How Searchestra measures AI visibility

AI visibility is only as trustworthy as the method behind it. This page explains how we measure, in plain terms, so you can judge our numbers rather than take them on faith. It is deliberately specific about what we do, and honest about what we do not claim.

Why we publish our method

Two AI visibility tools can measure the same brand and report different numbers, because they use different questions, engine coverage and definitions. That is exactly why the method matters more than the headline score. We would rather you trust a number you can interrogate than a confident figure you have to take on faith. This mirrors the transparency the IAB's Measuring Visibility in the AI Era framework asks measurement providers to compete on. See also directional vs decision-grade and what to ask a provider.

The AI answer engines we track

We maintain a catalog of AI answer engines spanning assistants and AI-search surfaces. Each is measured on its own, never blended into a single vanity score:

Honest caveat: an engine is only queried when access to it is configured. Coverage for a given account depends on which engines are enabled, so we report per-engine rather than implying every surface is always live. See multi-platform aggregation.

How answers are collected

We ask each engine through its provider's API with the provider's own web search switched on wherever it is offered: ChatGPT, Gemini, Claude, Meta AI and Grok. The model decides for each question whether to search; we never force a search. An answer given without searching is a real observation of how that engine behaves, so it stays in every metric. For every answer we record whether the model searched and how many searches it ran, and the Citations screen shows how often each engine chose to search. DeepSeek's API offers no web search, so its answers come from the model's own knowledge. On Perplexity, Google AI Overviews and Google AI Mode the search is part of the product rather than the model's choice: Perplexity's own answer setting instructs a search before every answer, and AI Overviews and AI Mode are Google Search features that build the answer from several related searches. Google shows an AI Overview only when it judges it helpful, so some questions have none. We record those questions and report how often an AI answer was shown, but they do not count as answers in the other metrics.

The unit of measurement

Our unit is not a single API call but a question paired with an engine, run as a real request, not a simulation. Results are reported per engine so a strong showing on one surface never hides invisibility on another. This is the difference between share of model and share of voice: we keep the layers separate rather than averaging them into one blurry figure.

Where the questions come from

The question set for a category is generated from a taxonomy of topics and a graph of realistic buyer scenarios, rather than hand-picked, so coverage is systematic instead of cherry-picked toward questions a brand happens to win. Each question carries its own metadata (intent, scenario, source type, version).

On sourcing, which the IAB framework asks providers to be explicit about: the universe combines the taxonomy and scenarios with evidence of real search demand, and uses synthetic expansion only to fill coverage gaps, never as the whole set. Search-demand signals inform the universe but do not solely define it, and every question records which source type it came from.

Questions are classified by intent: informational, commercial (including recommendation-style "best X" queries), transactional, navigational and comparison. This spans the intent types the IAB decision-grade approach asks to be covered and segmented, and the classifier runs in both English and Turkish, so a clearly commercial Turkish question is not silently treated as a general one.

The question set is brand-neutral by construction: the brand being measured does not influence which questions are generated. Brand identity is used only as a downstream accept-or-reject check on already-generated candidates, never to shape what is asked. That separation is what keeps the set from being quietly tilted toward a flattering result, the guardrail behind prompt and platform bias.

The set is versioned and immutable: a measurement run records exactly which version of the question set it used, so a change in results can be told apart from a change in questions. See measurement stability and reproducibility.

For each Prompt Universe we can disclose the transparency profile of the exact question-set version in use: how many questions it holds, how they break down by intent and by source type, the weighting basis, and the baseline and refresh policy behind it. These cover several of the query-set transparency dimensions the IAB framework identifies, so the question set behind a score is not a black box; we are still expanding what we expose here.

The four layers we report

We report along the four layers of the IAB framework rather than one score, so you can see where representation breaks down. The layer names are the IAB framework's; the metrics in the third column are how Searchestra measures each layer:

These are separately reported measures, not a black-box composite. Every one of them is defined below: what is counted, what it is divided by, and where our definition differs from the framework's. See the 4 P's of AI visibility, mention rate and citation rate.

The formulas, in the open

Every number we report is a plain ratio or average you could recompute from the answers behind it. There is no proprietary reweighting and no blended score. Ratios for a period are totals over totals: all matching answers in the period are added up first and then divided, rather than averaging daily percentages, so a day with few answers does not weigh as much as a day with many.

