Research reviewed 1 October 2026
How should AI visibility be measured when generative answers can change from run to run?
AI visibility should be measured as a distribution of repeated, time-stamped observations across a defined prompt set and a defined set of AI engines, then reported as rates with uncertainty—not as a single screenshot, a single prompt result or a permanent “rank”. A defensible measurement design separates prompt breadth, run-to-run repetition, engine differences and time effects so the reader can see what was observed, how often it happened, and how uncertain the estimate remains. 2026 GEO measurement preprintACL 2026 · generative-search stability
That distinction matters because generative systems are probabilistic and operationally dynamic. In a 2026 field study covering 30,504 prompt-engine responses, mention status changed between consecutive observations 9.7% of the time, while consecutive citation-source sets had a mean Jaccard overlap of only 0.396. Separate academic work has shown that even semantically small prompt changes can produce material output variation, and peer-reviewed research on ChatGPT found that a temperature setting of zero does not guarantee identical outputs. 30,504-response field studyACL 2026 · prompt sensitivityACM TOSEM · non-determinism
The practical rule: measure frequency, not anecdotes; measure across prompts, not one keyword; measure each engine separately; repeat over time; and publish the uncertainty around every important rate.
TL;DR
1. Define the unit
Treat one observation as one prompt × one engine × one run × one timestamp under documented conditions.
2. Repeat the cell
One answer shows what happened once. Repeated runs estimate how often the outcome occurs.
3. Cover the intent space
A large number of repeats cannot compensate for a prompt set that represents only one narrow wording or buyer intent.
4. Report intervals
For mention/citation rates, use an appropriate binomial interval; for composite metrics, use resampling or hierarchical methods where warranted.
5. Keep engines separate
Cross-engine source overlap is low enough that a pooled average can hide major platform differences.
6. Timestamp every claim
AI visibility is a measured state over a stated window, not a permanent property of a brand or page.
What does “AI visibility” mean in a measurement context?
AI visibility is the observed frequency and prominence with which a defined brand, entity, page or domain appears in relevant AI-generated answers under a specified prompt set, engine set, market context and time window. It is not one universal number unless the methodology explicitly defines how several different outcomes are combined. 2026 GEO survey · 45 studies
For measurement, it is safer to treat AI visibility as a vector of observable outcomes: brand mention, recommendation, own-domain citation, third-party citation, answer position or prominence, prompt coverage, competitor share of voice, source-set stability and—where the research question requires it—factual absorption or citation fidelity. A 2026 critical survey of 45 GEO studies similarly argues that Generative Engine Optimisation operates across a “stochastic, partially observable pipeline” rather than one fixed ranking task.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Metric | What it answers | Recommended reporting form | Main uncertainty |
|---|---|---|---|
| Mention rate | How often is the brand named? | Count / valid runs + percentage + interval | Sampling, prompt mix, drift |
| Own-domain citation rate | How often is the website cited as evidence? | Runs with ≥1 own-domain citation / valid runs | Citation panel volatility and source selection |
| Recommendation rate | How often is the entity positively selected, not merely mentioned? | Predefined binary/ordinal coding rubric | Semantic coding and evaluator consistency |
| Share of voice | How much of the competitive mention pool belongs to the brand? | Percentage plus the exact denominator definition | Denominator choice can dominate the result |
| Source stability | Do repeated answers cite the same evidence? | Jaccard overlap / recurring-domain rate | Retrieval and source-set volatility |
Why can a single AI answer mislead you?
A single AI answer can mislead because it cannot distinguish persistent visibility from a one-off retrieval outcome. If a brand appears once, you do not yet know whether its underlying appearance probability is high or low; if it fails to appear once, you do not know whether it is truly absent or was simply missed on that run. 2026 GEO measurement preprint30,504-response field study
The logic is easiest to see with a hypothetical brand that has a genuine 20% probability of being mentioned for one fixed prompt-engine condition. The chance of seeing no mention in one run is 80%. Even after five independent runs, the chance of seeing no mention is 32.8%; after ten runs it is 10.7%; after twenty runs it falls to about 1.15%. The calculation is P(no mention in n runs) = (1 − p)n. This is not a claim about any specific AI engine; it shows why low-frequency visibility is easy to miss with one-shot testing.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Repeated observations | Chance of missing a true 20% mention-rate signal | Interpretation |
|---|---|---|
| 1 | 80.0% | A negative result tells you very little. |
| 5 | 32.8% | Useful as exploration, still easy to miss intermittent visibility. |
| 10 | 10.7% | Much better detection, but not precise estimation by itself. |
| 20 | 1.15% | Intermittent visibility is unlikely to be completely missed under the independence assumption. |
Real AI observations may not be independent: runs close together can share the same index state, model configuration, retrieval cache or news cycle. NIST warns that when repeated measurements are autocorrelated, uncertainty calculations that assume independence are not appropriate. That is why spacing observations over time matters as much as simply increasing the raw run count. NIST · autocorrelated uncertainty
What is the correct unit of observation for AI visibility?
