Choose an AI search analytics tool by the decision you need to make and the grain of data required to make it. A free, one-time diagnostic can establish an initial snapshot. Recurring prompt monitoring helps compare the same buyer questions over time. Citation-level data shows which URLs and domains appear as sources, while page-level analysis helps diagnose an individual URL.
These tools do not measure the same thing. A brand mention, visible citation, recommendation, prompt-level observation, and aggregate share-of-voice score are different records or calculations. This guide compares documented plans as of October 10, 2026, then sets out a practical workflow for baselining, importing, deduplicating, and reviewing findings without treating vendor scores as standardized.
The implementation patterns are proposed operating designs, not vendor-published templates. Confirm the selected plan’s fields, exports, API access, retention, and refresh behavior before building around them.
Choose by the question and the data grain
Start with the question your team must answer. Then select the tool and plan that expose the necessary records, history, and access method.
- Snapshot: Is the brand present in a one-time diagnostic? This is useful for orientation, not for proving a recurring trend.
- Prompt-run observation: What happened for one prompt, engine or model, region, and execution or reporting time? Use recurring monitoring when the same buyer questions must be compared.
- Citation observation: Which URL or domain appeared as a source in a particular answer? Use citation-level data to investigate source patterns and content opportunities.
- Page diagnosis: What might explain the condition of one URL, including crawler activity, page health, or citation data? Use page-level analysis for this narrower task.
- Aggregate report: What is the summary across a defined prompt cohort, platform scope, and period? Use visibility or share of voice for directional reporting only when its population and calculation are documented.
If a competitor is cited for a recurring product-comparison question, a site-wide score is not enough. You need the relevant prompt observation and its associated citations.
A visibility number is not comparable until its prompt set, platform, region, time window, denominator, and aggregation method are known.
What the metrics mean, and why scores disagree
A prompt-run observation is one result for a specific prompt, engine or model, region, and execution or reporting time. One answer may contain no citations, one citation, or several. A brand mention means the answer names the brand. An explicit citation is a visible source reference. A recommendation presents the brand as an option or preferred choice. These states should not be collapsed into one field.
A page can be cited without the brand being recommended, and a brand can be mentioned without an owned page being cited. Some tools also distinguish content explicitly cited from content identified as a source used by the system. Store those states separately when the vendor exposes them.
Semrush documents separate metrics for mentions, citations, cited pages, sources, visibility, audience estimates, and average position. Profound defines Visibility Score as responses containing a brand divided by responses containing at least one brand. That is a vendor-specific aggregate over a selected population, not an industry standard or a per-response event.
Two tools can disagree because they sampled different prompts, engines, regions, or periods; captured responses on different schedules; or used different denominators and aggregation rules. Keep engine and model results separate before creating any combined summary.
Before reporting a metric, record its definition, grain, denominator, vendor, platform and model scope, region, prompt scope, date range, refresh or retrieval time, and aggregation version. If these details are unavailable, label the value as a vendor-defined aggregate and do not compare it directly with another tool’s score.
Semrush’s metric definitions provide a useful example of why the data grain matters. Profound’s visibility-score explanation illustrates why a documented denominator is essential.
Compare tools by documented coverage and plan limits
The following details reflect official vendor pages reviewed on October 10, 2026. Prices, quotas, coverage, and entitlements can change. Verify the selected plan before purchase.
| Tool | Verified starting point | Documented coverage or capability | Important qualification |
|---|---|---|---|
| HubSpot AI Search Grader | Free, one-time diagnostic | ChatGPT, Perplexity, and Gemini; five weighted dimensions | Its composite score is HubSpot’s proprietary diagnostic, not recurring monitoring. |
| HubSpot AEO | $50 per month advertised | Ongoing product positioned for ChatGPT, Perplexity, and Gemini; CRM-based prompt suggestions are described | Confirm cadence, history, exports, API access, and account entitlements. |
| Semrush AI Visibility Toolkit | $99 per month standalone; 25 prompt-tracking prompts listed | Reports cover ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, and Gemini; CSV exports are documented | Coverage and refresh schedules vary by report. Do not assume a uniform daily schedule. |
| OtterlyAI Lite | $29 per month | Base plans list ChatGPT, Google AI Overviews, Perplexity, and Microsoft Copilot | Gemini, Claude, and Google AI Mode are listed as add-ons. API and MCP access are plan-specific. |
| Profound | Starter $99 per month; Growth $399 per month, billed yearly | Starter tracks ChatGPT; Growth includes three answer engines; broader coverage is listed for Enterprise | Growth and Enterprise list CSV and JSON exports. API access is listed for Enterprise. |
| Peec AI | Price not verified | Prompt segmentation, daily execution for selected prompt and model combinations, citation analysis, CSV, Looker Studio, REST API, and MCP are advertised | Verify price, endpoints, limits, authentication, and plan entitlements before building. |
| SE Visible Basic | $99 per month | 200 prompts, three projects, five listed answer-engine surfaces, and a 10-day trial | The reviewed page does not establish a public API, export schema, or retention period. |
Shortlist two or three tools against the actual engine, model, region, language, prompt, citation, retention, export, and access requirements. A capability listed on a product overview or another plan is not proof that the evaluated plan supplies it.
