AI search behavior creates three different measurement questions: did an answer mention your brand, did it cite your site, and did someone later visit or convert? Treat these as separate outcomes. A response to “Which CRM suits a small B2B sales team?” might name your company without linking to it, cite a product page, or lead to a later visit. None of those observations proves the next one.
This distinction changes what your team records and what it can conclude. An answer is an observation for a specific prompt, engine, context, and time. It is not a permanent visibility score for a brand or page. A useful measurement program therefore combines documented product reports, controlled observations, and downstream analytics without forcing them into one unexplained percentage.
The practical goal is not to claim that a page “won AI.” It is to identify persistent gaps in important answers, preserve enough evidence to investigate them, and make content decisions that can be reviewed later.
A mention measures answer presence. A citation measures source selection. Traffic and conversion measure downstream behavior. Keep all three separate.
What AI search visibility measures and what it does not
AI search behavior is conversational discovery through answer engines and generative search features. A person may receive a synthesized answer before deciding whether to visit a website. A brand mention means the answer names the brand. A site citation means the answer identifies a page or domain as a source. A visit or conversion is downstream activity recorded in analytics or another business system.
These signals answer different business questions. Mentions show whether the brand appeared in an answer. Citations show whether its material was selected as a source. Analytics can show attributable sessions and conversions under the configured attribution model, but it cannot reveal every unclicked mention or citation. A citation does not prove that a person clicked, and a later conversion does not by itself prove that an AI answer caused it.
Google explains that its generative Search features use its Search systems, while other answer engines have different products, retrieval processes, and reporting. Google’s guidance on generative AI features recommends sound Search fundamentals rather than a special set of AI-only ranking tactics.
Choose the measurement stream for the question
No single dashboard represents every answer engine and every kind of visibility. Keep each evidence stream identifiable by source, period, and data grain.
| Evidence stream | What it records | Useful decision | Boundary |
|---|---|---|---|
| Google Search Console | Google generative-feature impressions and available page, country, device, and date dimensions | Review visibility of site pages in Google generative Search features | Not a measure of ChatGPT or Perplexity mentions |
| HubSpot AEO | Tracked prompts, brand and competitor visibility, citations, and HubSpot’s share-of-voice metric | Review tracked answers across supported engines | Product-defined reporting and limits, not a universal metric |
| Analytics | Attributable sessions and conversions under the analytics configuration | Assess measurable downstream activity | Does not reveal answers that received no click |
| Controlled observations | Recorded runs of exact prompts and their mentions or citations | Investigate changes for a stable prompt set | Results vary by engine, variant, context, and time |
Google’s Search Console generative-AI reporting announcement describes Google-specific performance reporting, including impressions and page-level dimensions. Keep those records separate from prompt observations for other engines.
HubSpot documents AEO tracking across ChatGPT, Gemini, and Perplexity, including prompts, competitors, citations, and its product-defined share of voice. HubSpot defines share of voice as the proportion of brand mentions belonging to the tracked brand across tracked prompts. Do not assume another tool uses the same definition.
HubSpot lists 2,500 monthly answers and 25 daily prompts for HubSpot AEO, its free trial, and Marketing Hub Professional. Marketing Hub Enterprise lists 5,000 monthly answers and 50 daily prompts. Tracked prompts run daily, and HubSpot recommends reviewing results across multiple days or weeks. Check the current HubSpot AEO setup and reporting documentation when planning coverage. The documentation describes product reporting, but the reviewed material does not establish a public AEO API or a general cross-engine export specification.
Teams assessing how their HubSpot environment fits reporting operations can review HubSpot systems consulting. That link is a service resource, not evidence of a particular AEO integration.
Build a prompt set that represents real discovery
Choose prompts from real buyer and customer questions rather than only high-volume keywords. Collect examples from sales conversations, service tickets, site search, and customer research. Group them by intent, such as discovery, comparison, implementation, and risk or support.
A prompt set should represent the decisions your audience is making. Include the initial category question and likely follow-ups about tradeoffs, pricing, implementation, security, support, and fit. A rewritten prompt is a new version, not an interchangeable substitute for the original.
Maintain a prompt register with prompt_id, exact_prompt_text, intent_group, prompt_set_version, owner, and active_status. In HubSpot AEO, setup includes brand and domain details, brand-name variations, competitors, products, ideal customer profiles, and prompts that can be reviewed or edited. HubSpot notes that brand-name variations are case-sensitive. Have sales and service review the prompt set before recurring monitoring begins.
Record observations at the right grain
For a proposed cross-engine observation model, one observation row means one run of one exact prompt against one engine or product variant, in a recorded context, at one time. A later answer is a new observation, not an overwrite. Store each citation separately from the run, and calculate period summaries separately from both.
The following is an illustrative internal data design. It is not a HubSpot or Google export format, vendor template, or API specification. Preserve a raw response reference before calculating metrics.
{
"observation_id": "obs_20261014_001",
"prompt_id": "crm_compare_001",
"prompt_set_version": "v1",
"exact_prompt_text_hash": "sha256:example",
"engine_name": "ChatGPT",
"model_or_product_variant": "visible variant if available",
"run_timestamp_utc": "2026-10-14T12:00:00Z",
"locale": "en-US",
"context_state": "anonymous temporary chat",
"personalization_state": "not observed",
"response_capture_reference": "capture_ref_001",
"source_record_hash": "sha256:example-response",
"experiment_id": "opening_answer_test_01",
"brand_mention_status": "present; reviewer pending",
"reviewer_status": "pending"
}observation_id uniquely identifies this prompt-run row. In a concurrent system, enforce uniqueness with a database constraint or transactional upsert based on the full run identity, such as prompt version, exact prompt, engine, model variant, timestamp, locale, context, and experiment. A lookup followed by an insert is not sufficient protection against races.