The safest unit is one valid response from one specified prompt on one specified AI engine at one timestamp, under documented market and session conditions. In compact form: prompt × engine × run × time. Where location, language, login state, search mode, model variant or device context can affect the answer, those conditions should be recorded as additional strata rather than silently mixed. AirPulse · measurement contract
This “cell” approach prevents a common measurement error: treating 100 heterogeneous outputs as though they were 100 interchangeable trials. Ten prompts asked ten times are not statistically identical to one prompt asked 100 times. The first design contains information about prompt breadth and repetition; the second contains only repetition of one wording.
Prompt
What information need, buyer intent and wording is being tested?
Engine
Which answer system produced the result, and in which mode?
Run
Which repeated observation is this within the cell?
Time
When was it measured, and over what reporting window?
Diagram: the four dimensions of a defensible AI visibility sample
A large raw response count is not enough if one of these dimensions is missing.
↔ Mobile note: swipe or scroll horizontally if needed to view the full measurement diagram.
Discovery · comparison · use case · pricing · evidence
Repeated runs estimate how often the outcome recurs
ChatGPT · Claude · Gemini · AI Mode · Perplexity · Copilot
Spread observations across the stated reporting window
■ repetition
■ engine separation
■ time coverage
How large should an AI visibility sample be?
There is no single correct AI visibility sample size: it depends on the metric, the precision you need, the expected visibility rate, the number of prompts and engines, and how independent the observations really are. For a simple binary metric such as “brand mentioned: yes/no”, classical proportion mathematics gives a useful precision benchmark, but correlated runs and changing AI systems mean real-world designs often need more observations—or more careful modelling—than the textbook number suggests. NIST · sample-size design
What does sample size actually control?
Sample size primarily controls sampling uncertainty: more independent observations narrow the range of plausible values around an estimated mention or citation rate. At a worst-case true proportion near 50%, an approximate 95% margin of error is about ±31 percentage points with 10 independent observations, ±19.6 points with 25, ±13.9 with 50, ±9.8 with 100 and ±4.9 with 400. NIST · sample-size designNIST · proportion intervals
Bar chart: sample size versus approximate worst-case 95% margin of error
Assumes independent Bernoulli observations and p≈0.50. This is a statistical reference, not a universal AI measurement guarantee.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
n = 10
n = 25
n = 50
n = 100
n = 400
■ moderate-high
■ moderate
■ around ±10 pp
■ around ±5 pp
How many independent observations are needed for a chosen margin of error?
At the conservative p=0.50 case, about 25 independent observations correspond to a ±20-point 95% margin, 43 to ±15 points, 97 to ±10 points and 385 to ±5 points using the standard planning approximation. These are planning numbers for one binary proportion, not prescriptions for how many prompts a GEO programme must track. NIST · sample-size design
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Target 95% margin of error | Approx. independent observations at p=0.50 | What that means for AI visibility |
|---|---|---|
| ±20 pp | 25 | Coarse directional estimate. |
| ±15 pp | 43 | Still too broad for small competitor differences. |
| ±10 pp | 97 | Reasonable precision for a single simple rate if independence is plausible. |
| ±5 pp | 385 | High precision; expensive and still vulnerable to clustering or drift. |
Why is 1,000 repetitions of one prompt still a bad sample?
Because repetition measures stability within a prompt-engine condition, while prompt breadth measures whether the tested condition represents the real question space. A thousand repetitions of “best X in the UK” cannot tell you how the brand performs for comparisons, alternatives, pricing, use cases, problem-aware questions or buyer constraints that were never sampled. More runs reduce one kind of uncertainty while leaving coverage bias untouched.
How often should each AI visibility prompt be repeated?
Each prompt-engine cell should be repeated enough to estimate its stability, and the repetitions should be distributed across time rather than fired as one tight burst. There is no universal minimum, but a 2026 30,504-response field study proposed at least seven runs per prompt and engine for brand-presence measurement and eight for source-coverage measurement, spread across two to four weeks, as a practical floor rather than a guarantee of fixed precision. 2026 GEO measurement preprint30,504-response field study
What does repetition protect against?
