Clear questions.
Visible evidence.
A short explanation of what the report measures and how to read it.
Independent measurement; not affiliated with OpenAI, Google, Anthropic, Perplexity, or other measured providers.
Method in five steps
1. Build a relevant question set
Questions are generated around your website, buyer tasks and buying decisions, then frozen for the report. The standard target is 18 buyer questions and 2 brand checks. Actual coverage is shown in each report.
2. Collect and compare answers
Kovmere samples available search-enabled engine interfaces. A benchmark uses repeated measurements. These observations do not represent all consumer conversations or a fixed position inside every consumer app.
3. Keep the counts understandable
Buyer answers determine the main rates. Brand checks and missing responses are kept separate. “Recommended” uses the current method’s definition: presented as an option a buyer should consider.
4. Turn observations into practical changes
Website information, mentioned competitors and cited pages inform the Fix package. The package shows what to change, where it belongs and what supports it. You review and publish the materials.
5. Ask the same questions again
Pro follows the frozen set across periods. The report can show a change after an implementation without claiming that one Fix caused it.
The five steps in detail: the same every time
These are search API measurements, not results from consumer apps.
1. Build a relevant question set
We read your public pages once to identify what you sell and to whom, note the competitors you named or the ones we identify, and write the questions a buyer asks before choosing — in the buyer's words, not yours. The set is frozen for the report, and the actual coverage is shown in each report.
2. Collect and compare answers
Each question is asked 3 times through the search APIs that return attributable citations (OpenAI, Perplexity, Gemini, Google AI Overviews, and Claude); every answer is stored verbatim. An engine that cannot be sampled is marked not measured, never zero. These observations do not represent all consumer conversations or a fixed position inside every consumer app.
3. Keep the counts understandable
Each answer is judged for whether you are mentioned, recommended as an option, and whether your own page was cited. Buyer answers determine the main rates; brand checks and missing responses are kept separate. “Recommended” means presented as an option a buyer should consider, not necessarily first choice.
4. Turn observations into practical changes
Every cited page is fetched and classified by owner. The gaps between what buyers asked, what the engines cited and what your pages say become the Fix package: page content, cited-source and profile opportunities, technical artifacts and brand-fact checks, each with where it goes, why, and what supports it. You review and publish the materials.
5. Ask the same questions again
Pro follows the frozen set across periods: a later reading asks exactly the same questions, and a movement is reported as observed once, a repeated signal, or within normal variation — never as caused by one fix.
Measurement scope
Each benchmark uses the same 20-question set: 18 unbranded buyer questions and 2 brand checks. Each question is sampled three times across up to five engines, for up to 300 answers. Primary visibility metrics use only the 270 buyer-question answers; the 30 brand-check answers are reported separately.
Pro and Agency Portfolio ask the same questions again every 30 days: each billing cycle includes one full monthly benchmark and three weekly single-repeat pulses on the same question set. That is a monthly benchmark of 20 questions × 3 runs × 5 engines = 300 answers, plus 3 pulses of 100 answers each (20 questions × 1 repeat × up to 5 engines) — up to 600 answers per billing cycle, per site. A single Check is one reading, so it cannot show change. Anything not measured is printed as “not measured”, never as zero. See a complete public example.
Every rate with its count, every answer kept as evidence.
What the three words mean
- Recommended — the answer presents the business as an option to consider for that question. An undecidable answer is excluded rather than counted as “no.”
- Mentioned — the business is named at all, including as a caution.
- Cited — a URL on the measured domain appears in the answer’s source list.
- Recommended
- Recommended means the answer presents the business as an option to consider for that question; an answer we cannot decide on is left out, never counted as a “no”.
- Mentioned
- Mentioned means the answer names the business at all, including as a caution.
- Cited
- Cited means a page on the measured website appears in the answer’s list of sources.
- Answers
- Answers are the replies we collected to the buyer questions — each question asked on each engine, more than once — with the brand checks and the missing replies kept apart.
- Monthly benchmark
- The monthly benchmark asks the full question set on every engine, three times each; the later readings of the period are compared with it.
- Weekly pulse
- A weekly pulse asks the same questions on a smaller scale between benchmarks, to show whether the benchmark’s number holds — it is not daily rank tracking.
- No clear change
- No clear change means the difference stays inside the range the engines’ own answer-to-answer variation produces, so it is not reported as a change.
1. Question generation and freezing
From your site and the competitors you name (optional), we generate the questions buyers ask before choosing you: best-for, comparison, how-to, and evaluation questions. They are grouped, not listed flat, so you read the set the way you think about your market.
Near-duplicates are replaced before the set is frozen; the core set is unbranded and the two brand checks name your brand on purpose (see the scope card above).
Google Keyword Planner is an external search-demand input used in question selection. It is not an estimate of how often anyone asks an AI engine a question, and it never enters a visibility figure.
Why 20 questions, 3 repeats, five engines and the same core set every month
- Why 20 questions in total. 18 buyer questions already surface most of the providers an engine names for a category, and a small set can be frozen and re-asked identically; the 2 brand checks test how engines describe you when asked directly, and are kept out of every primary figure.
- Why 3 repeats. An engine does not give the same answer twice. Three repeats per engine-question show the engine’s own noise, so a single lucky or unlucky answer is never printed as the rate.
- Why five engines. Buyers ask different assistants; the five we measure (OpenAI, Perplexity, Gemini, Google AI Overviews, and Claude) are the ones whose search APIs return attributable answers we can verify.
- Why the same core set is preserved. A comparison is only honest when both readings asked the same questions; the core set is frozen, and an edited question starts its own history rather than rewriting the old one.