Run a baseline that remains comparable
Build a controlled library of real buyer questions. Label each prompt by audience, buyer stage, intent, geography, priority, owner, first-seen date, and version. Keep vendor-suggested prompts in a separate discovery cohort because they can broaden research without representing customer demand.
For a one-time starting point, record the HubSpot grader date, platform scope, five dimensions, and weighted scoring context. The current page describes a composite score out of 100 weighted across sentiment, presence quality, brand recognition, share of voice, and market competition. Treat the result as a vendor-specific diagnostic, not a recurring baseline.
For recurring monitoring, hold prompt wording and scope steady where possible. Record additions, removals, prompt versions, model changes, and region changes. Review a defined multiweek period and require repeated observations or an adequately sampled change before assigning an action. There is no universal minimum sample size, so set a rule that reflects the cohort, cadence, and decision risk.
Design the data workflow: observations first, summaries second
Use a warehouse or relational database as the research system of record. Store each actual prompt execution separately from its citations, then calculate period summaries after ingestion. One row in prompt_run_observation represents one prompt executed for one vendor, workspace, engine or model, region, and run. One row in citation_observation represents one citation attached to that observation. A visibility summary is a separate aggregate, not a run event.
One answer can produce zero or many citation records. Link every citation to its originating observation, and keep share of voice in a period summary. Otherwise, an aggregate can be mistaken for an answer and the source relationship is lost.
The following is a proposed normalized model, not a vendor export format. Map only fields the selected vendor actually supplies. If an export has a report period but no exact execution time, store the report period and retrieval timestamp separately rather than inventing a run time.
{
"prompt_run_observation": {
"observation_id": "vendor-workspace-export-row-001",
"vendor": "illustrative-vendor",
"workspace_id": "workspace-01",
"prompt_id": "prompt-014-v2",
"engine": "Perplexity",
"model_variant": null,
"region": "US",
"executed_at": "2026-10-10T09:00:00Z",
"source_export_id": "export-2026-10-10-a",
"raw_response_hash": "sha256:illustrative-hash",
"ingestion_version": "1"
},
"citation_observation": {
"citation_id": "vendor-workspace-export-row-001-citation-001",
"observation_id": "vendor-workspace-export-row-001",
"cited_url_raw": "https://example.org/review",
"cited_url_canonical": "https://example.org/review",
"citation_position": 1,
"cited_domain": "example.org",
"extraction_version": "1"
}
}The proposed citation key includes the originating observation and citation detail, so two links in one answer can coexist. A proposed prompt-run key should include the vendor, workspace, prompt version, engine, model, region, and vendor run or response identifier where available. Do not use only brand, prompt, and date because multiple runs can occur on one day.
Keep summaries in a separate table with fields such as period_start, period_end, aggregation_version, prompt_scope, platform_scope, denominator, visibility_score, and share_of_voice. Keep CRM contact, company, deal, or task events separate again. A CRM record should represent an approved business action, not a raw answer or an aggregate row.
Semrush documents CSV exports. Peec advertises CSV, Looker Studio, REST API, and MCP. OtterlyAI lists API and MCP access on selected plans, while Profound lists CSV and JSON exports for Growth and Enterprise and API access for Enterprise. These facts do not establish common schemas, endpoint behavior, or native CRM writeback. Validate the selected vendor’s actual technical documentation before designing an import. CRM systems consulting can help define which approved summaries belong in an operational system.
Validate inputs and make replays safe
Validate required vendor, workspace, prompt, platform, model, region, reporting or execution time, URL, and numeric fields. Quarantine incomplete records instead of silently filling gaps. Preserve the original export row or permitted response alongside normalization, and record the source filename or export hash, retrieval timestamp, and ingestion version.
Prefer a vendor run or response ID as the unique key. If none exists, use a composite that distinguishes vendor, workspace, prompt version, engine, model, region, and execution or report time. Add a row hash when the source can repeat the same record. Enforce uniqueness with a database constraint and use a transactional upsert when concurrent workers can replay data. A lookup followed by an insert is not safe against concurrent duplicates.
Canonicalize URLs conservatively. Retain the raw URL, store the canonical value separately, and mark unresolved canonicalization rather than dropping a citation. Store vendor name, metric name, and metric version for vendor-generated sentiment or classification fields.
Send only approved summaries or tasks to a CRM. If middleware is used, verify the source export or API entitlement and document transformation, validation, retry, and exception behavior. Zapier automation consulting may help design a suitable handoff, but connector availability must be checked for the selected vendor and plan.
Turn a citation gap into a reviewed action
A single answer where a competitor appears and your brand does not is a lead for investigation, not proof of durable visibility loss. Verify the prompt, answer, engine, model, region, period, and citation URLs. Then classify the gap and assign it to the team able to investigate it.
