AI visibility can be measured, but treating it as the equivalent of an SEO position creates a false sense of precision. In ordinary rank tracking, the thing being observed is relatively clear: a URL has an ordinal place in a search result under specified conditions. In AI search, what we usually call visibility is calculated from a sample of generated answers. Those answers may change when the same prompt is repeated, and the sources behind them may change even more. Recent studies have measured substantial run-to-run and day-to-day variation, while official documentation from Google, OpenAI and Yandex shows that the products themselves differ in crawling, retrieval and answer-generation pipelines.
For practitioners, this does not make AI visibility useless. It changes what the percentage means. I find it more useful to read a figure such as 42% as a sample statistic: under a defined set of prompts, platforms and measurement rules, the brand appeared in 42% of the observed answers. That number becomes more useful when you keep the methodology stable, repeat measurements, preserve platform-level data instead of hiding it inside one average, and read visibility together with Share of Voice, answer sentiment, attributable AI referral traffic, external publication reach and branded traffic. This is especially important in the cases that cause the most questions in real projects: a brand appears in ChatGPT but not in another system, or the reported visibility remains almost flat for a month and then jumps by 20–30 percentage points several weeks later.
The problem starts with the word “position”
If you come to GEO from SEO, the temptation to reuse the old mental model is strong. SEO has trained us to turn a complicated search system into a simple observable variable: a page is fifth for a query, then ninth, then third. Every experienced SEO knows that a rank is not a law of nature. Location, device, search features, index changes and other factors can affect the result. Still, the object being measured is well defined enough for rank tracking to work: we observe the ordinal position of a document in a ranked result set.
AI search changes the object itself. A generated answer is not simply a SERP with the blue links removed. It is the output of a process that may include query interpretation, search activation, retrieval, source selection, context construction and generation.
Google explicitly says that AI Overviews and AI Mode may use “query fan-out”, issuing multiple related searches across subtopics and data sources, and that the two products may use different models and techniques, so their responses and supporting links can differ.
Suppose you ask an AI system for project-management tools suitable for a small construction company. One answer may recommend six products, another four. A brand might be mentioned first but receive one sentence, while another appears later and gets half a paragraph. A third answer may use a company’s documentation as a factual source without recommending the company at all. You can certainly define “average recommendation position” and track it, but you have created an analytical convention rather than discovered an exact analogue of a Google rank. This distinction was already implicit in the original GEO paper by Aggarwal and colleagues, first released in 2023 and accepted at KDD 2024. The authors did not reduce generative visibility to a single rank; they introduced a framework for defining and optimizing visibility inside generated answers and a benchmark for evaluating it. Their experiments showed that content interventions could change visibility within the experimental setup, in some cases by as much as 40%.
The field has become more explicit about measurement since then. The 2026 paper Quantifying Uncertainty in AI Visibility formalizes citation count, citation share and citation prevalence as statistics computed from a finite sample of responses rather than fixed properties of an answer engine. A second 2026 study, Don’t Measure Once, investigates the stability of brands and sources under repeated queries. A large production-data preprint published in June 2026 separates mention rate, ordinal position, sentiment and Share of Voice instead of collapsing them into one score. That progression matters because it shifts the central question from “What is our AI rank?” to “Which property of AI visibility are we trying to estimate, and from how many observations?”
A compact way to see the difference is this:
What is being measured | SEO rank tracking | AI visibility tracking |
|---|---|---|
Basic observation | URL at an ordinal position | Brand, domain or source inside a generated response |
Natural unit | query × SERP × time | prompt × platform × run × time |
Typical headline metric | rank, share of top positions, search visibility | mention rate, citation prevalence, Share of Voice |
Main statistical problem | changing SERPs and measurement conditions | changing retrieval plus stochastic answer generation plus finite-sample uncertainty |
The last row is important because it prevents an equally misleading opposite claim. SEO is not perfectly deterministic, and AI search is not pure randomness. The practical difference is that repeated generation introduces an additional source of variability, while the metric usually called AI visibility is already an aggregation across many answers. Sielinski’s 2026 paper makes this distinction explicitly: system-level stochasticity is the variability produced by the answer engine itself, while measurement uncertainty comes from trying to estimate an underlying rate from a finite sample. Both contribute to the percentage on the dashboard.