- Why an unavailable engine is not counted as zero. “No answer collected” and “answered and did not recommend you” are different facts; the first is printed “Not measured”, with the reason.
- Why these are API measurements. Search APIs carry no memory, location or personalisation, so the same question can be re-asked on the same method; a consumer-app session cannot be.
- Why we promise no ranking, citation, traffic or revenue. We measure what engines answered and what changed; no method we know can promise an engine’s future output.
- Why weekly pulses are not daily rank tracking. A pulse is 100 observed answers that report one movement vocabulary (observed once, repeated signal, within normal variation, reverted, not established); daily charts of a noisy series invite reading noise as change.
2. Repeats and variability
Each question is sampled 3 times through 5 search APIs. The five engines are OpenAI, Perplexity, Gemini, Google AI Overviews, and Claude. We record the model version, the buyer’s country, run time with timezone, answer, and attributable citations or URLs. The country is the one chosen at checkout — the United States, the United Kingdom, Canada, Australia, Ireland or New Zealand: ChatGPT, Perplexity and Google AI Overviews are asked as a buyer there, and the Google searches and search volumes behind the questions come from that country; Gemini and Claude take no location and answer from their usual one. An unavailable API is marked “not measured,” and no rate is printed for it. When Google AI Overviews shows no overview for a sampled question, that answer is marked “did not surface” and left out of every rate — it is never counted as zero. Google AI Overviews may not surface for some or all sampled questions. When no overview appears, we report that state and exclude it from visibility rates; it does not trigger a rerun or refund. A report is released only when at least 4 of the 5 engines are measured and at least 80% of the expected buyer-question answers are validly judged. Below that, the failed answers are asked once more. A report still below it is not published but re-run automatically, and if the re-run misses it too, the payment is refunded automatically.
Appendix: why each question is asked more than once (measured on our own readings)
Appendix: how accurate the judge is
Google AI Overviews is read through DataForSEO, a search-data provider: for each question it returns the AI Overview Google shows in its results.
On the roadmap and not measured today: Google AI Mode and Microsoft Copilot. An engine is named on a report only once we actually sample it.
Repeated answers are not treated as independent buyer questions. They show how the same question can vary from sample to sample; the buyer question remains the unit of measurement.
3. Facts per answer and the rate formulas
- Recommended — whether the answer presents the business as an option to consider. An undecidable answer is excluded rather than counted as “no.”
- Mentioned — whether the business is named at all, including as a caution.
- Your page cited — whether a URL on the measured domain appears in the answer’s source list.
recommended ÷ judged answersmentioned ÷ judged answersanswers citing your domain ÷ sampled answers
4. How much a rate could move with other questions
Each measured rate carries a range showing how much it could move if other buyer questions had been asked; the reports call it the question-resampling sensitivity range. It is computed over the sampled buyer questions, so repeated answers to one question never count as new, independent evidence.
On Pro and Agency Portfolio, from the second reading onward, the current reading is set beside the earliest reading measured the same way, on one timeline. At 20 questions the paired question-level test cannot establish a month-to-month change at any number of repeats, so we do not print one: a movement is reported as a movement, with the answers behind it, and where it is too small to call with this many questions we say exactly that instead of calling it. This describes repeated measurements; it does not promise a future ranking or citation, and it is not a claim that a change has been proven.
5. Citation verification and reachability
A citation is a URL returned in an answer’s source list. We fetch each cited URL and classify the page as your site, a competitor site, a directory, a list, a community thread, video, or news. Whether a page names you is checked on the cited page itself, not inferred from the answer text. A page that cannot be fetched is labeled “not fetched” rather than guessed.
6. Engine attribution and reachability gates
| Use | Engines | Rule |
|---|---|---|
| Sampled for answers | OpenAI, Perplexity, Gemini, Claude, Google AI Overviews | Only when the API returns attributable citations or URLs |
| Page reachability | Documented crawler user agents | A user-agent fetch shows what our request received; it does not prove crawling or indexing |
If DataForSEO, the provider Google AI Overviews is read through, becomes unavailable, Google AI Overviews is reported as not measured on the runs it misses. Every sampled engine also has its crawler reach checked; an engine that returns no usable answers on a run is reported as not measured on that run. Google AI Mode is not sampled.
Reachability is a separate observation. We fetch a page with a documented crawler user agent and as a browser to identify robots rules, server blocks, and browser-only content. The result says only what those requests received at that time.
7. Limits, and what is deliberately not scored
We do not commit that anything we write will change what an engine answers. We commit to asking the same questions again, on the same method, and telling you exactly what came back — including when nothing measurably moved.
- Prompt volume / how many people ask a question — we do not estimate it. Search APIs do not publish that population, so we report only the answers sampled for the questions in the report.
- Schema.org / JSON-LD as a citation tactic — it is not graded as evidence that an engine will cite a page.
- llms.txt and IndexNow — neither is treated as proof that a measured engine used a page.
- FAQ reformatting by itself — format alone is not scored as a cause of recommendation.
We do not promise rankings, mentions, citations, audience size, or sales. We deliver the described measurement and the fixes that follow from it, each written out as a block to copy into your site's editor, whatever it is built on, with the steps and who does them. A Check includes up to 3 evidence-supported, ready-to-use fixes; Pro and Agency Portfolio include up to 12 each period when the evidence supports them. Every fact about you in a fix comes from your own website, and each fix is checked against your pages before it is delivered. You do not fill in a questionnaire.
Method inputs and outputs stay attached to the dated report: the question set, model version, sampled locale, answer, citation list, page fetch result, and method version.