Repetition protects against mistaking stochastic retrieval, answer-generation variation or short-lived source selection for a stable visibility pattern. In the AirPulse dataset, 35.9% of prompt-engine cells observed at least three times changed from mentioned to not mentioned—or the reverse—at least once during the 30-day window. That means “we appeared” and “we did not appear” can both be true observations for the same cell at different moments. 30,504-response field study
Stacked bars: what repeated AI measurements reveal
Observed in AirPulse’s frozen 30-day 2026 study; these are dataset-specific results, not universal constants for every engine or category.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
Consecutive brand-mention status
Consecutive citation-source sets: Jaccard intersection over union
■ mention-state flip
■ source-set intersection
■ source-set non-overlap
How should repetitions be spread over time?
Spread repeats across the reporting window so the sample captures day-to-day variability, index refreshes and platform changes instead of repeatedly sampling one moment. If the reporting claim is “monthly visibility”, a burst of 30 runs on one morning is a weak representation of a month. Daily or systematically spaced observations are usually more defensible, and any major platform update should be annotated rather than silently folded into the series. ACL 2026 · generative-search stability30,504-response field study
If observations show autocorrelation—today’s result strongly predicts tomorrow’s because the underlying system state has not changed—then the effective amount of independent information is smaller than the raw run count. NIST’s statistical guidance is explicit that conventional uncertainty estimates can become misleading when the randomness/independence assumption fails.
What is a practical sampling architecture for an AI visibility benchmark?
A practical benchmark uses a balanced matrix: a predefined set of commercially relevant prompts, every target engine measured separately, repeated observations distributed across the reporting window, and the same rules preserved from period to period. The objective is not to maximise the raw response count; it is to allocate observations so they answer the actual visibility question without over-sampling one prompt, platform or moment. NIST · stratification & randomisationONS · sample design & weighting
For example, an illustrative operational design might use 20 fixed prompts × 5 engines × 8 time-spaced repeats = 800 raw responses. That does not mean the study has 800 fully independent observations. The 800 responses are clustered inside 100 prompt-engine cells, and each cell shares wording, business context and platform characteristics. The correct analysis respects that structure.
How should those 800 illustrative observations be interpreted?
Interpret them at several levels at once: per run, per prompt-engine cell, per engine, per intent category and across the whole reporting window. The run level captures volatility; the cell level estimates stability for one question on one engine; the engine level identifies platform-specific strengths; the intent level shows where buyer-question coverage is strong or weak; the full window supports trend reporting.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Layer | Illustrative count | Question it answers | Statistical caution |
|---|---|---|---|
| Raw responses | 800 | What happened on each observed run? | Not 800 independent market observations. |
| Prompt-engine cells | 100 | How stable is each specific question on each platform? | Eight repeats per cell still give broad cell-level intervals. |
| Engine strata | 5 | Which systems consistently surface or cite the brand? | Do not hide zeroes behind a cross-engine average. |
| Prompt set | 20 | How broad is visibility across the intended commercial question space? | Prompt representativeness is a design judgement, not solved by more repeats. |
Should a small business wait until it can afford hundreds of AI responses?
No: use a staged design and be proportionate about what each stage can prove. A small diagnostic sample can identify obvious invisibility or competitor dominance; a repeated operational benchmark can monitor direction; a large, carefully structured study is needed when decisions depend on distinguishing small percentage-point differences. The mistake is not using a small sample—the mistake is making a large claim from it.
How should uncertainty be calculated and reported for AI visibility?
Uncertainty should be reported with the metric itself: show the numerator, denominator, point estimate, interval method and sampling design, then disclose any dependence or coverage limitations the interval does not capture. A rate without its sample size is incomplete; a 50% mention rate from 2 runs and a 50% mention rate from 200 runs are not equally informative. NIST · proportion intervals
Which confidence interval should be used for mention and citation rates?
For binary outcomes, use a binomial proportion interval such as Wilson, Agresti–Coull or an exact binomial interval rather than relying blindly on the simple symmetric normal/Wald interval. NIST notes that the common normal approximation may be inaccurate when the sample or number of successes/failures is small; its statistical references provide exact and Agresti–Coull alternatives. NIST · proportion intervals
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Observed result | Point estimate | Approx. 95% Wilson interval | What a careful reader should conclude |
|---|---|---|---|
| 3 mentions / 5 runs | 60% | 23.1%–88.2% | Very broad uncertainty despite the impressive-looking 60% point estimate. |
| 5 / 10 | 50% | 23.7%–76.3% | Directional, not precise. |
| 15 / 30 | 50% | 33.2%–66.8% | Useful trend signal, still not enough for small differences. |
| 50 / 100 | 50% | 40.4%–59.6% | Much tighter under an independence assumption. |
When are simple binomial confidence intervals too optimistic?
