Methodology & data
What we measure. How we calculate. Where the limitations lie.
Original measurement wave: 46 industries
Data collection on 23–24 October 2026: 460 questions across 46 industries × three language/location profiles × three models × two search modes × three repetitions = 24,840 attempted calls. There are 24,822 complete responses; 18 failed calls remain missing. The separate pretest is not included.
The analysis uses 24,786 responses from 8,262 complete triplets. Another 36 complete responses are excluded because their repetition groups are incomplete. Visiting this website does not trigger AI queries.
- GPT:
gpt-6.1-sol - Gemini:
gemini-4-argon - Claude:
claude-opus-5-5 - Profiles: Zurich/German, Lausanne/French, Lugano/Italian.
- Search modes: web search available and no web search. Available does not necessarily mean executed. Provider-specific search metadata are not comparable measures of treatment.
MBA and Executive MBA: new collection using natural questions
The withdrawn supplementary rankings have been replaced by two new, separate collections: 540 complete responses each, from 10 questions × 3 language/location profiles × 3 models × 2 search modes × 3 repetitions. The questions contain no artificial exclusions of other programme types. The original 46 industries and the old collection archives remain unchanged; old responses are not attributed to the new questions.
All 2,273 name candidates were resolved against official institution and programme sources and independently counterchecked. The MBA wave contains 91 provider and 16 programme targets; the Executive MBA wave contains 89 provider and 38 programme targets. Institutions, organisational units, programmes and joint awards remain distinct. HSG is no longer removed by a blanket alias exclusion. Unresolved identities and unsupported components remain blocked; their omission is not treated as a negative assessment.
The four metrics are calculated separately for each wave. Non-mentions are counted only when identity and enumeration are sufficiently clear; they are not independently source-certified. Missing components are never replaced with zero. In the overall selection, two providers and two programmes in each wave have a complete AI Score. Other targets may have individual metrics or values under narrower filters, but no unsupported overall rank.
R and Python were reconciled separately for both waves; repeated runs produced byte-identical CSV results. Website reconciliation covers all 48 filter combinations. These checks and the identity counterreview are not a complete independent recoding of every answer or scientific quality approval. The measurement concerns portrayals in AI responses, not institution or programme quality.
Historical programme families do not establish current joint availability. Current EPFL-only mentions are not retrospectively attributed to UNIL. MBA ETH MTEC is announced for 2028; according to its official source, no MBA cohort has taken place yet. Its participant statistics concern MAS MTEC. Logos are programme or institution marks; for historical partnerships, both official partner marks are combined for display, not presented as a new official joint logo.
Four metrics, different denominators
The following fictional examples illustrate the metrics using the question: “Which bank is suitable for digital everyday banking?” Bank A, B and C are placeholders, not real brands. The excerpts should be read independently and are not taken from the study.
- Visibility
- Share of clearly mentioned targets among counted response–target pairs. A mention alone is not a recommendation.
Example: “Providers include Bank A, Bank B and Bank C.”
Bank A is visible but is neither evaluated nor recommended here. - Representation
- Share of positive portrayals among substantively described, evaluable cases. A description alone does not count as positive. Cases without a description are not applicable, not negative.
Example: “Bank A offers an intuitive app and accessible customer service.”
The brand is substantively described in positive terms. It has not yet been named the first choice. - Consideration
- Share of counted cases in which the brand is presented as suitable for the specific need.
Example: “If you want to manage your account entirely through an app, Bank A is a suitable option.”
The brand is explicitly linked to the stated need but is not singled out as the best choice. - Top Recommendation
- Share of counted cases with a clear first/best overall choice or rank 1 in an explicitly prioritised recommendation. Popularity, market leadership, merely being listed first or “check first” do not automatically qualify.
Example: “For your digital everyday banking needs, Bank A is my first recommendation.”
A clear first choice for the specific need. In contrast, “I would check Bank A first” alone is not sufficient.
The GEO funnel is a metric profile, not a conversion funnel. Stage width represents the percentage on a shared scale. Different denominators can make a later stage wider. Missing values are shown in grey, not as 0%.
Calculation: mean ± SD
- Use only complete triplets for each question, model, profile and search mode.
- For each model × profile × search mode × repetition: positive assessments / counted assessments × 100.
- Weight the selected strata equally within each repetition. With all filters included, there are 18 strata. If even one required denominator is missing, that repetition’s mean remains missing.
- Calculate the mean and sample standard deviation (SD) across exactly three complete repetition means. If a repetition is missing, the mean, SD and rank remain unavailable.
SD is reported in percentage points (pp). It is neither a standard error nor a confidence interval and does not capture all uncertainty. Summed numerators/denominators do not necessarily match the equally weighted mean when stratum sizes differ. Pooled diagnostics from the private analysis do not replace missing primary values on the website.
