evidence of a visit or a sale.
Define the metrics before building a dashboard
AEO metrics are not standardized across vendors. Before comparing results, document the numerator, denominator, prompt population, engine scope, date window, and formula version. These definitions are a practical measurement contract, not an industry-wide standard.
| Metric | Definition | Keep distinct from |
|---|---|---|
| Brand inclusion rate | Completed answer observations with at least one detected brand mention, divided by all valid completed observations in the stated scope. | Mention frequency, where repeated mentions within one answer can increase the count without changing inclusion. |
| Owned-domain citation rate | Completed answer observations with at least one citation to an owned domain, divided by valid completed observations. | Brand inclusion, because a cited page may not name the brand. |
| Citation share | Citations to owned domains divided by all captured citations in the stated scope. | Citation rate, because this denominator is citations rather than answers. |
| Share of Voice | A formula-defined share of brand mentions or citations across a stated population. | Other Share of Voice formulas. Mention-based and citation-based results are different measures. |
For example, mention-based Share of Voice could be calculated as a brand’s mentions divided by all brand mentions across tracked answers. Citation-based Share of Voice instead uses citations in its numerator and denominator. HubSpot documents a mention-based Share of Voice, so label the formula before comparing it with a citation-based calculation. Its documentation also distinguishes citations from brand mentions: a citation can point to a page that does not mention the brand. See HubSpot’s brand visibility and Share of Voice definitions and its explanation of citations versus brand mentions.
Prominence and sentiment are useful coded dimensions, not substitutes for inclusion or citation measures. Record where a brand appears and its context, such as primary recommendation or supporting option. If reporting HubSpot sentiment, use the displayed scale accurately: HubSpot describes a scale from -100 to +100. Do not turn a vendor score into a universal benchmark.
Capture prompt observations at the right grain
Keep prompt definitions, engine runs, answer observations, citations, and aggregates at separate data grains. A prompt definition records the exact prompt and its version. An engine run records one execution for that prompt, engine, and available model version. An answer observation records the response status and captured answer for that run. Each citation is a child record linked to its answer observation. A reporting aggregate is calculated separately across a named time window and population.
The following is an illustrative internal data contract, not a HubSpot schema and not a claim that every field is exposed in its interface or API. Store permitted answer text or a reference to it. If retaining text is not appropriate, retain a hash and the source record identifier where available.
{
"answer_observation_id": "internal UUID",
"prompt_definition_id": "p_104_v2",
"engine": "ChatGPT",
"model_version": "if exposed by source",
"run_started_at": "2026-10-12T09:15:00Z",
"response_status": "completed",
"source_tool": "illustrative monitoring source",
"vendor_run_id": "if exposed by source",
"raw_answer_reference": "permitted reference or null",
"answer_text_hash": "optional SHA-256",
"extraction_version": "1.0"
}Put citation details in separate child rows, such as answer_observation_id, citation_ordinal, original URL, normalized URL, cited domain, and whether the answer also mentioned the brand. Keep the original URL for provenance when normalizing it for analysis. If multiple mention positions or contexts matter, store individual mentions separately rather than reducing them to one answer-level Boolean.
Do not use prompt text plus calendar date as a unique key. Multiple engines, model versions, daily runs, and retries can share that date. Prefer a vendor run ID where available. Otherwise compose an internal key from prompt version, engine, model version when available, and run timestamp. A read-then-insert check can race when workers run concurrently, so enforce uniqueness in the database or use an atomic upsert.
Keep completed_no_mention, no_response, extraction_failed, not_run, and permission_error distinguishable. Exclude incomplete states from the inclusion denominator and report completeness separately. For every aggregate, store the formula and numerator and denominator, not just a percentage.
Choose a collection method that matches what is documented
HubSpot AEO is documented as a beta monitoring option for tracked prompts across ChatGPT, Gemini, and Perplexity. HubSpot describes daily prompt checks and reporting for visibility, prompt performance, citations, competitors, recommendations, and sentiment-related measures. It recommends reviewing results over multiple days or weeks because generated answers change. Product limits and features vary by edition.
- Standalone HubSpot AEO and Marketing Hub Professional: documented limits are 25 daily prompts and 2,500 monthly answers.
- Marketing Hub Enterprise: documented limits are 50 daily prompts and 5,000 monthly answers.
These limits are date-sensitive and may change. The standalone product page lists $50 per month, or $45 per month on annual billing, as of the research date of October 12, 2026. Confirm current pricing, packaging, and checkout terms before budgeting. Consult HubSpot’s AEO setup and analysis documentation for current edition details. Teams reviewing their broader operating setup can also consider HubSpot systems consulting; that link is not a claim of an undocumented AEO integration.
