Measure AI search performance across three evidence layers: observed visibility and citations, identifiable referrals or self-reported discovery, and business outcomes under a stated attribution rule. If a tracked answer names your company but sends no visit, record the visibility only. If a later session arrives from a recognizable AI referral, measure that session separately.
This framework helps teams define AI search performance KPIs, select the right denominator, preserve evidence at the correct data grain, and connect discovery signals to analytics and CRM reporting. It is a measurement design, not a claim that an AEO platform automatically feeds GA4, a warehouse, or a CRM.
Traffic can miss zero-click discovery, while visibility does not prove that anyone saw, trusted, or acted on an answer. Report the evidence layers together, but do not collapse them into one unqualified AI-impact number.
How to choose AI search performance KPIs
Choose each KPI to answer one business question. Visibility asks whether a brand appeared in a defined set of AI answers. Citation measures ask which sources were displayed. Referral data asks whether a measurable visit arrived from a recognizable source. Self-reported discovery records what a person says influenced them. Conversion, pipeline, and revenue describe outcomes only when their event definitions and attribution rules are explicit.
Before a KPI enters a report, document its question, unit of analysis, denominator, reporting period, source system, owner, and intended decision. A visibility rate might guide content review; a verified AI-referral conversion rate might guide channel investment; a self-reported discovery measure might inform sales and research. They are not interchangeable.
Define the measurement contract before tracking
Decide what counts as a brand appearance before calculating visibility. An explicit mention, recommendation, link, citation, and position in an answer are different observations. Record them separately. A mention without a citation is not a citation.
Version the prompt set and retain the exact prompt text, engine, model or variant when exposed, timestamp, geography, and language. AI answers vary by run and context, so these fields define what a comparison covers. The same prompt tested twice creates two prompt-run records.
Keep the data grains separate. A prompt is a reusable question. A prompt run is one test of that question. A brand observation records whether a particular brand appeared in that run. A citation record represents one displayed source item. A period aggregate summarizes eligible records over a defined scope.
A visibility metric is comparable only when its appearance rule, prompt-set version, engine scope, and denominator stay fixed or the change is explicitly versioned.
If prompts are added, removed, reworded, or reclassified, create a new prompt-set version or publish a controlled comparison. Do not present a denominator change as an improvement in visibility.
Measure visibility, citations, and answer quality
Prompt-run visibility rate
Use the prompt run as the denominator:
Prompt-run visibility rate = (eligible prompt runs with an explicit brand appearance / all eligible prompt runs) x 100.
If a brand appears in 18 of 40 eligible runs, its visibility rate for that prompt-set version and engine scope is 45%. Report mention, recommendation, link, and citation status as separate fields or rates. One answer is not a representative score for an entire prompt set.
Citation share
For a defined set of tracked brands, calculate citation share as:
Brand citation share = the brand’s citation instances / citation instances attributed to all tracked brands in the same eligible scope x 100.
State what counts as an instance. One practical rule counts each displayed citation item in each run. The same URL cited in separate runs counts again. Unique cited-domain share is a different metric and must be labeled separately. If one answer displays two citations from the same domain, the instance rule determines whether both count; do not change that rule between periods.
Keep citation existence separate from citation validity. A cited page may be inaccessible, irrelevant, outdated, or unable to support the answer. Accuracy and sentiment are reviewed fields, not proof that a citation is sound. Incorrect pricing, product features, legal statements, or security claims should go to a human reviewer.
HubSpot publicly describes its AEO product as beta and lists ChatGPT, Gemini, and Perplexity coverage, with prompt tracking, visibility, citation analysis, competitor visibility, sentiment, recommendations, and daily refresh language. The public product pages do not establish a general export schema or API for writing observations to another system. Recheck current plan and product details before relying on a capability. See HubSpot AEO product details and the AEO setup walkthrough.
Answer quality
Use a review rubric for product identity, pricing, feature availability, use-case fit, sentiment, and citation validity. Store the evaluator, evaluation status, and rubric version. A score should identify an operational issue to investigate, not create a false impression of statistical certainty.
Separate referrals from indirect discovery signals
Use GA4 source, medium, referrer, and campaign values to identify sessions with a recognizable AI referral. For known sources, deterministic rules are preferable to asking an AI model to classify traffic. GA4 custom channel groups use configured rules, and the first matching channel is applied. Google notes that custom channel groups are not included in the BigQuery Export schema, so preserve raw acquisition values if you use a warehouse.