The filters change the selected strata; the same rules apply to the remaining strata. The map shows the three profiles for the selected model and search mode; clicking selects a profile for the table and funnel.
AI Score: four metrics, each weighted at 25%
AI Score = (Visibility + Representation + Consideration + Top Recommendation) / 4. By default, the ranking is sorted by this score in descending order. Each of the four metrics receives a 25% weight, regardless of its denominator. The score ranges from 0 to 100 points; it is not a probability or market share.
For each of the three repetitions, the four stratum-balanced metrics are calculated first and then averaged with equal weights. The displayed values are the mean and sample SD of the three repetition scores, in points. SD is not the average of the four individual SDs.
If a metric is missing in any required measurement group or repetition, the AI Score, SD and rank remain unavailable. Missing values are neither replaced with 0 nor bypassed by averaging the remaining metrics. A genuine value of 0 remains valid. The four individual metrics and their respective denominators remain visible separately.
In the overall selection of the original 46 industries, 188 of 15,003 provider targets and 39 of 8,405 product/offering targets therefore receive an AI Score. All remaining targets can still be found without a score. Filters may change availability; providers and products remain separate. This equally weighted summary is a presentation choice, not an independently validated quality index.
Separate entities, conservative exclusions
Providers and product families/offerings are never mixed. Directly primary-coded product families and offerings can be entities. Individual models and variants require an evidenced family relationship. Product assessments are not automatically transferred to providers or vice versa, and no new aliases are invented.
Evidenced language and spelling variants of the same entity were cross-checked against sources and merged across the original 46 industries. As a result, Universität St. Gallen and ETH Zürich are each analysed jointly across the three language profiles within each industry. Departments, degree programmes and other distinct entities remain separate; unresolved identities remain excluded. Metrics were then recalculated from original response components, not by adding percentages.
There are 23,408 analysis targets across the 46 displayed industries: 15,003 providers and 8,405 product/offering targets. These are not globally unique brand inventories or a complete market directory. In the original 46 industries, 4,034 candidates were excluded. The three vehicle industries, Car Brands, Premium Cars and Electric Cars, have no product-ranking entities because evidenced family relationships are missing; provider results remain available separately.
Borderline cases are excluded from both numerator and denominator for the affected metric only, counting as neither yes nor no. Clear components of the same response may be retained.
Non-mentions and missing values
Non-mentions are derived from formally complete primary coding with resolved identities at the relevant level. Identity conflicts, unresolved roles, known name occurrences without an assigned row and enumeration blocks prevent unjustified negative values. A missing registry match or text match alone does not prove absence.
These non-mentions are derived from primary coding, not independently certified against sources. The coverage view distinguishes counted, excluded and not-applicable cases. A rate may be missing despite a positive total denominator if individual required strata are not covered.
In the overall selection, Representation is fully defined for only 188 provider targets and 39 product/offering targets. Rarely described brands are missing in individual strata; this gap is not filled with 0%. A narrower filter may yield evaluable values but changes the comparison context.
Source checks and reproducibility
For the original measurement wave, in addition to the existing coding, 14 purposively selected responses were checked in full, along with 15 entity cases checked against original text excerpts. This is not a random sample or a representative estimate of accuracy. Decisions are privately linked to source texts, exact lines, quotations and hashes.
Calculations were performed in R. A second R run produced byte-identical CSV results. All stratum rates and primary means/SDs were also independently recalculated arithmetically; 4,070 targets were reconstructed independently from the sparse inputs through to numerators and denominators. These checks demonstrate computational integrity, not independent semantic validity.
The website receives only aggregated counts for results. Original responses, source evidence, raw labels and private audit archives are not published. Selected original questions are shown as examples; English translations are provided for illustration and were not used as an additional study profile. There is no connection to the AIRank database.
Map, ranks and management guidance
The map illustrates the three language/location profiles using language-region shapes. It is not a cantonal or population-representative survey. Language and location vary together; differences cannot be attributed to either factor in isolation. Romansh was not studied separately.
Geometry: Swiss Federal Statistical Office / GEOSTAT / ThemaKart, map geometry set 2026, language regions as of 1 May 2022, generalised total area K4sprg20220501gf_ch2007Poly, proportionally converted from CH1903+ / LV95 to SVG. City markers are approximate.
Ranks are descriptive orderings by the selected metric, not overall quality scores. The overall comparison includes different question contexts and does not identify a universal overall winner. Management suggestions are general aids to interpretation; neither causes nor optimisation effects are established by this measurement. Brands without a stored logo appear with an initials icon. Logos are used solely for identification, not to imply a partnership.