They are too optimistic when observations are clustered or dependent—for example, repeated runs of the same prompt, several prompts expressing the same underlying intent, or measurements collected under one unchanged engine state. A binomial interval treats trials as if each contributes independent information. When that assumption is materially false, the nominal interval can understate real uncertainty. NIST · autocorrelated uncertainty
For serious comparative studies, model the structure you actually sampled. Options include cluster-aware bootstrap resampling at the prompt level, block bootstrap over time, hierarchical/mixed-effects models with prompt and engine effects, or a beta-binomial model when there is over-dispersion beyond simple Bernoulli variation. The method should be chosen before looking for a desired result.
“The randomness assumption is the most critical but the least tested.”
NIST/SEMATECH e-Handbook of Statistical Methods.
When should bootstrap methods be used?
Bootstrap methods are useful when the reported statistic is an aggregate whose sampling distribution is awkward—for example weighted share of voice, average prominence, source diversity or a multi-prompt composite. NIST research on information-retrieval measures found both standard and bootstrap confidence intervals useful, with bootstrap percentile intervals providing an equivalent practical option in the studied settings. NIST · bootstrap uncertainty
How should the prompt sample be designed so it represents real AI visibility?
The prompt sample should represent the decision-relevant question space, not merely the phrases most likely to make the brand look visible. Build the prompt set before reviewing results, document inclusion rules, preserve a fixed core for longitudinal comparability, and add clearly labelled exploratory prompts rather than quietly changing the benchmark when performance weakens. NIST · representative samplingONS · sample design & weighting
How many prompt types should be included?
Include enough distinct intent categories to cover the buyer questions the measurement claim is supposed to represent. A practical commercial set often needs several classes—category discovery, “best” or shortlist queries, comparison, alternatives, use-case fit, pricing/cost, problem-solution, local/market constraints and evidence/proof questions—but the correct mix depends on the business and should be justified rather than copied mechanically.
A useful distinction is between fixed benchmark prompts, which support month-to-month trend comparison, and coverage prompts, which test whether the benchmark generalises to adjacent ways users ask. The fixed set protects comparability; the broader set protects against overfitting to a small handful of phrasings.
Should paraphrases of the same intent be included?
Yes, when the goal is to estimate robustness to natural wording variation—but paraphrases should be treated as related observations, not as dozens of fully independent new intents. Peer-reviewed work shows that wording can affect model behaviour, while also warning that measured prompt sensitivity depends on how outputs are evaluated. EMNLP 2025 tested seven LLMs across six benchmarks and twelve prompt templates and found that some apparent sensitivity fell when semantically equivalent answers were judged more flexibly. ACL 2026 separately analysed meaning-preserving prompt variants and found that prompt templates can materially influence model logits, supporting paraphrase coverage as a genuine measurement dimension rather than a cosmetic wording exercise. EMNLP 2025 · evaluation sensitivityACL 2026 · prompt sensitivity
Should prompts be weighted?
Weight prompts only when you can defend the weights with a business-relevant basis, such as observed query demand, funnel importance or a pre-registered decision rule. Otherwise, equal weighting is easier to audit. If weighting is used, publish both the weighted and unweighted result so the reader can see how much the weighting scheme changes the conclusion.
Should ChatGPT, Claude, Gemini, Google AI Mode, Perplexity and Microsoft Copilot be pooled into one AI visibility score?
Not by default: report each engine separately first, then provide an aggregate only if the weighting rule is explicit and useful for the decision being made. Different systems can use different retrieval sources, citation interfaces, web-search triggers, freshness mechanisms and answer-generation policies, so a pooled average can conceal the exact platform where visibility is strong or weak. ACL 2026 · cross-engine variation
The scale of cross-engine difference is not merely theoretical. A July 2026 industry study of 161,286 prompts across ChatGPT, Gemini, Perplexity and Google AI Overviews reported that only 3.8% of sources were shared by all four engines in the comparison set, with pairwise source overlap around 17%. Treat that as vendor research rather than a universal law, but it strongly supports engine-level reporting. 161,286-prompt vendor study
A 2026 arXiv preprint analysing 100K+ prompt responses across 100+ brands also found that sentiment was substantially less stable than simple brand mention status, flipping about 6.7 times more often in its dataset. That is a useful warning against compressing mention, sentiment, citation and recommendation into one opaque score. 2026 preprint · 100K+ responses
What counts as a valid AI visibility observation, and what should be excluded?
A valid observation is a completed response produced under the study’s predefined conditions; a technical failure is missing data, while a completed answer in which the brand is absent is a legitimate negative result and must remain in the denominator. Confusing those two cases is one of the easiest ways to inflate visibility.
How should timeouts, refusals and unavailable citation panels be treated?
Predefine status codes and report them separately instead of silently deleting inconvenient runs. A timeout, platform error or blocked request is usually an invalid technical observation; a valid answer that searches the web but does not mention the brand is a zero for mention visibility; an answer that mentions the brand but provides no own-domain link is a mention success and an own-domain citation failure under the relevant citation metric.