Google defines direct/(none) as traffic without a clear referral source. Missing campaign data, removed referral information, privacy tools, redirects, and direct URL entry can all contribute. A direct session does not identify the original discovery path. Do not relabel direct sessions as AI traffic.
Use consistent UTM naming where campaign parameters are applied, and do not put personal information in Analytics parameters. See Google’s custom channel group guidance, Google’s explanation of direct traffic, and its UTM guidance.
Search Console can provide query, URL, country, device, and available search-appearance filters. Verify the actual property and current account options rather than assuming a universal AI Mode filter. Branded-query movement and direct-session movement can provide context, but advertising, PR, social activity, seasonality, and existing awareness can also move them. Treat these measures as directional indicators.
| Signal | Unit | What it supports | What it cannot prove |
|---|---|---|---|
| Prompt visibility | Eligible prompt run | Observed appearance in a defined test set | Audience exposure, visits, or revenue |
| AI referral | Session | Visit with a recognizable referral or campaign value | Untracked discovery before the visit |
| Self-reported discovery | Response or discovery event | What a respondent says they used or recall | Verified first touch unless asked directly |
| Branded or direct trend | Query or session over time | Directional movement for investigation | That AI caused the change |
Connect measurement to conversions and pipeline carefully
For a session-based AI referral conversion rate, define the eligible session first:
AI-referral conversion rate = sessions with at least one specified conversion event / eligible AI-referred sessions x 100.
State the key event, reporting window, session definition, and referral-classification rule. Counting converting sessions rather than raw event totals avoids a rate above 100% when one session triggers an event more than once.
Capture survey answers in a separate CRM property or discovery-event record. Retain the original response, question wording or version, and response time. Ask whether the person first heard about the company through an AI tool only if that is the intended measure. Otherwise label the answer as reported influence. Do not overwrite analytics acquisition fields with a survey answer.
Use distinct labels such as verified AI-referred session, self-reported AI discovery, and AI-assisted deal. Before reporting pipeline or closed revenue, specify whether the rule is contact-, company-, or deal-based, how associated contacts qualify, and the lookback window. If contacts linked to one deal give conflicting answers, send the case to Revenue Operations for a documented decision.
HubSpot’s CRM API documents batch upsert using a property configured as unique. That can support repeatable CRM writes when a suitable unique key exists; it does not establish an AEO data feed or native AEO-to-CRM connection. For CRM property and reporting design, see HubSpot systems support. For contact-to-deal rules and ownership, see CRM systems consulting. The relevant API behavior is documented in the HubSpot batch upsert reference.
Build a repeatable reporting workflow
The sequence below is a proposed measurement design. Use an AEO product’s supported reporting for observations where available; otherwise maintain a documented review store. The AEO measurement lead owns prompt definitions, the analytics owner owns acquisition rules, and Marketing Operations or Revenue Operations owns discovery and deal rules.
Observation model and grain controls
The following model is illustrative, not vendor-published. It keeps raw observations, reviewed records, and reported summaries distinct.
- AI_PROMPT_RUN: one execution of one prompt, with
prompt_run_id, source system, prompt ID, prompt-set version, engine, model or variant, timestamp, geography, language, and raw response hash. - AI_BRAND_OBSERVATION: one brand’s result in one prompt run, with
observation_id,prompt_run_id, brand ID, appearance flag, mention type, placement, accuracy score, sentiment score, evaluator, and review status. - AI_CITATION: one displayed citation item, with
citation_id,observation_id, cited URL, cited domain, citation type, ordinal or normalized fingerprint, validity status, and target brand. - AI_VISIBILITY_PERIOD: one aggregate for a stated brand, engine scope, prompt-set version, period, metric version, and denominator definition.
- CRM_DISCOVERY_EVENT: one reported discovery response, with an event ID, contact ID, question version, raw answer, normalized source, reported timestamp, first-touch flag where justified, and consent or privacy status.
A prompt can have multiple runs on the same day, a run can contain multiple brand observations, and a response can contain multiple citations. Do not use prompt ID plus calendar date as the only key. A proposed prompt-run key should include the source system and stable run ID where available. A citation key can include the prompt-run ID, citation ordinal, and normalized citation fingerprint. Period aggregates require a period, prompt-set version, and denominator definition.
For concurrent ingestion, enforce uniqueness in the database and use an atomic or transactional upsert. A lookup-then-create sequence can race when two workers process the same event. If writing to HubSpot, use a configured unique property with the documented upsert mechanism. If writing to another store, use its database-enforced unique constraint. Retain parser and evaluator versions so revised scoring can be distinguished from new observations.