Each citation belongs in a separate record with a citation_id, observation_id, cited_url, normalized cited_domain, and citation_position. The citation key must distinguish multiple citations in one response, for example by combining the observation identifier with a citation ordinal or normalized URL and position. Do not collapse four observed URLs into one row merely because they came from one answer.
Keep mention classification distinct from citations. If the question is whether the brand appeared, use one response-level flag. If you count individual occurrences, create a mention record for each occurrence and retain the counting rule. A daily or monthly summary is a different row grain again. It should include the brand, engine, model or product variant when available, prompt-set version, period, period start, and counting rule.
Use deterministic rules for URL normalization, approved brand-domain matching, date windows, and duplicate detection. An AI classifier can suggest a classification for an indirect or ambiguous brand reference, but a person should review low-confidence matches before they enter an operational record. Keep raw evidence separate from derived data, restrict access, and set retention rules for response captures that could contain personal or confidential information.
Do not attach a monthly share-of-voice value to an individual prompt run. A run is raw evidence; share of voice is a defined aggregate over a prompt set, engine scope, counting rule, and reporting period.
Turn raw observations into defensible metrics
For a response-level mention rate, divide eligible prompt responses containing an approved brand mention by all eligible responses in the stated scope. Report the engine, prompt-set version, date range, and whether a repeated mention counts once per response or once per occurrence.
Calculate citation rate separately. Choose and state the denominator, such as eligible responses or defined citation opportunities. Specify whether multiple citations from the same domain count once or more than once. Rates with different denominators or counting rules are not directly comparable.
HubSpot’s share-of-voice value is its product-defined proportion of brand mentions across tracked prompts. It is an aggregate, not a value for an individual prompt run. For every aggregate, retain the brand, engine, model or product variant when available, prompt-set version, counting rule, aggregation period, and period start.
A changed prompt mix, engine mix, competitor list, or model can move a metric even when no page changed. If prompts are added, removed, or rewritten, create a new prompt-set version and report the change rather than presenting the movement as a content improvement.
Use controlled reviews to decide what to change
To assess a content edit, preserve the baseline page version and observations, make one substantial editorial change, then rerun the same prompts under comparable conditions. Keep prompt wording, engine, locale, login state, and context as consistent as practical. Review the evidence before retaining the change, and record a competing explanation such as an engine update or changed citation availability.
Improve content for people first: give direct, accurate answers, use clear headings, make important information available as crawlable text, and provide useful internal links. Google says AI Overviews and AI Mode do not require special AI-only schema or mandatory content chunking. Structured data should describe visible page content and follow Google’s policies; it does not guarantee a Search feature. See Google’s guidance on AI features and structured-data policies.
A March 2026 arXiv preprint reported citation and extraction results for content-structure experiments across six generative engines. Its reported 17.3% citation improvement and 43% list and table extraction result belong to that study’s benchmark. They are experimental evidence, not a production guarantee, a Google requirement, or a reason to force a fixed number of headings or a particular page format.
Apply data-quality and ownership controls
Assign prompt-set and monitoring ownership to a named marketing or analytics owner. Page edits belong to the content owner. Keep the evidence and classification rule available so a result can be checked or recalculated if the metric definition changes.
- Confirm the exact prompt, prompt-set version, engine, timestamp, locale, context, and raw-evidence reference.
- Keep response-level mention status and individual citation records separate.
- Normalize URLs deterministically and check domains against an approved list.
- Send ambiguous entity matches, inaccessible citations, malformed URLs, and unexpected domains for human review.
- Retain the derivation rule, denominator, counting rule, and source record for each metric.
- Apply access and retention controls before storing full response captures.
Do not write an uncertain classification into a CRM simply because it appears in a monitoring result. Block write-back when provenance is missing, confidence is below the declared threshold, a cited page is inaccessible, an approved classification would be overwritten, or the response contains personal or confidential information. Teams clarifying CRM record ownership and operational controls can review CRM systems consulting.
Make measurement useful to the business
Review mentions and citations to see whether the brand appears in priority answers, whether its own material is used, and which competitors or third-party sources are present. Review referral sessions and conversions separately. A lack of clicks does not prove an answer had no influence, while a measured visit does not prove that a citation caused it.
Use a monthly review for a stable prompt set and a quarterly content review. Add a check when a material engine change is announced. For Google-specific generative-feature visibility, use Search Console. For tracked prompts across supported engines, use the product reporting available to your team or a clearly labeled observation process.
A practical action rule is to investigate a persistent gap in a high-priority prompt group, identify which page or third-party source appears in the answer, and assign an editorial action only after checking the evidence and its business relevance. Compare citation and mention outcomes with organic visibility, usability, and factual accuracy before deciding that an edit succeeded.
Frequently asked questions about AI search measurement
Does Google require special schema for AI Overviews or AI Mode?
No. Google says special AI-only schema is not required. Use structured data where appropriate, ensure it matches visible content, and follow Google’s policies.
Can Search Console measure ChatGPT or Perplexity visibility?
No. Its generative-AI reporting concerns Google Search features. Keep it separate from observations or product reporting for other engines.
How is a citation different from a brand mention?
A mention names the brand. A citation identifies a source, such as a page or domain. Record them as separate outcomes.
Does one prompt run establish a trend?
No. Compare multiple observations with prompt wording, engine, model or product variant, context, and timing as consistent as practical.
Can HubSpot AEO observations be exported through a public API?
The reviewed HubSpot documentation describes product tracking and reporting but does not establish a public AEO API or general cross-engine export specification. Confirm current capabilities with HubSpot before planning an integration.