Citation-capable and non-citation-capable surfaces require particular care. If one engine/mode routinely provides linked sources and another mode does not expose sources, a single pooled “citation rate” can become conceptually incoherent. Either restrict the citation metric to comparable surfaces or publish platform-specific definitions.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Observed event | Mention denominator? | Citation denominator? | Recommended status |
|---|---|---|---|
| Completed answer, brand absent | Yes | Yes, if surface is citation-comparable | Valid negative observation |
| Completed answer, brand mentioned, no own-domain citation | Yes | Yes | Mention success; own-domain citation failure |
| Completed answer, brand + own-domain citation | Yes | Yes | Positive mention and citation observation |
| Timeout / platform error / no completed answer | Usually no | Usually no | Missing technical observation; report count |
Why should missing-response rates be published?
Because missingness can be systematic rather than random. If one engine fails more often on long prompts, one region has intermittent access, or a tool retries failures until it gets a completed result without disclosing the retries, the final dataset may no longer represent the originally scheduled observations. Publish scheduled runs, valid runs, invalid runs, retry rules and the reasons for exclusion.
Which AI visibility metrics are worth reporting?
A defensible report should prioritise transparent, reconstructable metrics: mention rate, recommendation rate, own-domain citation rate, citation count, prompt coverage, prominence, share of voice with a stated denominator, and source stability. Composite “visibility scores” can be useful for dashboards, but they should never replace the underlying counts and formulas.
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Metric | Formula / rule | Report with | Do not confuse it with |
|---|---|---|---|
| Mention rate | Runs with brand mention ÷ valid runs | n, %, confidence interval, engine | Citation rate |
| Own-domain citation rate | Runs with ≥1 citation to owned domain ÷ valid runs | citation count + run rate | Brand mention |
| Recommendation rate | Runs meeting a pre-defined “recommended/selected” coding rule ÷ valid runs | coding rubric + reviewer agreement | Any mention |
| Prompt coverage | Unique prompts with ≥1 qualifying appearance ÷ tracked prompts | prompt count + intent distribution | Run-level mention rate |
| Share of voice | Brand numerator ÷ explicitly defined competitive denominator | denominator definition + competitor set | Market share |
| Source stability | Jaccard = |A∩B| ÷ |A∪B| for source sets | pairing rule + time interval | Citation volume |
Why must share of voice disclose its denominator?
Because the denominator is part of the metric definition, and changing it can radically change the reported percentage even when the underlying observations stay identical. In an August 2026 methodology analysis, one brand’s share of voice calculated from the same dataset ranged from 0.94% to 40.0% across four defensible denominator choices—a 42.5× spread. The correct response is not to declare one denominator universally right; it is to state the business question, define the denominator before analysis and publish it with the result. Aug 2026 · SoV denominator study
How do you tell whether AI visibility actually improved rather than merely fluctuated?
Call an improvement credible only when the comparison uses the same core prompts, engines, market conditions and measurement rules, and the observed change is large enough relative to sampling and time variability to survive an uncertainty-aware comparison. A movement from 23% to 27% is not automatically meaningful if each estimate is based on a small, noisy sample. 2026 GEO survey · 45 studiesNIST · stratification & randomisation
The strongest practical design is paired and longitudinal: compare the same prompt-engine cells before and after the intervention, preserve a control set where possible, and retain raw answers so changes can be audited. If a platform changes its model, search product or source interface during the test, annotate the event; otherwise the analysis can attribute platform drift to the optimisation work.
What comparison method is strongest for a before-and-after GEO test?
Use a paired comparison at the prompt-engine level, with repeated observations in both windows, and estimate the distribution of the change rather than comparing two headline percentages alone. For complex programmes, a cluster bootstrap or mixed-effects model can preserve prompt and engine structure. If a control brand or control pages are available, a difference-in-differences design can help separate broad engine changes from the intervention—provided its assumptions are reasonable. NIST · bootstrap uncertainty2026 GEO survey · 45 studies
This is also why Generative Engine Optimisation should be evaluated with more than citations alone. A page can become easier to retrieve but not more frequently cited; a brand can be mentioned more often without its domain receiving citations; source attribution can improve while recommendation rate stays flat. These are related outcomes, not interchangeable ones.
What does a defensible AI visibility report look like?
