No major AI discovery platform currently exposes a complete query stream comparable to traditional search-query reporting, so the “prompts your buyers use” in AI visibility reports are modeled rather than directly observed. Layer on citation sets that can change sharply between repeated runs, and a weekly rank screenshot carries little decision signal.
More useful measures are share of answers across a locked query panel, run-level consistency, source corroboration coverage, and evidence grades that travel with the number.
The three slides every vendor shows you.
Almost every demo in this category is built on the same three claims. Your buyers ask these questions. You appear here in the response. Your competitor appears above you.
The sequence works because it borrows the shape of rank tracking, which every marketing team already knows how to read. That borrowed familiarity is the problem, because the two things are not built the same way underneath. Ask a single follow-up question about where the question set came from, and most demos lose their precision.
The query data does not exist in any purchasable form
Some vendors are straightforward about this. Otterly documents plainly that AI search engines like ChatGPT and Perplexity do not publish query data and that no public keyword tool exists for AI search. Its prompt sets are assembled from Google Search Console data, SEO keyword research, and AI-generated brainstorming.
Ahrefs publishes its method too. Prompts are pulled from People Also Ask and its own keyword database, then expanded into related sub-questions through a semantic fan-out system. Both companies deserve credit for showing their work, because publishing a methodology is the only way a buyer can judge the number.
Others are less forthcoming. One well-funded platform advertises real prompts submitted by actual users without documenting anywhere I could find where those prompts originate. Another generates suggested prompts from your website and industry context, then scores those against Google search trends, which means a stand-in for buyer intent is being validated against a second stand-in.
The commercial distinction is straightforward. A modelled question set produces a modelled answer, however precisely the output is measured. That does not make the exercise worthless. It makes the exercise a hypothesis test, and a hypothesis should be priced, governed and reported as one. Very few of these reports are.
The platform operators hold demand data third-party AI visibility vendors cannot directly observe
Major platform operators such as Google, OpenAI and Anthropic hold demand data that third-party AI visibility vendors cannot directly observe. The important commercial question is not whether a vendor has data; it is which parts of the measurement are directly observed and which parts are modeled.
On 3 June 2026, Google shipped Search Generative AI performance reports inside Search Console, a dedicated view separating AI Overviews and AI Mode impressions from standard organic data for the first time. The report gives you impressions broken out by page, country, device, and date. It does not give you clicks, click-through rate, average position, or queries. There is no Search Analytics API access and no BigQuery export, and historical data only begins on 18 May 2026.
Read that carefully. Google built an entirely new reporting surface for marketers and left out the two fields marketers want most. Asked about adding query data, the company has said it is continuing to work with site owners to understand what insights would be most helpful.
OpenAI is equally clear on its advertising side that advertisers receive aggregate performance information only, with no access to chats, history, or memories. Anthropic does not run advertising in Claude at all, so there is no advertiser pipeline to petition.
When a platform operator does not expose the underlying query stream, a third-party AI visibility dashboard must model the prompt universe rather than observe it directly. That does not invalidate the dashboard. It means the prompt methodology has to be disclosed and the resulting score should be governed as a sample.
An AI visibility score is not a rank. It is a sample. If your vendor cannot hand you the prompt set, the run count, the date range, and a confidence grade behind the number, you are not buying measurement. You are buying a screenshot with a logo on it.
Even with a perfect prompt list, the answer keeps moving
Assume for a moment that the prompt problem is solved and your vendor somehow has the exact questions your buyers ask. The output still will not hold still.
In January 2026, SparkToro published crowdsourced research run with Patrick O’Donnell of Gumshoe.ai in which 600 volunteers ran twelve brand-recommendation prompts through ChatGPT, Claude, and Google’s AI Overview a combined 2,961 times. The odds of two responses to the same prompt returning the same list of brands came in under one in a hundred. The odds of returning that list in the same order were closer to one in a thousand.
Citations behave the same way. Parse measured repeat answers between 26 March and 25 April 2026 across 16,143 ChatGPT prompts and 15,805 Google AI Overview prompts, spanning 693,509 answers, and found that two repeat ChatGPT answers to the same prompt shared only 21.2 percent of their cited domains. Nearly four-fifths of the source set churned.
The instability is not only across time. Semrush found that only 25.6 percent of cited domains overlap between ChatGPT’s minimal-reasoning and high-reasoning modes for the same prompts. Same engine, same question, roughly three-quarters different sources depending on which mode the user happens to be in.
And a Somantra AI study of 2.4 million citation records found that 57.2 percent of websites referenced by ChatGPT and Google appear in exactly one month of results and are never surfaced again.
