Choose AI citation tracking tools by the evidence your team needs to collect and verify, not by the length of a vendor’s engine list. A report about whether a page appeared in an answer needs prompt-run records and source URLs. A report about visits from AI tools needs human referral sessions. Those are different questions and require different data.
This guide compares publicly documented capabilities and plans checked on October 14, 2026, then shows how to make the resulting data auditable. Vendor plans and prices can change, so confirm contracted limits before purchase. The vendor examples describe published capabilities, not independent performance tests.
What AI citation tracking measures, and what it does not
Start by naming the unit you are measuring. An observation is one answer to one prompt on one engine in one run. A citation is a source link or domain shown in that answer. A brand mention is a reference to a company, product, person, or other tracked entity, whether or not a source is linked. A citation can appear without a brand mention, and a mention is not automatically a citation.
HubSpot’s AEO documentation defines citations as webpage links or domain names embedded in an answer and measures brand mentions separately. It does not establish unattributed paraphrases as a confirmed citation type. Treat a suspected paraphrase as a lower-confidence research signal, not as verified source attribution.
Visibility and share of voice are aggregates, not individual citations. Their meaning depends on the formula and denominator. HubSpot describes share of voice using brand mentions relative to tracked competitors’ mentions. Profound distinguishes response-level visibility from mention-based share of voice. Report the calculation, cohort, and period alongside every result.
Keep two further signals separate. AI referral sessions are human visits attributed in an analytics system. Crawler or agent requests are server-side requests classified as AI-agent activity. Scrunch documents these sources separately. A retrieval request does not prove that a page appeared in the final answer, and a citation does not prove that anyone clicked it.
A visibility number is useful only when its observation, denominator, engine, cohort, and reporting period are explicit.
Choose the tool by the evidence you need
Before comparing prices, write down the engines, prompt volume, run frequency, regions, languages, response limits, integrations, and export requirements you actually need. Then compare total cost and effective observation capacity for that same scope. A prompt allowance is not automatically comparable to a response allowance. Ask whether prompts are counted once or multiplied by engine, region, run, or response.
| Tool | Documented fit | Public plan snapshot |
|---|---|---|
| HubSpot AEO | Daily prompt monitoring and citation analysis across ChatGPT, Gemini, and Perplexity. | $50 per month standalone for 25 daily prompts and 2,500 monthly answers. Additional tracking packs are listed separately. |
| OtterlyAI | Prompt monitoring with a defined base engine set and separately priced engine add-ons. | Lite $25 per month, Standard $160, and Premium $422. Base engines include ChatGPT, Google AI Overviews, Perplexity, and Microsoft Copilot. Google AI Mode, Gemini, and Claude are listed as add-ons. |
| AirOps | Visibility analysis alongside content operations. Coverage depends on product area and plan. | Solo is described as 100 tracked prompts and pages with ChatGPT-only insights. Confirm current pricing and broader plan coverage. |
| Profound | Answer-engine insights with broader capabilities described for Enterprise. | The public page lists a Trial of 10 prompts run once on ChatGPT and custom-priced Enterprise. Enterprise features include up to nine engines, CSV and JSON exports, and API access. |
| Peec AI | Focused visibility analytics with official materials describing model selection and regional and language support. | The reviewed official material does not establish current plan prices. Verify pricing and allowances directly. |
| Scrunch | Visibility analytics when referral and agent-traffic measurement may also matter. | Core is $250 per month for 125 unique prompts, 5,000 monthly responses, and four platforms. Enterprise lists nine platforms. |
HubSpot’s standalone allowance is 25 daily prompts across three engines. Its catalog lists additional tracking packs at $20 per month for 10 prompts and 1,000 additional monthly answers. HubSpot also offers a free AI Search Grader, described as a one-time diagnostic rather than continuous monitoring. Marketing Hub Professional and Enterprise have their own AEO limits, including 25 daily prompts and 2,500 monthly answers for Professional and 50 daily prompts and 5,000 monthly answers for Enterprise, according to HubSpot’s documentation.
