--- title: "AI visibility reports: what they measure and what to set up first" description: "What an AI visibility report measures, what public studies say about unstable answers, and what to ask and set up before paying for one." url: "https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/" markdown_url: "https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar.md" lang: "en" date_published: "2026-11-12" author: "Martín Garay" author_url: "https://cookingmetrics.com/about-martin-garay/" publisher: "CookingMetrics" publisher_url: "https://cookingmetrics.com/" keywords: ["SEO for AI (GEO)"] citation_apa7: "Garay, M. (2026, November 12). *AI visibility reports: what they measure and what to set up first*. CookingMetrics. https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/" --- # AI visibility reports: what they measure and what to set up first > What an AI visibility report measures, what public studies say about unstable answers, and what to ask and set up before paying for one. - **Author:** [Martín Garay](https://cookingmetrics.com/about-martin-garay/) · Founder of CookingMetrics - **Published:** November 12, 2026 - **Topics:** SEO for AI (GEO) - **Web version:** https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/ ## How to cite this article (APA, 7th edition) Garay, M. (2026, November 12). *AI visibility reports: what they measure and what to set up first*. CookingMetrics. https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/ In-text citation: (Garay, 2026) --- **An AI visibility report counts how often a brand appears in an assistant's answers to a set of questions chosen by the provider.** Public studies show that those answers change between repetitions, so the method and the number of repetitions matter. Before paying, it is worth measuring your own traffic and requiring the provider to explain how it measures. This guide is written for the person who has to decide whether to buy a report of this kind, not only for the person who sets it up. Technical terms are explained in plain words and the detail appears under the label "For the technical team". It is worth stating where this is written from: [CookingMetrics](https://cookingmetrics.com/) offers measurement audits and support to development teams, and develops its own applications. It does not sell or resell visibility tools, but it does have an interest in every site measuring its own data. That "visibility tools are not a reliable basis for deciding" is an opinion of CookingMetrics; what follows separates that opinion from what the sources say. ## In summary - A visibility report measures appearances in model answers to defined questions, not visits or sales. It is a sample, and its value depends on how many questions and how many repetitions it includes. - A SparkToro study (January 27, 2026, not peer-reviewed) found that, when the same question was repeated about one hundred times, the list of recommended brands almost never repeated. Its author holds that a position in an AI ranking is not a meaningful measure and that a visibility percentage over many repeated questions may be. - An Authoritas study (February 19, 2025) measured the change in the sources cited by Google's AI Overviews, not in chat assistants. Those changes were larger than those in organic results over the same period. - "Visibility is not revenue": a brand appearing more does not mean it sells more. The relationship exists or not depending on the business, and it is checked with your own conversions; it is neither taken for granted nor ruled out. - Before paying, it is worth having your own measurement (visits, crawling, citations in the consoles) and bringing the provider a short set of questions about its method. ## What a visibility report measures The following description is general and is a synthesis made for this guide; methods vary and each provider should document its own. A typical report works like this: the provider defines a list of questions that a customer might ask an AI assistant ("which is the best…", "which provider is best for…"). It sends them to one or more assistants, stores the answers and counts in how many of them the client's brand appears, sometimes with its position in a list or with a rating. The result is shown as a visibility percentage, an average position or a comparison with competitors. That is very different from what the site's own tools measure: - **A visibility report** estimates appearances in answers to chosen questions. It does not record any real person. - **Analytics and server logs** measure visits and the activity of real robots on the site. They do not know what was answered. - **Search engine consoles** report impressions, clicks and, in the case of Bing Webmaster Tools, citations, for the verified domain. The guide [How to measure AI assistant traffic](https://cookingmetrics.com/blog/como-medir-trafico-asistentes-ia/) describes those layers and their limits. None of the three replaces the others. Analytics and logs describe activity on the site; search consoles report appearances or clicks on their surfaces; a visibility report observes answers within a selected sample. Impressions and citations can happen without a visit to the domain. ## What the public evidence says about the stability of answers Two public studies serve as a reference. Neither proves that all tools are unreliable, and both have limits that their own authors state. ### SparkToro, January 2026: chat assistants - **What they did:** [Rand Fishkin (SparkToro)](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), with Patrick O'Donnell, who according to the article itself works at a company in the AI tracking sector, recruited 600 volunteers. They