Position, in this environment, is not a stable property of your brand. It is one outcome drawn from a wide distribution, and it will be a different outcome an hour later. Aggregate AI visibility across hundreds of prompts over months does describe something real and worth managing. A single captured placement does not, and it is the single captured placement that keeps ending up in quarterly reporting.
IAB gives the market a shared measurement vocabulary and a higher bar for evidence
On 3 August 2026, the Interactive Advertising Bureau published Measuring Visibility in the AI Era. IAB’s own materials describe it as a standardized set of measurement guidelines and introduce the four Ps of AI visibility, plus Directional and Decision-Grade quality standards. The document also notes that more than twenty companies use different methodologies that can produce different answers for the same brand.
There is an important nuance. In a same-day AdExchanger interview, IAB’s VP of AI said the market is still too unstable for the document to function as a formal industry standard in the traditional sense. The practical reading is straightforward: shared definitions and quality thresholds now exist, while vendor implementation and reproducibility still vary materially.
IAB also treats fewer than 50 queries in a measurement program as Exploratory rather than Directional. And although IAB includes Position under Prominence, that is a property of a measured response, not proof of a durable search rank. The defensible CMO position is therefore narrower than “rank does not exist”: single-run position should not be governed like a stable ranking metric.
- Concentration. A 5W synthesis of more than 680 million citations concluded that the top fifteen domains capture roughly 68 percent of all citations. Evertune, analysing 200 million prompts across five months, concluded that even the single most-cited domain on any platform rarely exceeds 5 percent of total citations. Those two findings cannot be treated as measurements of the same underlying population.
- Which domain leads. Peec AI ranked Reddit the most-cited domain across five engines from 30 million sources. LLM Pulse gave YouTube the top spot at 26.47 percent with Reddit second at 17.39 percent. Attrifast, running 1,200 buyer-intent prompts across four engines, ranked Wikipedia first by share of citation slots at 8.9 percent.
- Whether phrasing matters. SparkToro found brand lists barely repeat. Peec AI, analysing 37,804 AI responses across 1,754 prompts, found that polished keyword-style prompts and casual conversational questions frequently surface the same brands, which implies far more stability than the SparkToro framing suggests.
None of these teams are wrong. Each picked its own prompt set, sample size, engine mix, model version, refresh cadence, and counting rule, then reported a single headline number. Different populations, different instruments, different answers. That is exactly the disclosure gap the IAB guidelines were written to close, and it is why a vendor quoting a definitive figure without publishing its method is overstating what it can know.
The practical read for a CMO: treat every AI visibility statistic, including the ones in this article, as directional evidence with a stated method attached. If the method is not stated, do not put the number in a board deck.
Your website is one part of a broader evidence environment
Your website matters because it is the source you control most directly, but it is not the only evidence layer an answer system can use when responding to category questions.
Muck Rack’s May 2026 Generative Pulse study analysed more than 25 million links from ChatGPT, Claude and Gemini responses across seventeen industries and found earned media accounts for 84 percent of all AI citations, with paid and advertorial content at 0.3 percent. Ahrefs found that 67 percent of the top 1,000 pages ChatGPT cites are structurally off-limits to brand SEO, meaning Wikipedia, government and educational institutions, app stores and major news media. You cannot pitch your way onto most of them.
Review platforms matter more than most B2B teams have budgeted for. G2 reports that AI models now cite roughly 80 percent of its product profiles more often than humans visit them. That finding comes from G2, which sells vendor AI visibility, and was authored by a paid G2 advisor, so weigh it accordingly. It still points the same direction as everything else.
The practical implication is not that third-party sources always outweigh your website. It is that buyer-facing AI answers can be assembled from a mix of owned and independent sources, so consistency across that wider evidence environment deserves active management.
Five metrics that survive the volatility
Volatility does not make measurement impossible. It makes single-point measurement impossible. These five hold up.
- Citation share across a locked prompt panel. Fix a panel of buyer questions, do not change it, and run every prompt multiple times per engine on a fixed cadence. The panel composition matters less than never changing it mid-flight. Report share of answers in which you appear, not position.
- Run-level consistency. Track how often you appear across repeated runs of the same prompt, not whether you appeared once. A brand that shows up in nine of twenty runs is in a materially different position than one that showed up in one of twenty, and a single-run report cannot tell them apart.
- Source corroboration coverage. Count the independent, non-owned sources that describe your company consistently. In Bullzeye’s methodology, this is a leading indicator we monitor because it creates a clearer evidence environment. We are not aware of a published correlation study proving that higher corroboration coverage causes durable AI inclusion, and the article should not claim one.