- Owned-content gap: The question is not answered clearly on an appropriate owned page. The content owner checks the evidence and proposes an editorial change.
- Third-party-authority gap: A relevant publisher, review site, forum, or other external source appears. The PR or partnerships owner evaluates whether a credible contribution or correction is appropriate.
- Product-information gap: The answer uses incomplete or ambiguous product facts. Product marketing checks approved source material before proposing a change.
- Reputation gap: A source or answer raises a material sentiment concern. The reputation owner verifies the underlying source and escalates sensitive claims.
- Technical-access issue: A priority URL appears inaccessible or incomplete. Technical SEO checks page and crawl conditions without assuming that a technical change guarantees a citation.
- Sampling artifact: The prompt set, platform, schedule, or method changed. The analytics owner annotates the comparison and prevents it from being presented as a like-for-like trend.
AI can group paraphrased prompts or suggest a label from a controlled taxonomy. Preserve the source answer, evidence identifiers, classifier version, and review status. Use deterministic rules for URL validation, domain matching, duplicate detection, engine and date checks, and numeric thresholds. A person should approve factual, regulated, comparative, sentiment-risk, and positioning decisions before public content or CRM updates.
| Trigger | Bounded AI job | Validation gate | Action or fallback |
|---|---|---|---|
| Competitor appears and brand is absent on a priority prompt twice | Suggest a gap category and prompt theme | Review both answers, cited URLs, prompt version, engine, region, and observation count | Create a human-reviewed content, PR, product, or reputation task. If evidence is inconsistent, hold the task. |
| A cited URL cannot be canonicalized | None required | Retain raw URL and check redirect, format, domain, and access status deterministically | Store as unresolved citation. Do not discard it or infer that the source is absent. |
| Sentiment changes materially | Suggest context classification | Inspect source wording, vendor metric version, sample scope, and repeated observations | Route negative or ambiguous results to reputation review. Do not write sentiment directly to a CRM. |
| Import is replayed or a vendor schema changes | None required | Validate schema, source hash, row grain, and database uniqueness constraint | Transactional upsert valid rows and quarantine conflicts for data engineering review. |
Measure the workflow by verified observations, correctly linked citations, completed owner reviews, and documented actions. Treat referral traffic and conversions as separate acquisition measures. A visibility change alone does not establish that it caused a lead or sale.
Questions to ask in a vendor demo
Ask the vendor to demonstrate the records your team will actually use, not only an executive dashboard. Confirm whether the plan supplies response-level observations, citation-level rows, page diagnosis, or aggregates only. Ask how every metric is defined and what denominator it uses.
- Are prompts user-defined, vendor-suggested, or drawn from a prompt database?
- Are results live answer-engine runs, a vendor-owned database, or a mixture?
- Which engine, model, region, language, and location controls are available on this plan?
- What is the refresh cadence, and is an exact run timestamp available?
- How much history is retained, and is it raw response history or aggregate history?
- Do exports contain raw answers and citation-level rows, or only dashboard summaries?
- Is a documented API available on the selected plan? Ask for authentication, pagination, limits, stable identifiers, and retry behavior.
- Can the platform distinguish brand mention, explicit citation, source usage, and recommendation?
- What permissions, retention, compliance, and personal-data controls apply?
- Required grain: snapshot, prompt run, citation, page diagnosis, or aggregate.
- Required engines, models, regions, languages, and user-controlled prompts.
- Metric definition, denominator, report scope, and refresh cadence.
- Historical retention and whether history is raw or aggregated.
- Export fields and API entitlement on the specific plan.
- Permissions, data handling, and retention requirements.
- Named owner for validation, review, and action approval.
- Documented destination and success measure for approved findings.
Choose a one-time diagnostic when the need is initial orientation. Choose recurring monitoring when the team can maintain a comparable prompt cohort and review it. Consider a custom pipeline only after confirming exports or API access and estimating engineering, maintenance, and governance costs against the subscription.
Frequently asked questions
Which AI platforms should a team monitor first?
Start with platforms relevant to your audience and business questions. There is no universal order supported by the reviewed evidence. Add platforms when the team can maintain a comparable prompt and reporting scope, and report each engine separately before combining results.
How can a team measure business impact?
Track AI referral traffic as a separate acquisition segment where your analytics system can identify it. Report conversions with an explicit source definition, time window, and conversion event. Referral attribution does not prove that a visibility change caused a visit or conversion, so keep visibility and downstream outcomes as related but distinct measures.
How should a team handle model volatility?
Retain observation dates, prompt versions, engine and model history, and method-change annotations. Review a defined multiweek period rather than reacting to an isolated response. Escalate a change when it is repeated or sufficiently sampled for the decision at hand.
Can one vendor’s visibility score be compared with another’s?
Not safely unless the prompt set, platform and model, region, period, denominator, and aggregation method are aligned and documented. Otherwise, report each vendor-defined score in its own context.