Do not transfer broader platform coverage to a specific AirOps plan. Do not assume Profound Enterprise features apply to the Trial. Do not assume Scrunch Core includes Enterprise’s nine-platform coverage. For Peec AI, treat the reviewed instructional page as product guidance rather than a pricing contract. These are vendor statements, not comparative tests of accuracy or performance.
Normalize the comparison using an effective-capacity calculation:
effective monthly observation capacity = tracked prompts multiplied by engines multiplied by runs per month
Report the subscription, prompt add-ons, engine add-ons, response or credit caps, regional and language expansions, export or API charges, and overage rules separately. If the vendor does not define a unit clearly, the apparent price advantage is not yet comparable.
Buy for the decision you need to support. A monitoring-first product may suit a team that already owns its content and reporting workflow. An execution platform may suit a team that needs content operations and visibility analysis together. A crawler or referral product is relevant only when agent access or human sessions are separate requirements.
Build a comparable prompt set before reading trends
Make prompts reflect real buyer questions. Keep branded prompts separate from category and problem-led prompts. Record each prompt’s language, region, persona, funnel stage, target engine, and effective date. Maintain a versioned prompt set and competitor list. A wording, engine, region, or competitor change can alter the comparison, so do not silently apply a new configuration to historical results.
Compare like with like: the same prompt-set version, engines, region, language, date window, competitor set, and calculation rule. HubSpot documents daily runs across ChatGPT, Gemini, and Perplexity and recommends evaluating trends over multiple days or weeks because answers change. Weekly review of high-priority prompts and monthly review of a broader set are reasonable starting policies, not universal requirements.
When comparing tools, ask whether a prompt is counted once or multiplied by engine, region, run, or response. Request a sample export and field definitions before procurement. A dashboard score without its underlying grain cannot tell you whether a change came from an answer, a prompt-set revision, an engine change, or a denominator change.
Turn answer captures into auditable citation records
Keep data at its natural grain. One observation row should represent one prompt, one engine variant, one execution, one region, and one language. Each cited URL or domain occurrence belongs in a separate citation record linked to that observation. Store brand mentions separately. Calculate visibility and share-of-voice summaries later, with their own cohort and calculation version.
The following is a hypothetical independent data design, not a vendor export schema. It represents one prompt run and contains no citation or aggregate fields:
{
"observation_id": "obs_8f31",
"execution_id": "run_20261014_0900_042",
"prompt_id": "p_042",
"prompt_set_version": "2026-10-v2",
"engine": "ChatGPT",
"engine_variant": "record-if-exposed",
"region": "US",
"language": "en",
"collected_at": "2026-10-14T09:02:00Z",
"raw_answer_uri": "secured://observations/obs_8f31"
}The execution ID makes repeated runs of the same prompt distinguishable. If a system can run twice at the same timestamp, include a collection-run ID or another source-generated execution identity. Do not use a key containing only date, prompt, and engine.
For citation occurrences, retain the raw cited URL before normalization. Store the normalized domain, final URL after redirects, reachability status, citation ordinal, and parent observation ID separately. Do not invent a page URL from a domain-only citation. For mentions, retain the answer span, matched variant, classifier version, and confidence. Use an AI classifier only where interpretation is needed, such as ambiguous context or sentiment.
For concurrent collection workers, a search-before-create check is not sufficient. Two workers can both see no record and insert duplicates. Use a database uniqueness constraint or transactional upsert. A suitable observation key includes prompt ID, engine and variant, execution ID or precise run identity, region, and language. Citation occurrences should include the parent observation ID plus a citation ordinal or normalized URL. Aggregate records should include the brand, prompt-set version, competitor-set version, engine, region, language, reporting period, and calculation version.
These controls are proposed implementation patterns, not documented vendor features. Their purpose is to preserve provenance when a dashboard export, manual capture, or monitoring product supplies the original answer data.
