How we measure.

Every result belongs to one cell: an engine, a market, a language and a collection method. Results from different cells are never combined into one number. Arabic and English, different markets, and consumer-interface and API answers are always reported separately.

Within one market and language, an “all engines, equal weight” average is offered only as a secondary view, only across engines collected the same way, and always after the per-engine results.

Customers track themes: a core need (“dental implants”) plus the words that change who can serve it, such as a place, quality, price or service (“dental implants Muscat”). Each theme belongs to one market and language and is asked as 15 to 25 questions, 20 by default. Every question asks the engine to recommend or name companies; the theme itself is always one of them.

Questions are generated from organisation research and approved by a person with the theme. Arabic themes are written as Arabic speakers search, not translated from English, with Saudi-dialect variants alongside Modern Standard Arabic. Questions are sent exactly as written: no format instructions, personas or system prompts, because each of those changes the answer.

Questions and themes carry equal weight. There is no reliable search-volume data for AI questions, so a rarely asked wording counts as much as a common one. This is a known limitation.

Collection runs every Saturday in each market's local time: three runs per question on every engine. Most engines are measured from the answers a logged-out consumer sees, captured by specialist data providers from the market's location and in the question's language; the location and language each provider reports are stored, and mismatched answers are excluded. Claude is measured through Anthropic's official API with web search.

Answers from official APIs can differ substantially from consumer answers, so every answer records its collection method and the two are never combined. A consumer answer is never presented as an API answer, or the reverse.

Public answers are shared: an identical question, asked of the same engine in the same market and language on the same day and slot, is collected once and used for every customer who tracks it. Each customer's themes, results and reports stay private to their workspace.

Each result records its collection method, market and language. Answers that do not match the intended market and language, or cannot be measured reliably, are excluded.

Each answer is stored exactly as returned. It is then turned into one structure: the companies, products, places and people it names, in reading order, each with its role and the reason the answer gives. Code checks every extracted name against the stored answer, which never changes; answers that cannot be structured reliably are flagged and excluded.

Only options take positions: entities the answer offers as something to choose. Regulators, rating bodies and sites the answer only cites count as named, without a position. Client reports do not show full answer text. Access to full answers for review is limited to authorised staff.

A position counts entries in reading order across the whole answer. Each list item or table row that names a new option is one entry; an option named only in prose takes its first mention. Companies in the same entry share its number, and a branch or product takes its company's position.

Question rank is your average position over the runs of one question that give you a position, with its range and “named in x of n runs”. Theme rank is the equal-weight average of your question ranks across the theme. Positions 1, 3 and 5 on three questions give 3.0; adding a fourth question where you are not named keeps 3.0 and shows “named in 3 of 4 questions”. A brand that is not named is shown as “not mentioned”, never as zero or last place.

Ranks always sit beside mention rate. A brand named once, in first place, has a good rank and a low mention rate; you need both to judge it.

For a single question we use the Wilson score interval, which stays sensible with small samples and at 0% or 100%. For aggregates we use a cluster bootstrap that resamples whole questions, with all their runs, 2,000 times. Runs of the same question are correlated, so treating them as independent would overstate precision.

Questions within a theme are near-identical wordings, so above theme level we resample themes first and then questions within them. Adding similar questions cannot buy false precision. Bootstrap seeds are fixed, so any result can be reproduced exactly.

Every result carries one of three labels, modelled on the IAB's August 2026 guidance for measuring AI visibility, with machine-readable reasons. A result gets the highest label whose conditions are all met.

The IAB's guidance asks for coverage of four query intents. Because every question we track asks for companies, we use theme coverage instead, and say so here. A single theme has fewer than 50 questions, so theme-level results are Exploratory; Directional and Decision-grade apply to results across a market and language's themes.

Each week we re-run a calibration set, ten questions ten times on the same day, to measure how much results move when nothing has changed. That noise band is scaled to each programme's size.

A change is reported as real only when its paired 95% range excludes zero and it is larger than the noise band. Otherwise the view says “no clear change”. When an engine changes its model, affected windows are annotated and old and new baselines are shown side by side for 28 days.

GA4, Search Console and Bing's AI Performance data are shown as a floor, never as a total: Google's AI Overview clicks cannot be separated in GA4, and a share of AI-referred visits arrives without a referrer. Correlations, such as branded search alongside mention rate, are labelled as correlational.

• A single-run position presented as a rank
• A number that blends languages, markets or collection methods
• A rate without its range and label (simple counts excepted)
• A composite score without its published weights and components
• Full answer text in client reports
• A result attributed to a personal agent such as OpenAI dots when it came from a stand-in; those are always labelled simulated

• Equal weights: questions are not weighted by how often people ask them, because that data does not exist for AI assistants.
• Generated questions: research-based questions can favour well-known brands. Results identify this limitation.
• Accuracy: extracted names are checked against hand-labelled Arabic and English answers; cases that cannot be structured reliably are excluded.
• Coverage: some Google searches show no AI Overview at all; we report how often one appears.
• Terms of service: engine terms differ by product and collection method; each result identifies how it was collected.