Metrics
How Ranqo measures
How Ranqo asks your prompts on each AI engine every week, reads every answer, and turns those answers into the numbers on your dashboard.
Every number in Ranqo comes from one source: real answers from AI engines to the questions your buyers ask. This page explains how those answers are collected and read. The rest of this section covers each metric in detail.
A run
A run asks every tracked prompt once on each engine it is set to, from the engines your brand tracks. Ranqo can track 6 engines: ChatGPT, Claude, Perplexity, Gemini, Grok, and Google AI Overviews. Which of them a brand uses depends on your plan and the brand's settings. See Plans and limits.
Runs are weekly. Ranqo checks every day at 00:00 UTC and starts a run for each brand that has had no run in the last 7 days, so each brand keeps its own day of the week. A run that fails is picked up again at the next daily check.
- Paused prompts sit out. They keep their history but are not asked. See Prompts.
- A brand with tracking paused, or with no tracked prompts, is skipped.
- Scheduled runs need an active plan or trial. See Billing and trials.
Each completed run appears in Snapshots with its own scores.
Live answers
ChatGPT, Claude, Perplexity and Gemini search the web on every run, so their answers draw on what is published now, not only on what the model learned in training. Google AI Overviews is read from Google's live results page for the prompt, including an overview that loads after the page does. When Google shows no AI Overview for a prompt, that counts as an answer that does not name you.
Every engine except Google AI Overviews, which is read from Google's results page rather than asked, receives the same instructions, and they never name your brand, because naming it would steer the answer toward you. The instructions ask for full product names, a numbered list when products are compared or ranked, and the sources the answer used. That keeps answers comparable across engines and makes the order of a recommendation readable.
Unbranded prompts by default
A prompt that names your brand almost always gets an answer that names your brand, so it says little about whether AI would recommend you to someone who has not heard of you. The prompts Ranqo writes when you set up a brand never name it. Generate with AI on the Prompts page stays unbranded too, unless you select the Comparison or Brand Research category, and it warns you when you do.
Prompts you add yourself, and prompts written for a theme, can name your brand. The prompt filter on the Visibility page, which reads All Prompts by default, can show Branded or Unbranded prompts only. It reads the branded flag a prompt was saved with: generated prompts are flagged when they name your brand, but a prompt you add yourself is saved as unbranded whatever it says, so Unbranded still includes it. See Prompts and Themes.
Location and language
Each prompt is asked from a location: the prompt's own location if it has one, otherwise the brand's target location. A prompt set to Global is asked with no location at all. Every engine except Google AI Overviews is told where the brand's buyers are, and some engines also take the location natively:
| Engine | How the location is applied |
|---|---|
| ChatGPT | Native search location (country and city), plus the instruction |
| Claude | Native search location (country and city), plus the instruction |
| Perplexity | Native search location (country and city), plus the instruction |
| Gemini | The instruction only |
| Grok | The instruction only |
| Google AI Overviews | Google results for the country; the United States when no country is set |
A brand can also have an answer language. Ranqo offers 8 answer languages: English, German, Spanish, French, Italian, Dutch, Polish, and Portuguese. English is the default. When another language is set, every engine is asked to answer in that language and to keep brand and product names in their original spelling, so a mention can still be matched. Google AI Overviews is fetched in that language. See Brand settings.
How an answer is read
Each answer goes through the same steps:
- Cleaned and reformatted. The answer is reformatted into a consistent shape, a numbered list when it names a set of products, and the reformatting is told to keep every brand, product name and link in the order the engine wrote them. The reformatting reads only the first 8,000 characters of the cleaned answer, and step 2 reads what it returns, so a brand or competitor that the engine names only after that point is recorded as not named. If the reformatting adds or removes your brand's name, comes back nearly empty, or stops partway through, it is thrown away and the whole cleaned original is used instead.
- Read by a model. A model reads the text from step 1, up to its first 12,000 characters, and records whether it names your brand (by name or by one of your aliases, when the alias clearly refers to you), your position in it, and its sentiment toward you.
- Checked against the original. If the reformatted answer names your brand but the engine's original answer does not, the mention is withdrawn, so the reformatting step cannot create a mention.
- Competitors listed. The model lists every competitor the answer names, up to 20 per answer, with each one's position and sentiment. It keeps companies that do the same job as you, plus any competitor you track wherever it appears, and leaves out directories, review sites and marketplaces (unless you are one), the AI assistants themselves, and your own brand and the companies it owns. Names are normalized to the parent brand, and each competitor's spellings are mapped to one name, so a competitor does not split into several rows. See Competitor identity.
- Citations recorded. The links the engine attached to its answer are stored as citations.
When a query fails
An engine can time out or return an error. Ranqo retries a query that fails with a temporary error. If it still fails, Ranqo looks for the most recent successful answer to the same prompt on the same engine from the last 14 days and reuses its reading: whether it named you, your position, the sentiment and the competitors named. The answer text and its citations are not reused. If there is no recent answer to reuse, the query has no answer.
A run's snapshot says how many queries used historical data (the reused answers, each marked imputed in its results) and, under Partial coverage, how many could not be answered at all.
Not every page treats these two cases the same way:
- A run's own scores count reused answers and leave failed queries out. That covers a run in Snapshots, the visibility figure and engine bars on Home, the Visibility card in the Action Center, and the points of the Visibility trend chart when no filter is applied.
- The Visibility page's cards and breakdowns, every figure on the Competitors page, Share of Voice and the competitor signal on Home, and trend chart points with a filter applied count only answers that came back in their own run. Reused answers and failed queries are both left out.
- The Tracking tab of the Prompts page, a prompt's insights page and the free AI Visibility Report read every result, reused answers included. Where they count answers, a failed query counts as one that did not name you.
So a run with failed queries can show different figures on these pages.
Principles
- One point per run. A trend chart plots one point for each run that happened and never fills in the days between. See One point per run.
- No window shorter than a run. When the date range you pick holds no run, most pages show the last completed run rather than an empty page, with an As of note giving that run's date. See Carried values.
- Changes inside the noise are not reported as changes. AI answers vary from run to run even when nothing changed, and the Action Center only reports a lift larger than that variation. See Noise bands.
- Every reason a number moves is written down. New runs, filters, prompt edits, competitor names and failed queries can all move a figure. See Why numbers change.