asked 12 different questions, each repeated between 60 and 100 times, until reaching 2,961 answers from ChatGPT, Claude and Google's AI Overviews (with AI Mode when the former did not appear). The data was collected in November and December 2025. - **What they found, according to the authors:** when asking for the same recommendation about one hundred times, the probability of getting the same list of brands twice was below one in one hundred; that of getting them in the same order, on the order of one in one thousand. With questions written by different people for the same need, the similarity between the texts was very low (0.081 on the measure the authors used), but some brands appeared with high frequency (from 55% to 77% of the answers, for four headphone brands). - **What the author concludes:** that a "position in the AI ranking" makes no sense as a measure and that a visibility percentage over dozens or hundreds of questions, each repeated several times, is a reasonable measure. He recommends checking that the provider publicly documents its method. - **Stated limits:** it is not peer-reviewed; the authors say they are not professional researchers; they did not control for location, device or history; the data is from November and December 2025 and models change; they did not record how the answers describe the brands; they ask for much larger samples. ### Authoritas, February 2025: Google AI Overviews - **What they did:** the company's chief executive, [Laurence O'Toole](https://www.authoritas.com/blog/serp-organic-and-ai-overview-volatility-research), gathered 11,203 keywords on Google.com in the United States, on desktop, and queried them on three dates: August 23, 2024, October 17, 2024 and January 17, 2025. Only 2,104 showed an AI Overview, and the analysis of those answers uses only that group. - **What it measured:** how much the list of pages cited in the AI Overview changes between two dates, on a scale from 0 to 1 where a higher value means more change. The author calculates it as a distance between lists that gives more weight to the first positions and does not publish the exact weights on the page. - **What it found:** between August and October (about eight weeks), the change in the cited sources was 0.68, against 0.49 in organic results; between October and January (about thirteen weeks), 0.73 against 0.55. The author summarizes that in two or three months a proportion close to 70% of the cited pages can be expected to change. - **Stated limits:** desktop only and United States only; three dates; a small subset of keywords with an AI Overview; no significance tests or confidence intervals are reported; the author warns that some of his metrics are not comparable with each other. ### What can and cannot be concluded - **It can be said** that, in those two studies, the answers and the cited sources changed a lot between repetitions and between dates, and that a single data point from a report may not represent what another person will see. - **It cannot be said** that all visibility tools are unreliable. The studies did not evaluate any specific tool, and a tool that repeats each question many times and publishes its method answers, at least in part, to what the SparkToro author considers reasonable. - **It cannot be extrapolated** from the second study to chat assistants, or from the first to AI Overviews: they measure different things, on different dates and with different methods. - **Both studies** are more than nine months old (the Authoritas one, almost two years) and models change frequently; it is worth looking for more recent work before citing figures. **For the technical team.** The "volatility" in the second study is a distance between lists (the authors describe it as a weighted Levenshtein distance): it counts how many position changes, additions and removals are needed to go from the list on one date to the list on another. It does not measure the quality of the answer or whether the change was good or bad for a brand. If the provider reports a visibility percentage, the technical question is how many repetitions back that percentage: with few repetitions, the margin of error is large. [Section for signed-in readers, level C: https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/#bloque-1] ## What to ask a provider before paying These questions are an editorial criterion of CookingMetrics, not an industry standard. They help compare proposals without needing technical knowledge. 1. **Who defines the questions, and can I see and change them?** The questions should resemble those your customers ask, not only those the provider chose. In the SparkToro study, different people wrote the same need in very different ways. 2. **How many times is each question repeated and how is the variation reported?** A percentage without the number of repetitions is hard to interpret. SparkToro suggests on the order of 60 to 100 per question; it is a recommendation by that author, not a norm. 3. **Under what conditions is it queried?** Country, language, device, whether the interface people see is used or a technical connection (API) to the model. The cited studies did not control for these conditions, and there is no way to know from here whether they change the result. 4. **Which assistant and which version?** Models change. A report without a query date and without a model is hard to compare with the next one. 5. **How is it decided that a brand "appears"?** Text mention, link, recommendation. And what if the answer is wrong or negative? The SparkToro study did not record that information. 6. **Can I export the complete answers, with dates?