A defensible report makes reconstruction possible: another analyst should be able to understand exactly what was tested, when, where, how often and how every headline number was calculated. The raw answer evidence can be private if necessary, but the methodological contract should not be. AirPulse · measurement contract
↔ Mobile note: swipe or scroll horizontally to view the full table or chart.
| Report field | Minimum disclosure | Why it matters |
|---|---|---|
| Research question | What decision the study is intended to support | Stops broad claims from outrunning narrow tests. |
| Prompt frame | Prompt list or sampling rules, intent categories, fixed vs exploratory | Shows coverage and protects against cherry-picking. |
| Platforms | Engine, mode/model where known, logged-in state if relevant | Different engines are different strata. |
| Time | Exact reporting window and run schedule | Makes recency and drift visible. |
| Sample size | Valid responses by prompt and engine, plus exclusions | Lets readers judge precision and missingness. |
| Metric definitions | Numerator, denominator, coding rules | Prevents denominator games and metric drift. |
| Uncertainty | Interval method or resampling method, dependence caveats | Separates signal from noise. |
| Evidence retention | Raw responses, URLs, timestamps, screenshots/recordings where lawful | Allows later verification after engines change. |
What are the most common ways AI visibility measurement fools people?
The biggest errors are one-shot testing, cherry-picked prompts, pooled engines, hidden denominators, over-precise percentages, unstable competitor sets and before/after comparisons that change the measurement conditions. These mistakes can make a dashboard look scientific while weakening the inference behind it. 2026 GEO survey · 45 studies2026 GEO measurement preprint
One screenshot = “rank”
A screenshot is evidence of one observed output, not a sampling distribution.
Prompt cherry-picking
Testing only prompts known to surface the brand inflates apparent coverage.
Pooling engines
A high average can hide zero visibility on an important platform.
Changing denominators
Share of voice can move dramatically without any change in raw observations.
False precision
Reporting 27.4% from a tiny sample suggests more certainty than exists.
Ignoring failed or uncited runs
Excluding inconvenient observations biases the numerator and denominator.
Mixing time periods
Model or index changes can masquerade as campaign effects.
Opaque “visibility score”
A composite score is not auditable if weights and inputs are hidden.
Industry Expert Quotes
“A defensible AI visibility score needs breadth, repetition and a timestamp. In NeuralAdX Ltd’s Month 10 benchmark, the 24 August to 23 September 2026 reporting window recorded 179 brand mentions, 28% share of voice, 15% brand coverage and a 1.13 average brand position. Our Validation Interval 5 live retrieval studies, recorded on 19 September 2026, produced 31 brand mentions and 44 neuraladx.com domain citations across three repeated studies. The lesson is to report repeated evidence and uncertainty—not a screenshot as if it were a permanent rank.”
— Paul Rowe, Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd
How should GEO testing and methodology be observed?
GEO testing should be observed as a controlled sequence of prompt, retrieval, citation, representation and repetition evidence—not as a single screenshot, ranking or isolated answer. The purpose is to identify what changed, where it changed and whether the pattern recurs often enough to be distinguished from normal generative variation.
Within Generative Engine Optimisation, this means separating retrieval from citation, separating brand presence from recommendation prominence, and separating a one-off observation from a repeatable measurement. NeuralAdX Ltd uses this distinction as part of its specialist Generative Engine Optimisation methodology.
GEO methodology also sits within a wider search and digital-strategy environment rather than operating in isolation. For a broader practitioner perspective, Top Spot Digital’s Insights & guides knowledge base brings together practical articles covering SEO, GEO, PPC and digital strategy, providing useful wider context around how generative search is being incorporated into modern search practice.
Observation principle
Start by defining the observation unit
Record the exact prompt, engine, date/time, location or market state where relevant, model or mode where visible, and the repeated-run number. Without that frame, two answers that look comparable may actually have been produced under different conditions.
Layer 1 · Prompt
Hold the question frame steady
Use a defined prompt set that covers the intended user tasks. Track paraphrases separately rather than silently mixing them into one metric.
Record: prompt ID, intent, wording, market and test date.
Layer 2 · Retrieval
Observe whether the brand or page enters the answer
First establish whether the entity is surfaced at all. Retrieval failure and weak representation are different problems and should not be collapsed into one score.
Record: mentioned yes/no, URL surfaced, answer position and competitors present.
Layer 3 · Citation
Separate being mentioned from being used as evidence
A brand can appear in an answer without its website being cited. Citation testing therefore needs its own count, source URL and fidelity check.
Record: cited yes/no, source URL, citation location and whether the source supports the claim.
Layer 4 · Representation
Evaluate how the entity is described
Presence alone is not enough. The answer may describe the business inaccurately, omit its specialism, or place it behind competitors despite citing it.
Record: entity description, recommendation strength, prominence, accuracy and missing qualifiers.
Layer 5 · Repeatability
Repeat the observation before interpreting the result
Repeat across runs, time and engines. Report recurrence and uncertainty so the reader can tell whether a pattern is stable, intermittent or platform-specific.
Record: repeated-run rate, engine, timestamp, reporting window and interval estimate.