- Ninety-day rolling trend. Bullzeye uses a ninety-day executive view to smooth short-term retrieval noise and make trend discussions more disciplined. This is a methodology choice, not an industry constant. Report the window explicitly and keep per-platform data available underneath the aggregate.
- Evidence grade on every number. Use four levels in this package. Exploratory is the IAB category for programs below its minimum query-volume threshold or with insufficient category coverage. Directional follows the IAB concept for pattern and trend monitoring. Supported is Bullzeye’s operational middle tier for repeated multi-engine evidence that is informative but has not yet met full decision-grade reproducibility and validation. Decision-Grade follows the IAB quality standard for data robust enough to support budget or strategy decisions. Credit IAB wherever its terms are used.
The corroboration audit, step by step
Consistent information across independent sources gives a retrieval system a clearer corroborating evidence environment to work from. That makes a corroboration audit a practical, high-control workstream: align positioning, review profiles, customer stories, analyst coverage and owned pages before chasing short-term answer positions. Most companies carry some degree of positioning drift because different teams updated different assets at different times.
- Pull ten real questions from your own data. Sales call recordings, support tickets, win and loss debriefs, community threads. This is the one prompt source nobody can sell you, because you are the only party who has it.
- Run each question at least five times per engine across ChatGPT, Gemini and Perplexity. Five is a Bullzeye floor rather than an industry rule, chosen because it is the smallest count that reliably exposes run-to-run variance while remaining workable by hand. Record the reference list from every response, not only the text of the answer, because the reference list is the part you can influence.
- Build a source inventory from those reference lists. Your own pages, your review profiles, the forum threads, the press coverage, the team profiles on LinkedIn. Deduplicate it and put it in one column.
- Read each source against three tests. Who does it say you serve, what problem does it say you solve, and what category does it place you in? Log every answer that differs from your positioning, and log the difference rather than a pass or fail.
- Sequence the remediation by control. Pages, review profiles, professional profiles, and sales collateral are yours to change this month. Customer stories, analyst coverage, and press are influence work measured in quarters. Doing them in the wrong order is the most common way this audit stalls.
What comes back is an inconsistency log, and it is almost always longer than the team expects. Most of the entries have nothing to do with AI. They are the accumulated residue of a repositioning nobody propagated, a product name that changed on the website but not the review profile, a founder bio written for a different market. Those inconsistencies were always visible to anyone who read everything. Until recently, nobody read everything.
Twelve questions to ask before you renew
Take this list into your next vendor call. The answers separate the category quickly.
- Where do the prompts in my report come from, specifically, and can you show me the documentation?
- How many times is each prompt run, per engine, per reporting period?
- Which model version and which reasoning mode was used, and what happens to my historical data when the vendor changes it?
- Can I export the raw answer text and the full citation list, not just the score?
- What is the confidence interval or variance on the headline number?
- Can I supply my own prompt panel and lock it?
- How do you handle brand mentions that appear without a corresponding citation?
- What is your refresh cadence, and is it the same for every customer?
- Do you report share of answers, position, or a composite score, and how is the composite calculated?
- What happens to my score when an engine changes its retrieval mix, as ChatGPT did with Reddit in late 2025?
- Do I own the prompt panel and the historical data if I leave?
- Which of your published numbers are observed and which are modelled?
What to put in front of your board
Three changes, and they are governance changes rather than tooling changes.
Stop reporting position. Report share of answers across a locked panel, with the panel size and run count on the same slide. A number without its denominator is a claim, not a metric.
Grade every figure. Use Exploratory, Directional, Supported, and Decision-Grade. Exploratory, Directional and Decision-Grade reflect IAB’s measurement-quality language; Supported is Bullzeye’s operational middle tier. Executives are entitled to know which numbers can carry a budget decision and which cannot, and the label should travel with the number.
Report corroboration coverage as a leading indicator you can manage, not as a proven causal driver of citation share. Track the count of independent sources that describe the company consistently, then test whether improvements coincide with stronger inclusion over time.
None of this is a reason to stop investing in AI visibility. The buying behaviour is real. 6sense research covering more than 4,000 buyers found that 94 percent of buying groups rank a preferred vendor before first contact and purchase from that preliminary favourite 77 percent of the time. AI-assisted research is one of several inputs that can shape early consideration, which makes measurement discipline more important rather than less.
Frequently asked questions
Are AI visibility reports accurate?
Partially. The citation and mention data these tools collect are observed from actual model outputs. The prompt universe is not directly observed because major AI discovery platforms do not expose a complete query stream comparable to traditional search-query reporting. Vendors therefore build question panels from search data, keyword databases, first-party inputs, and generated expansion. That can be directionally useful when the methodology is disclosed.