What happens when you ask the same question again
The most useful evidence for practitioners comes from studies that simply repeated the measurement.
Don’t Measure Once collected daily results for four AI search engines across four commercial verticals for roughly 45 days. The researchers used eight prompts per campaign and tested ChatGPT, Gemini, Google AI Mode and Perplexity. They compared both cited sources and brands mentioned in the generated text. Across more than 4,000 consecutive-day source comparisons, the average Jaccard similarity of cited-source sets was only 0.34 to 0.42, depending on the campaign. A Jaccard similarity of 0.35 means that only about 35% of the union of sources appeared in both observations. Brand mentions were more stable than citations, but still far from fixed: in the three verticals that passed the study’s brand-detection quality threshold, average day-to-day brand Jaccard similarity ranged from 0.45 to 0.59.
Day-to-day change could be blamed on fresh pages, updated indexes or platform changes, so the authors also repeated identical prompts up to ten times within a 24-hour window. That is the more interesting result for anyone running an AI visibility dashboard. Across the campaigns, average source-set similarity remained only 0.32 to 0.43. Depending on the vertical, brand similarity was approximately 0.33 to 0.48. The authors conclude that a substantial part of the observed instability cannot be explained simply by the web changing overnight; variation remains when the time window is compressed. The experiment has boundaries, of course. It covers four Swiss-German commercial verticals, a limited prompt set and four products at a particular point in 2026. It does not establish that every industry or language will show exactly the same amount of variation. It does show that a single answer is a poor basis for treating visibility as a fixed quantity.
Quantifying Uncertainty in AI Visibility approaches the same issue from a statistical direction. It studies Perplexity Search, OpenAI SearchGPT and Gemini using repeated samples, including daily collection and ten-minute sampling. Instead of asking only whether answers differ, the paper asks how uncertain the resulting visibility estimates are. The author reports that confidence-interval widths of roughly 5–7 percentage points were common for SearchGPT citation-share estimates in the studied data, and that apparent differences between domains could fall within the measurement noise. The paper therefore argues that citation visibility should be reported as an estimate with uncertainty rather than a point value that appears exact to the decimal place. This is a preprint, not a settled industry standard, and its scope is narrower than all of AI search, but the statistical argument is straightforward: if the process you sample varies, a finite sample should carry an estimate of that variation.
This helps explain something I have seen in real reporting. A brand’s ChatGPT visibility can be roughly unchanged between the beginning and end of a month, then increase by 20–30 percentage points about a month and a half later. The observation is real in the limited sense that the second batch of answers contains more mentions than the first. What you cannot infer from those two measurements alone is why. The underlying brand presence may genuinely have strengthened; the mix of sources available to the system may have changed; the platform may have changed; competitors may have moved; or part of the difference may come from sampling a variable response process at two different moments. The right comparison is therefore not “30% versus 55%, so our work added 25 points.” It is “30% versus 55%; is the second value outside the range we normally observe under comparable conditions, and does that shift persist?”
This is also why sample design matters as much as the percentage itself. Imagine measuring 20 prompts once. Eight mentions give you 40% visibility. Now imagine 200 prompts, grouped by intent, with repeated runs and a stable platform configuration. The dashboard may show 40% in both cases, yet the evidential weight is very different. The 2026 production-data paper makes a related point from another angle: it stores observations at the level of a (brand, prompt, platform, run) tuple and reports mention visibility, position, sentiment and Share of Voice separately because these signals behave differently. Its dataset contains more than 100,000 prompt responses across more than 100 brands, and it finds that sentiment is much less stable than the binary fact of whether a brand was mentioned. That finding should not be turned into a universal constant, because the work is a preprint based on one platform’s production data, but it is a useful warning against putting a mention, its position and its tone into a single opaque score.
Why a brand appears in one AI system and disappears in another
The client question I hear most often is simple: “Why are we present in ChatGPT but not in another neural network?” I would correct one word before answering it. In practical visibility work, we usually compare products or answer systems, not just foundation models. What the user sees is produced by a stack that can include a crawler, an index or third-party search provider, retrieval logic, ranking or reranking, a language model and product-specific rules about when web search is triggered. Two products can use excellent models and still expose different brands because the information reaching those models is different. Official documentation from the major platforms makes this visible, even though none of them publishes the full ranking logic.