Use a practical workflow for each evidence type
The same answer can produce several records, but those records answer different operational questions. Keep the workflow explicit rather than writing every result into a single visibility field.
| Trigger | AI job | Validation and destination |
|---|---|---|
| A scheduled prompt run or approved manual capture | Generate or receive the answer being observed. A classifier may identify candidate mentions, but it must not invent absent URLs or fields. | Validate execution identity, prompt-set version, timestamp, and required context. Store the observation in an append-only analytics or data store. |
| An answer contains a source link, domain, or possible brand reference | Parse URLs deterministically and classify ambiguous mentions only when needed. Return evidence spans and confidence. | Preserve the raw URL, validate redirects and ownership, and send low-confidence or reputation-sensitive findings to an SEO, content, communications, or legal reviewer. |
| A reporting period closes after validation | Draft an explanation of already calculated changes. Deterministic code calculates counts, rates, and denominators. | Check source observation IDs and all cohort versions. Publish to BI or analytics; write only reviewed, scoped summaries to a CRM. |
| A referral or agent request is recorded | No AI inference is required to call a session or request a citation. Classify the source as analytics referral or server-side agent activity. | Keep GA4 sessions and CDN or hosting requests in separate tables. Do not merge either signal into citation counts. |
Escalate alleged misinformation, negative executive or product sentiment, and revenue interpretations before they enter executive reporting or a CRM. AI can draft an explanation of an already calculated change, but it should not silently redefine a metric or convert an uncertain classification into a business event.
Keep visibility, referrals, and crawler activity separate
Publish separate measures for answer presence, citation rate or count, brand mention rate, share of voice, human referrals, and bot requests. Each aggregate should name its denominator, engine, prompt cohort, competitor set where relevant, region, language, period, and calculation version.
Profound defines Visibility Score as the number of responses containing the brand divided by the number of responses containing at least one brand. Its Share of Voice measure counts mentions rather than merely response presence. HubSpot’s documented share of voice is also based on tracked brand mentions relative to competitors. These definitions should not be substituted for one another.
Scrunch separates AI Referrals measured through Google Analytics 4 from Agent Traffic measured through a CDN or hosting integration. Referral data concerns sessions, views, configured conversions, sources, and landing pages. Agent traffic concerns requests classified as training, indexer, or retrieval activity. A retrieval request is an accessibility or retrieval signal, not proof of final-answer inclusion. A referral session records a visit, not proof that a citation caused a conversion.
Assign a system of record for raw observations, validated citations, and approved business summaries. If teams need CRM context, write back reviewed summaries with the source IDs, reporting period, and calculation version. This supports governed CRM systems without treating an aggregate visibility measure as an attributable contact or deal event.
A practical selection and launch decision
Do not sign off on a dashboard demonstration alone. Run representative prompts, inspect answer and source records, change a prompt to test versioning, export a sample, and confirm the exact plan’s engines and limits. Ask what one tracked unit means, what denominator each metric uses, and what changes when prompts, competitors, engines, or regions change.
- Confirm included engines, add-ons, prompt counting, run frequency, and response or credit limits for the contracted plan.
- Inspect a sample answer and export with prompt, engine, region, timestamp, source URL or domain, and brand mention represented distinctly.
- Run the same prompt twice and verify that separate executions remain separate observations and that repeated citations retain parent observation IDs.
- Change a prompt or competitor list and confirm that the new prompt-set or competitor-set version does not silently rewrite historical reporting.
- Document export or API fields, access terms, retention, authentication, and limits from current vendor documentation or the order form.
- Assign a prompt-governance owner, a validation and metric-definition owner, and a business owner for approved actions.
A monitoring product is useful when its evidence fits a defined decision and your team can preserve the context behind the metric. Start with a small, versioned prompt cohort, validate the records, and expand only when the resulting data can be interpreted consistently.