Four implementation patterns
1. Prompt-run visibility and citation observations
Trigger and input: a scheduled prompt-set review in an AEO product or documented manual test. Input includes the exact prompt, prompt ID, version, engine, timestamp, language, and geography.
Decision and output: record whether each tracked brand appears, its mention type and placement, and each displayed citation as a child record. If an AI tool helps extract candidate mentions or URLs, limit it to extraction and retain human review for material accuracy judgments.
Validation and destination: validate the prompt-set version, engine, timestamp, citation URLs, and unique run ID. Store the records in supported AEO reporting or a proposed observation store. Calculate period rates only from records with matching scope.
Fallback: if a field is unavailable, record it as unknown rather than infer it. If wording, engine coverage, or a high-risk product claim changes, route the run to the AEO measurement owner for review.
2. GA4 reporting for identifiable AI referrals
Trigger and input: a session arrives with a recognizable referrer or consistently applied campaign parameters. Preserve raw source, medium, referrer, campaign, landing page, timestamp, and configured conversion event.
Decision and output: apply deterministic custom-channel rules for known values, such as a documented source and medium combination. Record the channel-group version used for the report.
Validation and destination: test real source and medium values, rule order, and conversion configuration. Use GA4 reports or segments, while retaining raw acquisition fields for warehouse analysis.
Fallback: send unrecognized or missing-referrer sessions to an unclassified group. Never include direct/(none) in the AI channel without independent evidence. If channel rules change, publish the effective date and preserve the previous definition for historical interpretation.
3. Append-only self-reported discovery
Trigger and input: a lead form, demo request, onboarding form, or survey asks how the respondent first heard about the company. Store the question version, original response, selected sources, and response time.
Decision and output: normalize controlled selections such as ChatGPT, Gemini, Perplexity, another AI tool, search, referral, social, or unknown. Preserve multiple selections and free text rather than forcing one source.
Validation and destination: write a discovery event or CRM property subject to privacy review. Set a first-touch flag only when the question explicitly asks about first discovery. Keep the response separate from analytics acquisition data.
Fallback: route ambiguous free text, changed question wording, conflicting answers, or sensitive content to Marketing Operations. Preserve the raw answer and do not silently replace an earlier response.
4. CRM pipeline reporting under an explicit rule
Trigger and input: a validated discovery event is associated with a contact and, where applicable, a deal that reaches a defined stage. Required inputs include contact ID, deal ID, event ID, source type, reported timestamp, deal stage, and amount.
Decision and output: choose one rule, such as self-reported first discovery, verified AI-referred session, or AI-assisted deal. Report qualifying contacts, deals, pipeline, and revenue separately, with the rule displayed alongside the totals.
Validation and destination: confirm the source belongs to the controlled option set, the event key is unique, the contact-deal association is valid, and the lookback window applies. Use CRM reporting or a governed reporting store.
Fallback: quarantine duplicate events, missing associations, malformed dates, and conflicting contacts. Revenue Operations decides whether the deal qualifies; a model should not assign deal source from an unstructured association.
If a source system cannot provide a stable run or event identifier, create one from the source, scope, timestamp or supplied run ID, response hash, and citation ordinal where relevant. Use it for idempotency, but retain the original payload so a later reviewer can distinguish a duplicate from a genuinely repeated run.
What to report, and what not to claim
An executive view can show a visibility trend with its prompt-set version, verified AI-referred sessions and converting sessions, self-reported discovery as a separate measure, and pipeline or closed revenue under a published attribution definition.
Keep citation validity, sentiment, prompt-level placement, answer-quality exceptions, and ambiguous branded or direct trends as diagnostic signals. They can guide investigation without being presented as revenue evidence.
Do not treat third-party conversion comparisons as universal benchmarks. A reported result may be useful when its sample, period, population, and method are disclosed, but first-party measurement should drive your own reporting. Do not combine self-reported discovery, verified referrals, and AI-assisted deals into one number.
- Its question, unit, denominator, source system, reporting window, and owner are documented.
- Visibility includes the prompt-set version, engine scope, appearance definition, and run schedule.
- Prompt runs, brand observations, citations, CRM discovery events, and period aggregates are stored at their correct grains.
- Referral and self-reported discovery are separate evidence classes with tested validation rules.
- Contact-to-deal association, lookback window, and conflicting-evidence handling are published before pipeline or revenue is labeled AI-influenced.
- Direct sessions and branded-query movement remain directional signals rather than AI-attributed revenue.
Promote a KPI to routine reporting only when its source, denominator, time window, version or attribution rule, exception handling, and decision use can be reproduced.