HubSpot’s free AI Search Grader is a one-time diagnostic, not a continuous prompt-monitoring dashboard. HubSpot also documents a connector through which approved ChatGPT or Claude users can retrieve AEO performance, tracked prompt responses, citations, and competitor Share of Voice. A Super Admin must approve the connector and choose permissions. Treat it as permission-controlled assistant access, not a generic export, CRM sync, or unrestricted API.
HubSpot announced an AEO Public API in public beta, but the announcement alone does not establish endpoint schemas, scopes, quotas, history, pagination, or production readiness. Before planning automated ingestion, verify a current API reference for the exact resources and controls required. If those details are not available, use the documented dashboard or connector path, or describe automation as a proposed design rather than an available integration.
Use AI classification only where rules are insufficient
Use deterministic rules for exact brand and domain matching, URL normalization, engine matching, prompt membership, timestamps, deduplication, and response-state assignment. These tasks have explicit inputs and allowed values, so a parser or validation rule is easier to test and audit than a language model.
AI classification can help with context-dependent labels such as sentiment, framing, recommendation strength, or an ambiguous paraphrase. Keep the source answer or permitted reference with the label, along with classifier version, confidence, and review status. Validate model output against allowed values before saving it. Require a person to review low-confidence or legally sensitive negative claims and any classification that would change customer-facing copy or a CRM record.
Unlinked paraphrase can be recorded as an observation, but it does not prove that content entered a training set, knowledge graph, or persistent model signal. Keep that distinction in any analysis of content reuse.
Connect visibility to search and business outcomes carefully
Search Console can show Google Search impressions, clicks, and queries. It does not identify prior exposure in ChatGPT, Gemini, or Perplexity. Use it as a downstream signal, alongside analytics and CRM evidence, not as an AI-referral detector. Google notes that Search Console data typically has a delay of roughly two to three days; query anonymization and differences between Search Console and Analytics counting systems also affect comparisons. The Search Analytics API guide describes query access and its limits.
A practical analysis starts with agreed definitions. List branded query terms and spelling variants, define high-intent queries such as pricing or demo searches, select a sustained AEO visibility period, then compare Search Console and onsite conversion measures across equivalent periods. Join CRM opportunity records only where the business has a documented source or exposure signal. Keep query classification, dates, and event joins reproducible. An AI-generated summary may describe patterns, but it should not make the attribution decision.
A rise in branded searches after tracked visibility improves can justify investigation, but it does not identify which answer engine influenced a searcher. Preserve a dated prospect statement or verified AI-source record before assigning exposure to a lead or opportunity.
Use evidence labels that match what the data supports:
- Direct exposure evidence: a tracked answer mentions or cites the brand, or a prospect documents seeing the brand in an AI answer. This establishes an observation or reported exposure, not automatically a conversion effect.
- Downstream correlation: branded searches, organic clicks, direct sessions, or conversions move during a period of sustained AEO visibility. Report the timing and comparison, not a causal claim.
- Stronger attribution: a lead or opportunity has a verified AI source or documented prospect response, or an appropriate controlled design supports incremental impact.
For teams responsible for source fields and opportunity records, CRM systems consulting is relevant to data ownership and evidence quality. Do not label unattributed direct traffic as AI-sourced merely because it rose during the same period.
Set a review cadence and a decision rule
Monthly review and quarterly content refreshes are reasonable starting hypotheses, not universal benchmarks. Adjust frequency for category volatility, prompt coverage, answer completion, and meaningful changes in visibility, citations, sentiment, or engine behavior. Report observation counts and completion rates with the metric. Where sample volume supports it, show a distribution such as median, mean, and spread rather than selecting the strongest answer.
Changes to the prompt set or engine mix create a new measurement population. Version the change instead of silently comparing it with an earlier period. A practitioner case study tracked one prompt over a short period and reported Share of Voice ranging from 25% to 63.16%, with a reported mean of 43.25% and median of 40.84%. That single-prompt, uncontrolled observation illustrates volatility. It is not a benchmark or proof that a content update caused the change. See the case study and its reported observations.
- Confirm the prompt-set version and reporting window are stable and named.
- Show completed observation count and answer completion rate; do not treat failures as zero visibility.
- State whether Share of Voice counts mentions or citations, with its numerator and denominator.
- Check that engine scope and model details, where available, are comparable across periods.
- Inspect the underlying answers and citations before deciding what content to change.
When a sustained change passes those checks, investigate the answers and cited sources first. Then choose a specific content action, document its date and rationale, and continue observing the same prompt population before interpreting the result.