Google states that a page must be indexed and eligible to appear in regular Google Search with a snippet to be eligible as a supporting link in AI Overviews or AI Mode. It also says that those products may issue multiple related searches and may use different models and techniques, which means their answer and link sets can vary even inside the same Google ecosystem. Yandex describes Search with Alice as a sequence in which its algorithms find content matching the query, a neural network analyzes the content and selects relevant pieces, and the system compiles them into an answer with source links; the pages used for this process come from the Yandex index. Yandex’s Webmaster documentation goes further and explicitly states that Alice generates a new answer from the current search results each time, so both the answer and its sources can differ for the same query at different times.
OpenAI exposes a different part of the plumbing. Its publisher FAQ says that any public website can appear in ChatGPT search and recommends allowing OAI-SearchBot so that content can be discovered, surfaced and cited. OpenAI also adds utm_source=chatgpt.com to referral URLs from ChatGPT search, which makes some downstream visits measurable. Allowing the crawler, however, is an eligibility condition rather than a promise that a page or brand will be selected for an answer.
Once you think in terms of the whole pipeline, disagreement across platforms stops looking mysterious. A page can be available to Google but poorly represented in another system’s retrieval layer. Two systems can retrieve the same page and assign it different importance.
They can retrieve different third-party reviews about the same company. They can both have enough evidence to discuss the brand, but one answer may contain six recommendations while another stops after four. Even source identity and brand identity are not the same measurement: a system may cite an article that discusses your company without linking to your own domain, or it may mention your brand while citing a publisher, marketplace or competitor comparison page. The production-data study from 2026 found substantial platform differences in visibility and source composition, while Don’t Measure Once found that source sets were less stable than brand mentions. These results come from different datasets, but both point in the same methodological direction: “present in AI” has several layers that should not be collapsed into a yes/no property of the brand.
This distinction changes how I answer the client question. If a brand is visible in ChatGPT and absent elsewhere, I do not begin with “we need to improve the score in the second model.” I first ask where the difference enters the pipeline. Is the brand absent from generic discovery prompts but present when explicitly named? Is its site cited, or are third-party publications doing the work? Is the discrepancy limited to one intent cluster? Are competitors different too? Does the same gap survive repeated measurements? Only after that does it make sense to decide whether we are looking at a content problem, a discoverability problem, a source-ecosystem problem or simply normal platform disagreement.
A measurement stack that is more useful than one visibility number
In my own work, I therefore track AI visibility as part of a wider set of signals.
The first layer is visibility on the promoted prompt set: how often the brand appears in the questions we have deliberately selected to represent relevant discovery, comparison and decision scenarios. Next comes Share of Voice, because a 40% mention rate means something different when the main competitors are at 15% than when they are at 75%. I also track the sentiment of AI answers, but separately rather than blending it into the visibility score; the 2026 large-scale study provides empirical support for that caution by finding sentiment far more unstable than mention status in its dataset. Alongside these AI-output metrics, I look at attributable traffic from AI services and relevant publications, the reach and views of external publications, and branded site traffic. Google and Yandex generative search answers require separate treatment from ordinary external AI referrals because they are embedded in the search products themselves rather than behaving like independent referring websites. Google, for example, says that AI Overviews and AI Mode activity is included in Search Console’s web-search reporting rather than exposed as a separate external referral channel.
The reason for combining these signals is not that every one of them proves GEO impact. Quite the opposite. Each has a different blind spot. Prompt visibility depends on your prompt sample. Share of Voice depends on which competitors you define and how you count mentions. Sentiment depends on classification and can fluctuate even when mention status is stable. Referral traffic records only visits that can be attributed technically. Publication reach tells you that a piece of information has an audience but does not prove an AI system consumed it. Branded traffic is influenced by every other marketing channel as well, so a rise in branded searches cannot simply be credited to ChatGPT. Treating the collection as a measurement stack is useful because several imperfect signals can tell you whether a change is confined to one dashboard metric or is accompanied by changes elsewhere. The word “accompanied” matters here: without an experimental design, correlation across these metrics still does not establish causation.
