Methodology

We generate the questions your buyer might ask, put the same questions to every assistant in the tier separately, and record for each answer: whether your brand was mentioned, in which position, in what tone, which brands appeared, and whether the information given contradicts your verified information.

The questions do not come from real user query logs; they are scenario queries representing purchase and category intent. The free scan uses 2 assistants (ChatGPT and Claude); the full analysis uses 4.

The score combines all of this. It is called answer share: how much of the answers given to those same questions you appear in.

numerator   = Σ(value × category_weight × assistant_weight)
denominator = Σ(category_weight × assistant_weight)

share       = numerator / denominator
raw_score   = share × 100
penalty     = min(inaccurate_findings × 5, 15)
score       = clamp(raw_score − penalty, 0, 100)

A single measurement's value starts from visibility; position and tone multiply it. Where you are not mentioned there is no position and no tone — that measurement is simply zero and does not enter the average.

value = position_weight × tone_multiplier   (if mentioned)
      = 0                                 (if not mentioned)

tone_multiplier = 0,30 + 0,70 × (tone / 100)

Even with a tone of zero, being mentioned still earns 30% of the value: being mentioned in a poor context is not the same as not being mentioned at all.

Buyers usually look at the first three options. Being mentioned at the bottom of a list is only slightly better than not being mentioned.

Not every question is worth the same. Missing from a purchase-intent question costs more than missing from a category-discovery one.

Brand verification does not count towards the score, which is why it is absent from the table. Those questions already contain the brand name; being mentioned in the answer would be a tautology, not a measurement. They produce a separate metric: an accuracy score.

The weights reflect estimated usage share in Türkiye. They sum to 1,00.

If an assistant fails entirely during a measurement, that question's total weight falls below 1.00. The remaining assistants' weights are scaled back up to 1.00 while keeping their ratios: what stopped is the assistant, not the value of the question. A scan like that is marked partial in its result.

Where an assistant's information contradicts your verified information, each finding subtracts 5 points, up to a maximum of 15. A contradiction claim requires evidence: the contradicting sentence itself is recorded.

The horizontal axis is visibility, the vertical axis is tone. The thresholds are 50% and 50 points.

The score is calculated from core questions only. Two kinds of question are deliberately excluded:

Brand verification questions contain the brand name directly; being mentioned in the answer would be a tautology rather than a measurement. They produce a separate accuracy metric and are never asked in a free scan.

Custom questions — the ones you want to track yourself — do not count towards the score and appear in their own section of the report. If they did count, the score would be comparable neither with another brand nor over time, and it would be the easiest way to inflate it.

A measurement we could not obtain does not enter the denominator. If an assistant does not answer, that cell is not recorded: planned measurements stay the same, completed ones drop, failed ones rise. Counting a provider error as “not visible” would hold something we never measured against the brand.

The seven numbers used in reports and on screen are defined in one place; the screen, the PDF and the admin report use the same words.

Free            6 questions × 2 assistants = 12 planned measurements
Full analysis   40 questions × 4 assistants = 160 planned measurements
Enterprise      100 questions × 4 assistants = 400 planned measurements

Every report states how much of the measurement could be obtained: High / Medium / Low / Insufficient.

This is not a statistical confidence interval. The percentage says how much of the planned measurement was obtained; it is not an estimate of sampling error. The level is rule-based: completion rate, number of failed assistants and the question categories that could not be measured are assessed together.

Low visibility does not lower the confidence level. “You are not visible at all” is a valid measurement result; the confidence level describes only the coverage of the measurement.

The free scan runs 6 questions across 2 assistants (ChatGPT and Claude): 12 measurements in total. This limit is the same whatever your industry. Because the measurement base is narrow, the score is marked estimated.

The analysis may draw on industry measurements collected earlier. The scope of the result does not change: the score, visibility and assistant breakdown shown are in every case calculated from these 12 measurements.

The full analysis runs 40 questions across 4 assistants. The “measurement base” block at the foot of every report states the date, the question and assistant counts, and the measurement confidence level.

The two tiers also say different things. The free scan gives you a score, the assistant breakdown, the brands mentioned in AI answers, one example of a lost query, a high-level signal about AI access to your site, and a high-level diagnosis. The full analysis explains why: a page-level site audit, detailed brand and competitor analysis, technical reasoning, and a roadmap carrying priority, effort and time-to-impact.

This is not an error; it is the thing being measured. Given the same question, Claude runs a broad search and names more than twenty companies, while ChatGPT issues targeted site queries and answers briefly.

In our measurements, the overlap between the company lists assistants recommend is 21%. The spread is sharp too: 73% on comparison questions, 0% on purchase questions. Being visible in ChatGPT does not mean you will be visible in Gemini.

That is why the assistant breakdown sits at the centre of the report rather than in a footnote. A single “AI visibility” number hides this difference.

We measure through the providers' APIs. The answer a user sees in the ChatGPT app is not identical to the one returned by the API: consumer apps apply different system prompts, personalisation, memory and location.

What we measure is the model's answer to the same question without personalisation. That is the only base that is comparable between brands and trackable over time — personalised answers are not suited to comparing two brands.

This is a weakness the whole category shares. We say so, because a limit you do not know about will mislead you.