How Dfeelings Measures AI Search Visibility

How are prompts selected?

Prompts come from evidence of what buyers actually ask, not from keywords chosen to flatter a brand. Each prompt set is written for one market and one language, grouped by intent, reviewed, then frozen with a version number before any testing.

• Sources: Search Console queries, questions from sales calls, support messages and contact forms, and the wording customers use in their own messages.
• Intent groups: learning (what is…), comparing (X or Y), choosing a provider (best company for…), and brand checks (what is <brand>).
• Brand names appear only in the brand-check group, so mention rates in the other groups are not inflated by the prompt itself.
• Near-duplicate prompts are merged before freezing, so one question cannot be counted twice.
• Any change to the prompts creates a new version, and a new version needs a new baseline.

How are industries selected?

For a client, the industry is the client's own service lines and the questions their buyers ask. For a Dfeelings Research study, the industries are fixed in the published protocol before fieldwork, with the reason for each choice, and are not changed after results are seen.

How are countries handled?

Saudi Arabia and Jordan are measured as separate markets and never pooled. Answers can differ by location, so the test location is recorded for every run.

• Each prompt set belongs to one market; a Saudi set is not reused for Jordan by swapping the country name.
• Where an engine allows it, the region or location setting is fixed to the market tested; where it does not, that limitation is recorded with the run.
• Market-level rates are reported per country, which is what the market coverage metric means.

How are Arabic and English handled?

Arabic and English prompts are separate sets written by native speakers for the same intents. Arabic prompts are written in the Modern Standard or Gulf/Levantine phrasing buyers actually use, not translated from English, and results for the two languages are reported separately.

Which engines are tested, and under what settings?

Each engine is tested and reported separately because they retrieve and cite sources differently. The engine list is fixed per study or client before testing.

• Typical engines: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Microsoft Copilot.
• For every run the product, mode and model name are recorded exactly as displayed, plus whether web search or browsing was on.
• Tests run in fresh sessions with chat history and memory off, signed out where the product allows it, so earlier conversations cannot shape the answer.

How many times is each prompt tested?

Each prompt is run a fixed number of times per engine, set in the protocol before testing, because the same prompt can produce different answers. Every run is one observation, and the number of runs per prompt is kept identical between a baseline and its retest.

How is a brand mention counted?

An answer counts as one mention when it names the brand or an alias on the brand's pre-agreed alias list, however many times the name appears. Answers that describe the brand without naming it are not counted, but are noted.

How is a citation counted?

An answer counts as one linked citation when any source link in it points to the brand's own domain or subdomains. Every cited URL is also logged and reduced to its registrable domain, so the sources shaping answers can be analysed even when the brand is not one of them.

How is prominence defined?

A reviewer codes each mention against a written rubric: primary when the brand is the main or first recommendation, listed when it is one of several options presented together, and passing when it is mentioned without being recommended. Prominence is reported as a distribution, never as a ranking position.

How is accuracy checked?

Each mention is checked against the brand's approved company facts: what it does, where it operates, and for whom. The definition of accurate is written down before coding and the same reviewer codes the baseline and the retest.

How are duplicate results handled?

Repeated identical answers across runs are kept, because each run is an observation. Within a single answer, a brand named several times is one mention and a domain linked several times is one citation. Prompts that turn out to be duplicates after freezing are flagged and excluded from both baseline and retest.

What are the limits of this method?

AI answers vary by session, account, location, time and model version, and engines change without notice. The method reduces this variation but cannot remove it, so results are rates over many observations with their counts shown, not guarantees, and a single screenshot is never treated as evidence.

• Results describe the prompts tested, not every question a buyer could ask.
• Logged-out and fresh-session testing differs from what a signed-in user with history may see.
• A change in visibility after GEO work shows association, not sole cause; other changes are recorded alongside it.

How are dates and model versions recorded?

Every run records the start and end date of fieldwork in ISO format, the engine product, mode and model name as displayed, the prompt-set version and the runs per prompt. Known engine updates during or between tests are noted, because they can change results on their own.

How does retesting work?

A retest repeats the baseline exactly: the same prompt-set version, engines, settings, market, language and runs per prompt. If any of these change, the result is a new baseline, not a before/after comparison. Retests are published with the dates of both tests and a dated list of what changed in between.

How are competitors compared?

When a client wants competitive context, a list of competitor brands and domains is agreed and frozen before testing, and the same mention, citation and prominence coding is applied to each. Competitor data is reported to the client and is published only when it was actually collected and its publication has been approved; it is never estimated.