The traffic layer deserves particular caution. Ahrefs examined an anonymized sample of 3,000 websites in 2025 and found that 63% received at least one visit attributable to an AI chatbot, while AI referrals accounted for only a small proportion of average site traffic in that dataset. More importantly for measurement, the researchers demonstrated that some visits from AI products can arrive in analytics as “Direct” when referrer information is not passed. In their small manual checks, Copilot, Mistral and Jasper visits appeared that way. The study is commercial industry research rather than peer-reviewed science, and the product landscape has changed since early 2025, so its platform shares should not be treated as current market constants. The attribution limitation remains relevant: absence from an AI-referral report is not equivalent to absence of AI influence.
That pipeline also gives you a cleaner definition of the headline number. If a report says “AI Visibility: 42%”, I would want the methodology to make clear that this means something like: the brand appeared in 42% of the observed answers to the monitored prompt set under the stated platform and measurement conditions during the selected period. Sielinski’s statistical framing is useful here because it treats that 42% as an estimate of an underlying response distribution. Yandex has independently adopted a conceptually similar frequency-based approach in its Webmaster product: its Alice AI Share of Voice is calculated as the proportion of relevant queries with a site mention among queries for which Alice produced an answer. The implementations are not identical, but both are based on frequency of observed presence, not a fictional universal “AI rank”.
How I would read a 20–30 point jump
Suppose you open the monthly report and visibility has moved from 31% to 55%. The interesting part is not calculating the 24 percentage-point increase; any spreadsheet can do that.
The analytical work begins when you ask whether the two numbers are comparable. I would check whether the prompt set, intent distribution, brand aliases, language, geography and platform definitions remained unchanged. I would keep each platform separate before looking at a cross-platform aggregate, because Google’s own documentation confirms that even AI Mode and AI Overviews can differ in models, techniques, answers and links. I would then examine the underlying runs: did a broad set of prompts move, or did ten prompts inside one category suddenly change? Did competitors move at the same time? Did the sources change? Most importantly, does the higher level survive the next observations? The repeated-measurement research gives good reason to make persistence part of the definition of a meaningful change rather than something checked after a conclusion has already been written.
This is where I think GEO reporting can borrow from scientific measurement more profitably than from classic rank tracking. You do not need to turn a client report into a statistics paper or attach a bootstrap distribution to every slide. You do need to distinguish signal from the precision supplied by the interface. If the system displays 47.6%, that decimal does not prove that the underlying phenomenon is known to one tenth of a percentage point. With enough data, confidence intervals or repeated-run ranges are preferable. With smaller programmes, even showing min, max and average across repeated observations is more informative than pretending one run is canonical. Sielinski’s study provides the formal statistical argument for this, while Don’t Measure Once provides the empirical demonstration that repeated results can differ substantially even over short intervals.
The same discipline helps with attribution. If visibility rises after a new publication campaign, I would describe the sequence honestly: visibility increased after the campaign, the change persisted across several measurements, Share of Voice also improved, the new publication began appearing among cited or relevant external sources, and branded demand moved in the same period. That is a much stronger story than “the articles caused a 24-point GEO lift”, but it is still observational evidence unless the campaign was designed with a meaningful control. The foundational GEO study did perform controlled content interventions inside its benchmark and showed that source modifications could change measured visibility in that environment. The 2026 GEO at Scale paper, by contrast, explicitly distinguishes its production observations from causal evidence and proposes follow-up protocols for testing interventions. Keeping those two levels of evidence separate is important for a field where dashboards are being commercialized faster than measurement standards are stabilizing.
So, can AI visibility be measured like SEO positions? I would say that the useful part of the SEO tradition survives: define your query universe, control the measurement conditions, compare competitors and follow the series over time. The part that should not survive is the assumption that every percentage is a stable coordinate waiting to be read from the system. AI visibility is closer to estimating the frequency with which a brand appears across a specified population of generated answers. Once you treat it that way, the awkward practical questions become easier to answer. A brand can be visible in one product and not another because the products use different discovery and retrieval pipelines. A 20–30 percentage-point jump can be meaningful, but only after you establish that it exceeds ordinary variability and persists. A referral report can miss part of AI-driven discovery. Sentiment can move differently from mention frequency. None of these limitations makes measurement pointless; they simply tell you what kind of measurement you are actually doing.