** If the provider does not hand over the answers, there is no way to review the percentage. This, and the method being public, is what the SparkToro author recommends checking. If the provider cannot answer three or more, the figure remains an estimate, but not one that can be defended before a board of directors. [Section for signed-in readers, level C: https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/#bloque-2] ## Visibility is not revenue A brand may appear in many answers and not sell more, or sell more without appearing in them. The phrase "visibility is not revenue" does not claim that visibility does not generate sales. It claims that they are not equivalent: the relationship has not been measured in general and depends on the business, the product and who is asking. This guide cites no study that quantifies that relationship and, therefore, neither affirms nor denies it. What an organization can do is check it with its own data: - **What to look at:** the conversions attributable to visits from AI assistants and from the search engine, and how many of them correspond to the pages or topics that the report says improved. - **With what care:** if no recorded visit arrives from assistants, or if there are few, there is no basis for drawing conclusions in either direction. The detail of what can be measured is in [the layered measurement guide](https://cookingmetrics.com/blog/como-medir-trafico-asistentes-ia/). ## What to set up before paying Before buying a report, it is worth having the following in place. It is a recommendation from CookingMetrics and its order is a proposal. - **The verified consoles.** The domain verified in Google Search Console and in Bing Webmaster Tools, to see impressions, clicks and citations on the site itself. - **Analytics with the AI assistant channel reviewed.** See which visits already arrive and which cannot be attributed. - **Server or CDN logs.** To know which AI robots read the site. The guides on [accessibility for AI robots](https://cookingmetrics.com/blog/test-accesibilidad-bots-ia/) and on [Cloudflare for SEO and GEO](https://cookingmetrics.com/blog/cloudflare-seo-geo-guia-tecnica/) explain how to check access. - **A framework for interpreting the results.** The note [what GEO is](https://cookingmetrics.com/blog/que-es-geo-generative-engine-optimization/) describes the concept. With that in place, the report becomes a complement and not the only source of the data. [Section for signed-in readers, level C: https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/#bloque-3] ## When it may make sense to evaluate a tool It may make sense when the organization already measures its own data and needs to see something that the consoles do not offer, for example how it appears against competitors in specific questions from its customers, in several countries or languages. A reasonable way to decide is a **repetition test**, a method proposed in this guide that CookingMetrics has not yet run: 1. Choose between 5 and 10 questions that your customers really ask, written by them, not by the provider. 2. Ask each one several times, in new sessions, in the same country and language, and note in how many the brand appears and with which sources. 3. Ask the provider for the same data for the same questions. 4. Compare. If the provider and your own test differ greatly, ask why before paying. 5. Record the date, the assistant and the version, so the test can be repeated later. With few repetitions, the test is not statistically solid; it serves to detect large differences, not small ones. [Section for signed-in readers, level C: https://cookingmetrics.com/blog/reportes-visibilidad-ia-que-miden-cuando-pagar/#bloque-4] ## What not to conclude - That a high visibility percentage means more sales. There is no evidence cited here for that relationship. - That a position in an AI ranking is a stable measure. According to the author of the SparkToro study, it is not. - That the two studies represent all tools or all assistants. - That one tool is better or worse than another. This guide does not compare providers or recommend any. ## What to do next If your team is evaluating a visibility tool, bring the six questions to the meeting and compare the answers in writing. If you do not yet measure the visits, crawling and citations of your own site, start there. If you would like us to review that base, learn about our [SEO for AI search service](https://cookingmetrics.com/services/seo-ai-geo/): we review the setup and support your team, and we do not sell visibility tools. ## Frequently asked questions ### What is an AI visibility report? It is a report that estimates how many times a brand appears in AI assistants' answers to a set of questions chosen by the provider. It does not record visits from people or sales. ### Why do an assistant's answers change between repetitions? The cited studies observed that the lists of brands and the cited sources change between repetitions and between dates. This guide does not explain the cause; it only reports what was observed. ### Is a visibility percentage a reliable measure? According to the author of the SparkToro study, it can be if it is calculated over dozens or hundreds of questions repeated several times and the provider documents its method. With few questions or few repetitions, the margin of error is large. ### Does AI visibility increase sales? This guide has no data that quantifies it. It can be checked with your own conversions from visits that arrived from assistants; if there are few visits, no conclusion can be drawn. ### Do I need to pay for a tool to measure AI visibility? Not necessarily. The Google and Bing consoles, analytics and server logs measure visits, crawling and citations on your own site, without