How should changes in GEO test results be interpreted?
Interpret the pattern, not just the headline number. The same movement can mean different things depending on which measurement layer changed.
↔ Mobile note: swipe or scroll horizontally if needed to view the full interpretation table.
| Observed pattern | What it may indicate | What to inspect next |
|---|---|---|
| Mentions rise; citations flat | Retrieval or entity recognition may be improving faster than source selection. | Citation readiness, source authority and claim-level evidence. |
| Citations rise; prominence flat | The site may be trusted as evidence without becoming a stronger recommendation. | Entity differentiation, comparative evidence and recommendation context. |
| One engine improves; others do not | The gain may be platform-specific rather than general GEO improvement. | Engine-specific retrieval, citation sources and index state. |
| A single run jumps sharply | It may be normal generative variation rather than a durable change. | Repeat the same cell across additional runs and time points before concluding. |
Methodology inspection
What evidence should a reader inspect before trusting a GEO result?
A defensible GEO result should let the reader inspect the methodology, the longitudinal outcome data and at least some underlying observation evidence. The resources below are presented as examples of those evidence types rather than as substitutes for the measurement principles above.
Method specification
11-Factor GEO Methodology
Inspect the optimisation factors and how they are structured.
Citation outcomes
AI Citation Benchmark
Inspect longitudinal citation counts and comparative movement.
Visibility outcomes
AI Answer Visibility Benchmark
Inspect mentions, coverage, share of voice and position over time.
Observation evidence
Live GEO Retrieval Evidence
Inspect screen-recorded examples of time-specific AI retrieval and citation behaviour.
For additional raw examples, the NeuralAdX Ltd AI visibility video series shows observations across ChatGPT, Claude, Google AI Mode, Perplexity, Microsoft Copilot and Google Gemini. The Generative Engine Optimisation service page explains how this measurement logic is applied operationally.
When should a business start with a free AI visibility assessment?
A free assessment is most useful before committing to a large benchmark when the immediate question is whether the business appears at all for a small set of commercially important prompts and whether obvious retrieval or citation-readiness problems are present. It should be treated as a diagnostic starting point, not as a statistically precise estimate of long-run AI visibility.
If the initial tests show inconsistent surfacing, competitor dominance or weak citation evidence, the next step is to move from diagnosis to a repeatable benchmark with a fixed prompt frame, repeated observations and explicit uncertainty. The NeuralAdX Ltd assessment below combines an initial website check against the 11-Factor GEO Framework with five live AI retrieval tests.
AI Visibility Assessment
NeuralAdX Ltd
Request Your Free AI Visibility Assessment
Start with a focused diagnostic: an initial website check against the 11-Factor GEO Framework plus 5 live AI retrieval tests using commercially relevant prompts.
11-Factor GEO Framework
Initial website check
Live AI Retrieval Tests
Commercial priority prompts
Find out whether major AI engines recommend your business, cite your website, prefer competitors, or fail to surface you for the prompts that matter.
The email button opens a prepared request with placeholders for your website, contact number, five priority prompts and any useful context.
Initial assessment only · No obligation · Serious business enquiries answered within one UK business day · View live AI retrieval proof
What should you remember before trusting an AI visibility number?
Trust the number only to the extent that you can trust its sample, repetition schedule, denominator, platform coverage, coding rules and uncertainty disclosure. The question is not “is this percentage real?” but “what precisely does this percentage estimate, from which observations, during which period, and how uncertain is it?” NIST · representative samplingNIST · proportion intervalsACL 2026 · generative-search stability
The clearest mental model is: AI visibility is a sampled behaviour of a changing system. A good benchmark does not pretend volatility has disappeared. It measures the volatility, preserves the evidence and makes claims narrow enough to remain true.
A useful minimum disclosure sentence is: “Across [N] valid responses generated from [P] predefined prompts on [E] named AI engines between [start date] and [end date], the brand was mentioned in [x/N] responses ([rate]%; [interval]), using [interval method]; results are reported by engine and should be interpreted as period-specific observations.”
What are the most important questions to ask about AI visibility sample size, repetition and uncertainty?
Ask whether the sample represents the prompt space, whether each prompt-engine condition was repeated, whether results are separated by platform and time, and whether every rate is accompanied by its denominator and uncertainty. The following answers summarise the measurement standard developed throughout this article.
Is one AI visibility check enough?
No, not for estimating stable visibility. One check is useful as event evidence or a diagnostic snapshot, but it cannot estimate run-to-run frequency or uncertainty.
What is a good minimum number of repeats per prompt?
There is no universal minimum, but current field evidence supports treating roughly seven repeated brand-presence observations per prompt-engine cell as a practical floor rather than a precision guarantee. More may be needed for low-frequency signals, small differences or correlated runs.