Where do AI visibility tools get their prompt data?
From Google Search Console exports, third-party keyword databases, People Also Ask expansions, semantic fan-out generation, your own website content and industry context, and in some cases AI-generated brainstorming. Otterly and Ahrefs both publish their methodology. Several other platforms do not.
Can I see the queries behind AI Overviews and AI Mode in Search Console?
No. Google launched Search Generative AI performance reports on 3 June 2026, showing impressions by page, country, device and date. Clicks, click-through rate, average position and queries are not included, and the data is not available through the Search Analytics API or BigQuery export.
How volatile are AI citations?
Highly. Parse found two repeat ChatGPT answers to the same prompt shared only 21.2 percent of cited domains across 693,509 answers. Semrush found only 25.6 percent overlap between ChatGPT reasoning modes. SparkToro found under a one-in-a-hundred chance that two runs of the same prompt return the same brand list.
Does rank exist in AI search?
Not in any stable sense. Position within a generated answer changes between runs of the same prompt, so a ranked position captured at one moment is not a repeatable measurement. Share of answers across many runs is the meaningful equivalent.
What should I measure instead of AI rank?
Citation share across a locked query panel, run-level consistency across repeated runs, source corroboration coverage, a clearly stated trend window, and an evidence grade attached to every figure. If you use Bullzeye’s ninety-day executive view, label it as a methodology choice rather than an industry rule.
Do AI engines read my website when answering questions about my category?
Yes. Your website is an important owned source, but it is only one part of the evidence environment. Multiple citation studies show substantial use of earned media, reference sites, review platforms, and community content. Treat your website as the narrative you control, then audit the independent sources that AI systems may also reference.
Is it still worth paying for an AI visibility tool?
Yes, if you use it as a directional instrument with a locked prompt panel, raw data export, and a documented methodology, and no, if you use it to report weekly rank movements to executives. The tool is not the problem. The reporting convention around it is.
What does the IAB AI visibility framework actually standardise?
IAB describes Measuring Visibility in the AI Era as standardized measurement guidelines and a quality standard for evaluating AI visibility data. It defines the four Ps of AI visibility and separates Directional from Decision-Grade measurement. IAB’s VP of AI has also cautioned that the market is not stable enough to treat the document as a formal industry standard in the traditional sense. IAB further treats programs with fewer than 50 queries as Exploratory.
Does FAQ schema still help AI visibility?
FAQPage remains valid schema.org markup, but Google removed FAQ rich results from Search on 7 May 2026 and states that no special structured data is required for AI Overviews or AI Mode. Use FAQ markup only when it accurately describes genuine FAQ content; do not treat it as a Google visibility lever.
Sources
Google Search Central, Search Generative AI performance reports, 3 June 2026. https://developers.google.com/search
SparkToro and Gumshoe.ai, AI brand recommendation consistency study, January 2026. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
Parse, AI citation volatility by industry, 8 July 2026. https://parse.gl/research/ai-citation-volatility-by-industry
Semrush, ChatGPT reasoning modes and AI visibility, 30 June 2026. https://www.semrush.com/blog/chatgpt-reasoning-ai-visibility/
Peec AI, prompt variance and brand mentions, reported by Search Engine Journal, 6 July 2026. https://www.searchenginejournal.com/ai-prompt-intent-keywords-peec-spa/576201/
Peec AI, 30 million source citation analysis, reported by Search Engine Land, 31 March 2026. https://searchengineland.com/ai-search-engines-cite-reddit-youtube-and-linkedin-most-study-473138
Attrifast, AI citation rates by industry, 1,200-prompt benchmark, 26 May 2026. https://attrifast.com/blog/ai-search-citations-by-vertical-2026
Ahrefs, why ChatGPT cites one page over another, 1.4 million prompt study, 31 May 2026. https://ahrefs.com/blog/why-chatgpt-cites-pages/
G2, AI citations versus human pageviews, Kevin Indig, May 2026. https://sell.g2.com/resources/success/g2-products-ai-citations-vs-human-pageviews
6sense, 2025 Buyer Experience Report, November 2025. https://6sense.com/science-of-b2b/buyer-experience-report-2025/
IAB, Measuring Visibility in the AI Era, 3 August 2026. https://www.iab.com/guidelines/measuring-visibility-in-the-ai-era/
AdExchanger, IAB’s New Advice On How To Measure AI Search Visibility, August 2026. https://www.adexchanger.com/ai/iabs-new-advice-on-how-to-measure-ai-search-visibility/
Google Search Central, AI features and your website. https://developers.google.com/search/docs/appearance/ai-features
Google Search Central, FAQ structured data deprecation notice, 7 May 2026.