depending on questions chosen by a provider. A visibility tool can complement that base, not replace it. --- ## About the author **[Martín Garay](https://cookingmetrics.com/about-martin-garay/)** — Founder of CookingMetrics I'm Martín Garay, founder and Martech Advisor at CookingMetrics and Chair of the Technology Committee at CACE, Argentina's ecommerce chamber. For more than 20 years I've worked in Analytics, development, technical SEO and ecommerce: I implement measurement, find where it breaks and turn it into information you can act on. ### My work at CACE, Argentina's ecommerce chamber - **Chair of the Technology Committee** (SINCE AUGUST 2025): I set the committee's agenda together with Gabriel Topola, vice-chair. We look at the platforms and tools an online business needs to grow: CMS, integrations, customer service and omnichannel operations. - **Curating the technical forum** (IT CAMP · ECOMMERCE DAY ARGENTINA): With CACE's Events team I put together the IT Camp programme: talks for engineering and development teams on platform architecture, code and good practice, given by speakers who do the implementation work. - **Marketing and Data Analytics committees** (SINCE 2016): I started in the Marketing committee and helped create the Data Analytics committee. I gave Google Analytics webinars, including the migration to GA4, and worked on the Google Analytics implementation for Hot Sale and Cyber Monday. - **Ecommerce metrics dashboard** (FOR MEMBERS): With the Data Analytics committee and the CACE Community team we built a dashboard in Google Analytics and Looker Studio so each member can check the key indicators of their own store. ### Talks - [From SEO to GEO: what it takes for a brand to be cited by an answer engine](https://www.youtube.com/watch?v=mlyjuSBdMzI) — ECOMMERCE DAY ARGENTINA 2026 · IT CAMP · AUGUST 27, 2026 · BUENOS AIRES CONVENTION CENTER - [SEO for AI: how to prepare a brand to show up in ChatGPT and Google's AI overviews](https://www.youtube.com/watch?v=4dI7jqnLg-M) — AI IN LATAM 2025 · BUENOS AIRES · OVER 3,500 ATTENDEES AND 100 SPEAKERS, PER THE ORGANIZERS ### Profiles and links - [Full profile](https://cookingmetrics.com/about-martin-garay/) - [LinkedIn](https://www.linkedin.com/in/seospecialistmartingaray/) - [Instagram](https://www.instagram.com/garaymartin/) - [YouTube](https://www.youtube.com/@MartinGaray) - [martingaray.com.ar](https://www.martingaray.com.ar/) - [Chair of the Technology Committee at CACE, Argentina's ecommerce chamber](https://cace.org.ar/pages/comision-de-tecnologia?utm_source=cookingmetrics.com&utm_medium=referral&utm_campaign=perfil-martin&utm_content=caja-autor) - [Ecommerce metrics dashboard, free for CACE members](https://cace.org.ar/pages/dashboard-de-metricas?utm_source=cookingmetrics.com&utm_medium=referral&utm_campaign=perfil-martin&utm_content=caja-autor) --- ## About CookingMetrics CookingMetrics provides Google Analytics consulting, measurement implementation and SEO. Founder Martín Garay brings more than 20 years of experience in Analytics, development and ecommerce. The work connects data with business decisions. ### Services - **[Google Analytics audit](https://cookingmetrics.com/data-analytics-consulting/)**: Do purchases or revenue differ from your store? We review your measurement and deliver a correction plan in 10 business days. [See the audit scope](https://cookingmetrics.com/data-analytics-consulting/). - **[Google Tag Manager](https://cookingmetrics.com/services/gtm/)**: Did events stop being tracked after a website change? We review tags, triggers and versions to find duplicates or missing data. [Ask about Google Tag Manager audits and implementation](https://cookingmetrics.com/services/gtm/). - **[Server-side measurement](https://cookingmetrics.com/services/cookieless-tracking/)**: We assess how to send data from your server, respect consent and prevent duplication. [What's included in server-side measurement](https://cookingmetrics.com/services/cookieless-tracking/). - **[Technical SEO](https://cookingmetrics.com/services/technical-seo/)**: Have you migrated your site, or are pages missing from Google? We review crawling, indexing, structure and performance. [How we approach technical SEO](https://cookingmetrics.com/services/technical-seo/). - **[SEO for AI search (GEO)](https://cookingmetrics.com/services/seo-ai-geo/)**: We review whether search engines and assistants can access your pages and understand their content. [What we can improve for AI search](https://cookingmetrics.com/services/seo-ai-geo/). ### Solutions - **[Google Analytics self-check](https://cookingmetrics.com/google-analytics-self-check/)** (Status: Available) - **[GA4 Report Generator](https://cookingmetrics.com/product/ga4-report-generator/)** (Status: Beta): A tool in beta that already generates Google Analytics reports. Ask about available features and the scope for your project. [Explore GA4 Report Generator](https://cookingmetrics.com/product/ga4-report-generator/). - **[IndexNow Connect](https://cookingmetrics.com/product/indexnow-connect/)** (Status: Available for Tiendanube): Notifies search engines participating in IndexNow of new, updated or deleted URLs. [How IndexNow Connect works](https://cookingmetrics.com/product/indexnow-connect/). - **[AI SEO Analysis](https://cookingmetrics.com/product/ai-seo-analysis/)** (Status: In development): *Coming soon.* An automated audit of how AI assistants and search engines access and use your content. [Learn about AI SEO Analysis](https://cookingmetrics.com/product/ai-seo-analysis/). **Let's talk about your measurement:** [Book a call](https://cookingmetrics.com/about-martin-garay/) · [CookingMetrics](https://cookingmetrics.com/)