How many total AI responses are needed for a reliable visibility percentage?
For one independent binary proportion, about 97 observations give a worst-case approximate 95% margin of error near ±10 percentage points, while about 385 give ±5 points. Real AI measurement may require more because responses can be clustered by prompt, engine and time.
Should every prompt be run many times on the same day?
Usually no: distribute repetitions across the reporting period. A same-day burst can over-sample one engine/index state and make the effective sample less diverse than the raw count suggests.
Should ChatGPT and Google AI Mode results be averaged together?
Report them separately first. An aggregate can be added only when the weighting is transparent and the combined number answers a real decision question.
That separation also reflects the wider way GEO is being practised across multiple answer environments rather than as a single-platform ranking exercise. SugarNova’s guide to what a GEO agency does describes Generative Engine Optimisation across platforms including ChatGPT, Perplexity and Google AI Overviews, alongside entity authority, structured content and citation building.
What confidence interval is best for AI mention rate?
Wilson, Agresti–Coull or exact binomial intervals are more defensible than a naïve symmetric Wald interval for small or extreme samples. If repeated observations are correlated, use a cluster-aware or time-aware method rather than pretending every run is independent.
Can share of voice be compared between tools?
Only after confirming that the numerator, denominator, prompt set, engines and counting rules are equivalent. Different definitions can produce very different percentages from the same underlying observations.
Does a higher AI citation count mean higher AI visibility?
Not necessarily. Citation volume, brand mention frequency, recommendation prominence and prompt coverage measure different parts of AI visibility and should be reported separately before any composite is created.
How often should an AI visibility benchmark be refreshed?
Refresh on a cadence that matches the business decision and platform volatility, while preserving a fixed core methodology. Monthly reporting is useful for many commercial benchmarks; higher-frequency tracking may be justified during launches or major platform changes.
The same principle appears in sector-specific GEO implementation. XTRACT’s guide to improving AI-search visibility for law firms includes ongoing monitoring of AI mentions, sentiment and citation frequency as part of a structured GEO process, illustrating why visibility should be tracked over time rather than treated as a one-off check.
What is the biggest warning sign in an AI visibility report?
A precise-looking percentage with no sample size, denominator, prompt list, engine split or date range is a major warning sign. Without those fields, the result cannot be meaningfully audited.
What should you investigate if ChatGPT recommends a competitor but not your brand?
Treat the result as a diagnostic signal, not evidence of a permanent ranking. Repeat the test across relevant prompt variants and observation windows, then investigate why the competitor may be easier for the system to retrieve, understand or trust. Useful checks include entity clarity, relevant third-party mentions, topical content coverage and technical accessibility. Panovista’s analysis of why ChatGPT recommends competitors examines those evidence gaps from a practical SEO and AI-search perspective.
Which sources support this AI visibility measurement methodology?
The methodology combines established statistical guidance, peer-reviewed LLM evaluation research, current 2026 GEO research, large-scale industry measurement studies and clearly labelled NeuralAdX Ltd first-party benchmark evidence. Source types are kept distinct so readers can judge evidential weight rather than treating every citation as equivalent.
Statistical standards & methods
NIST exact binomial confidence limits
NIST sample sizes for proportions
NIST uncertainty with autocorrelated measurements
Peer-reviewed / academic LLM evaluation
ACM TOSEM: non-determinism of ChatGPT
EMNLP 2025: prompt sensitivity
ACL 2026: prompt sensitivity
Current GEO research
2026 critical survey of 45 GEO studies
2026 GEO measurement preprint: 100K+ responses
Broadcastwell 2026 open GEO dataset
Large-scale industry measurement
AirPulse: 30,504 repeated responses
Writesonic: 161,286-prompt source-overlap study
GetIntel: share-of-voice denominator analysis
NeuralAdX Ltd longitudinal evidence
AI Answer Visibility & Share of Voice Benchmark
AI Citation Benchmark
Month 10 Citation Benchmark transcript
Live GEO proof centre
Practical prompt-tracking perspective
Omni Eclipse argues for smaller fixed prompt sets that are actually read in full, with repeated runs across days and mentions tracked separately from citations. That practical approach complements the statistical methodology above by emphasising consistency of the prompt set and separation of distinct visibility outcomes rather than relying on raw prompt volume alone.
Continue the research
For related Generative Engine Optimisation research, measurement notes and evidence-led articles, browse the NeuralAdX Ltd blog archive.
Editorial note: research reviewed on 14 September 2026. Statistics are tied to the cited study’s own sample and should not be assumed to generalise unchanged to every AI platform, category, country or future model version. Preprints and commercial field studies are labelled by source type and are not presented as equivalent to established statistical standards or peer-reviewed research.