Share of Voice: our two choices

The IAB framework asks providers to state two choices, because they change the number. The competitive category is the set of brands tracked for your market: your brand plus the competitors listed for that market, which you can see and edit. Platforms, publishers and other names outside that set are left out. The mention universe is the sum of mentions across exactly that set, so the shares of all brands in the set add up to 100%. Two tools can report different Share of Voice for the same brand purely from these two choices, which is why we state ours.

Citation Rate: what counts as a citation

We count linked citations to domains registered as yours. A brand named as a source without a link is not counted, which is narrower than the framework's definition. A citation is a source the engine itself returns with its answer. Engines return sources when they search the web, so an answer given without a search, and every DeepSeek answer, has no citations. Those answers still count in the denominator: an answer that does not cite you is a real outcome, not missing data.

Recommendation strength rubric

For each answer that mentions you, a language model classifies the three dimensions the IAB framework names: endorsement (specific reasoning, generic, or none), prominence in the answer (primary, one of several, or peripheral) and qualification (unconditional or qualified). The score is then computed from those labels by a fixed rule: no endorsement scores 0; otherwise specific reasoning scores 2 and generic endorsement 1, plus 1 if the recommendation is primary and plus 1 if it is unconditional, for a range of 0 to 4. Answers scoring 3 or 4 count as active recommendations. A 3 does not require being the first choice: specific, unconditional praise for one of several listed options also scores 3.

A negative mention is not a recommendation; tone is reported under Portrayal. The rubric is versioned: when it changes, earlier scores are kept separate and never mixed into the new measure. If the model's output does not fit the rubric after one retry, the answer is left unscored rather than guessed.

Sentiment and flagged inaccuracies

Sentiment is scored by a language model that reads the whole answer and returns one score for its overall tone toward the brand. It is a score for the answer as a whole, not for each sentence, and we do not publish an accuracy benchmark for it, so read it as a signal rather than a verdict. If scoring fails, the answer has no sentiment value instead of a neutral 0.

Inaccuracies are flagged by a language model review of answers that mention you, plus fixed checks against product data where a brand has supplied it, for example a wrong price, an unsupported health claim or outdated product details. Framing errors, where the brand is positioned wrongly, are flagged but not counted in the rate, because they are a matter of positioning rather than fact. A recommendation given without a reason is not a portrayal issue at all; it is measured under Persuasion. Flags are not yet verified against brand facts or the cited sources, so this is not the framework's Factual Inaccuracy Rate, and it does not separate errors an engine made up from errors repeated from an incorrect source.

Momentum: when two periods are comparable

Momentum compares the selected period with the previous period of equal length. If the previous period has no data or a Mention Rate of 0, no change is shown, because a percentage change from zero is undefined. When the question set moves to a new version that breaks comparability, reports mark it as a baseline break. Engine model updates are not yet detected automatically, so a jump that appears for every brand at once should be read as a possible platform shift before it is read as a gain.

Reproducibility and change

Because AI answers vary between runs, a single answer proves nothing. Each question is asked on each engine at most once a day, and a period's numbers add up every daily answer in it, so one volatile response does not masquerade as a trend. Results are always tied to the versioned question set they came from, and every answer records the version of the collection method behind it, so a change in how answers are collected can be told apart from a change in the engine. See visibility momentum and understanding AI non-determinism.

The honesty principle

We report what we collect and flag what we cannot. When a signal is not measured, for example if recommendation judging is not enabled for a run, the field is left empty rather than filled with a guess. We do not present a confident single number where the data only supports a directional read. A documented limitation is part of the rigor, not a weakness to hide.

What we do not claim

An API result is not identical to what a signed-in person sees in a product's interface, and we are explicit about which layer a given number reflects. The same applies to web search: the search an engine runs through its API is the provider's developer search tool and can differ from the search its consumer app runs. Sentiment, recommendation strength and flagged inaccuracies come from a language model's judgment and are treated as signals rather than verdicts. Post-citation click-through is not measured, because AI platforms do not share click data per mention. Where a number is directional, we say so plainly. See directional vs decision-grade measurement.

Frequently asked questions