One may evaluate an AI visibility platform on four things, and price is not one of them. Can you lock your own prompt panel? How many times is each prompt run? Can you export the raw answers and citations, and is the methodology documented publicly? A tool that clears all four is an instrument. A tool that clears none is a monthly screenshot subscription.

There is a version of this purchase that goes badly, and it is common enough to be predictable. A team buys a platform on the strength of a demo, reports the score weekly for two quarters, watches it move without explanation, and quietly stops opening the dashboard. The tool was never the problem. The evaluation was.

Here is how to run the evaluation properly.

Start by separating what is observed from what is modelled

Every platform in this category observes something real when it queries an engine and records the response. What it generally cannot observe is the complete set of questions buyers are actually typing, because major AI discovery platforms do not expose a full query stream comparable to traditional search-query reporting. Thus, AI visibility may actually fall short in the face of search engines.

So the prompt list is a model. Vendors build it from Search Console exports, keyword databases, People Also Ask expansions, semantic fan-out, your site content, or AI-generated brainstorming. Otterly and Ahrefs document their approach openly. Several competitors do not.

Your first question in any demo: where do these prompts come from and can you show me the documentation? A vendor who answers precisely is worth more than a vendor whose dashboard looks better.

The twelve-point checklist

Score each vendor out of twelve as a Bullzeye procurement screen. Eight or more means the vendor clears most of the governance basics below. Do not translate this checklist score directly into IAB Directional or Decision-Grade status; IAB quality tiers depend on query volume, intent coverage, reproducibility, validation, and platform coverage.

1. Prompt provenance and prompt taxonomy are documented, not just described on a sales call.
2. Total unique query volume is disclosed, with programs below 50 queries labeled Exploratory rather than Directional, consistent with IAB guidance.
3. You can supply and lock your own query panel built from first-party buyer language and documented category questions.
4. Each query is run multiple times per platform per period, and the provider discloses the response count.
5. Model version, reasoning mode, and collection method are recorded with every observation or reporting batch.
6. Raw answer text and complete reference lists can be exported, not just the composite score.
7. The headline metric is share of measured answers, or the composite formula and denominator are published clearly.
8. Variance, confidence, or reproducibility is disclosed alongside the headline number, with acceptable variation defined for repeated runs.
9. Brand mentions without visible citations are tracked separately from cited or referenced mentions.
10. Platform coverage is justified for the target market, per-platform results are available, and aggregation does not hide platform-specific differences.
11. Historical data, rebaselines, and methodology changes are disclosed, and you can take your query panel and raw historical data with you.
12. The provider states plainly which inputs are observed versus modeled and publishes material methodology changes before they alter your baseline.

Four contract clauses worth the redline

• Data portability. Raw observations, not summary exports, delivered in a machine-readable format within ten business days of termination.
• Methodology change notice. Written notice before any change to prompt generation, run frequency, engine mix or scoring, with the right to terminate if the change materially alters your historical baseline.
• Prompt panel ownership. Your panel is your confidential information. It came from your sales calls, and it should not persist in the vendor’s shared corpus.
• Rebaseline disclosure. If your score changes because the vendor changed something rather than because your AI visibility changed, that has to be disclosed in the reporting interface, not in a changelog nobody reads.

What good reporting looks like once you have bought

Three rules, and they matter more than which platform you chose.

Bullzeye default: use a ninety-day rolling view for executive reporting and avoid treating week-over-week movement as performance. Parse found that two repeat ChatGPT answers to the same prompt shared only 21.2 percent of cited domains across 693,509 answers. The ninety-day window is a methodology choice designed to reduce short-term noise, not an industry constant.

Report share of answers with the denominator visible. Twenty-two percent means little until it reads twenty-two percent of 1,000 observations across 50 unique queries, four platforms, and five responses per query.

Grade every number before it leaves the marketing team. Exploratory is the IAB category for programs below its query-volume floor or with insufficient coverage. Directional is suitable for pattern and trend monitoring when the program clears the IAB minimums and discloses its method. Supported is Bullzeye’s internal middle tier for repeated multi-engine evidence that is informative but not yet fully decision-grade. Decision-Grade is the IAB standard for evidence rigorous enough to support budget or strategy decisions.

The tools are not the fraud. The reporting convention is. A vendor selling you a modelled prompt set is being reasonable about a genuinely hard problem. A marketing team presenting that output to a board as a rank is not.

The one input no vendor can sell you.

A high-value starting query set is already inside your company: sales calls, support tickets, win-and-loss debriefs, and community threads. Those questions capture real buyer language that competitors cannot purchase. Use them as first-party input, then broaden the panel enough to represent the category you are trying to measure.

For an internal exploratory pass, pull ten recurring questions and run them repeatedly across a defined platform panel while recording the reference list from every response. For a Directional category read, expand to at least 50 unique queries with multiple intent types, consistent with the IAB query-volume floor. Then build one source inventory and test each source against who you serve, what problem you solve, and what category it assigns you.

The inconsistency log can be more actionable than the composite score alone, and a first exploratory pass can be completed quickly. Treat it as a diagnostic, not as proof of causal ranking impact.

Frequently asked questions

How much should an AI visibility tool cost?

Pricing varies from low-cost single-brand plans to enterprise contracts. Cost is not the useful first filter. A lower-cost tool with a locked query panel, documented methodology, and raw-data export may be more decision-useful than an expensive platform without them because you can audit the output.

Can I build my own AI visibility tracking?

For a small prompt panel, yes. Ten questions run five times across three engines on a monthly cadence is a spreadsheet and a few hours of disciplined work. You lose automation and historical charting. You gain complete control over the prompt panel and full AI visibility into method. Many teams should start here before buying.

Which AI engines should I track?

For many B2B programmes, ChatGPT, Google AI Mode and Perplexity are a practical starting panel, with Claude added where the audience skews technical, clinical or regulated. Treat that as a methodology choice rather than a market-share claim. Tracking every platform equally may waste budget or obscure platform-specific differences, so justify the panel and report results by platform.

How often should I report AI visibility internally?

Use monthly reporting for the marketing team and quarterly reporting for executives. Avoid presenting week-over-week movement to a board as performance because repeated-answer research shows substantial source churn; if weekly data is collected operationally, label it as short-term monitoring rather than a durable trend.

Do these tools track brand mentions without citations?

Some do, many do not, and the distinction matters. A mention that carries no visible citation may reflect parametric knowledge from training, a retrieval step the interface did not surface, or a synthesis across several sources. You cannot tell which from the output alone, so treat uncited mentions as a separate measurement category rather than assuming a cause. Ask vendors specifically how they classify them.

What is the difference between AEO and GEO?

In this framework, answer engine optimization focuses on making owned content clear, useful, and easy to surface as a direct answer, while generative engine optimization extends to the broader evidence environment that can influence whether and how a brand is referenced. For Google Search specifically, Google treats optimization for generative AI visibility features as part of SEO rather than a separate special-markup discipline.

Sources

Parse, AI citation volatility by industry, 8 July 2026. https://parse.gl/research/ai-citation-volatility-by-industry

Otterly, AI citations report, February 2026. https://otterly.ai/blog/the-ai-citations-report-2026/

Ahrefs, why ChatGPT cites one page over another, 31 May 2026. https://ahrefs.com/blog/why-chatgpt-cites-pages/

Google Search Central, Search Generative AI performance reports, 3 June 2026.

IAB, Measuring Visibility in the AI Era, 3 August 2026. https://www.iab.com/guidelines/measuring-visibility-in-the-ai-era/