Product

Methodology

How Queryosity counts: the four AI search engines it measures and how each is collected, what a check is, how answers are read, how each of the six metrics is calculated, how topics count, how the AI Search Audit is judged, what the coverage states on every screen mean, and what it does not measure.

What Queryosity measures

AI search visibility, also called answer engine optimization or generative engine optimization, is how AI search engines mention, recommend, rank, cite, and portray your brand when people ask the questions your customers ask. Queryosity measures it by asking those AI search engines your prompts and reading every answer.

The numbers describe what the AI search engines answered. They are not counts of real users, traffic, or clicks, and every figure links to the answers behind it.

The four AI search engines

Each AI search engine is its own filter and its own line on every chart. An overall figure pools the AI search engines you select.

  • ChatGPT with web search, ChatGPT with web search switched on. It searches the web and cites its sources.
  • Gemini, Google's assistant.
  • Google AI Overviews, the AI summary Google shows above some search results. Google does not show one for every search. A check where it showed none still counts, as an answer without your brand, and how often an Overview appeared is shown beside it.
  • Google AI Mode, Google's conversational search, measured separately from AI Overviews.

Every plan asks these four. Enterprise adds Perplexity, asked through Perplexity's API with web search rather than read from the Perplexity app. Its answers are labeled API wherever they appear; the next section explains the difference.

Prompts are asked in English, from the market you choose for each brand: the United States, Sweden, or India. Claude, with the customer's own Anthropic key, is in development.

How each AI search engine is collected

Each AI search engine is collected in one of two ways, and every answer records which.

  • From the product's interface. The question is asked in the product, as a person would ask it, and the answer is captured as the product shows it, with the sources it shows. ChatGPT with web search, Gemini, Google AI Overviews, and Google AI Mode are collected this way.
  • Through the model's API, with web search. The question is sent to the company's own model, which searches the web and answers. This is the model behind the product, not the product's interface: it is not signed in, has no history, and does not use the options a person can choose in the app. Perplexity, on Enterprise, is collected this way. So will Claude, which is in development.
How each AI search engine is collected
AI search engineHow it is collectedModel and searchMarket
ChatGPT with web searchFrom the interfaceChatGPT with web search switched onThe market you choose
GeminiFrom the interfaceGemini as it answers by defaultThe market you choose
Google AI OverviewsFrom the interfaceA Google search on a desktopThe market you choose
Google AI ModeFrom the interfaceGoogle's AI ModeThe market you choose
Perplexity, on EnterpriseThrough the model's APIPerplexity's Sonar model, with web searchTold the market you choose; it reads the same web pages in every market

What an answer asked through an API can tell you, and what it cannot:

  • It can show which brands the model names and recommends for the question, how it ranks them, and which pages it cites, asked the same way every day. That makes it a steady signal to follow over time.
  • It cannot show exactly what a signed-in person sees in the app, where their history, plan, and settings can change the answer. Perplexity also reads the same web pages whichever market it is asked from: its answer is written for the market, but its sources are not local.

The two ways are never pooled without saying so:

  • Every answer asked through an API carries an API badge wherever its AI search engine is named.
  • The AI search engine filter shows the two apart.

The trend charts mark the day an AI search engine is added and the day the way it is collected changes. A change against the previous period leaves out any AI search engine that was not tracked in both periods, and says so.

What a check is

A check is one prompt, asked in one AI search engine, in one market, on one day. The screens count checks as answers. Paid plans run every check once a day, overnight. The free report runs each check once.

A check ends in one of three ways:

  • Successful: the AI search engine answered, and the answer was captured. A Google search that showed no AI Overview is a successful check with no AI answer.
  • Failed: something went wrong collecting it. A failed check is left out of every metric and counted in coverage instead. It is never counted as your brand being absent.
  • Not yet collected: the day's checks are still running. They are shown as collecting, never as zero.

AI search engines can answer the same prompt differently from one day to the next. One check a day for each prompt, AI search engine, and market shows how the answers move over time rather than resting on any single reply.

How answers are read

A language model reads every captured answer against a written rubric. For each brand you track, it records whether the answer mentions it, whether the answer recommends it, its position if the answer ranks options, and the sentiment, each with the excerpt from the answer that supports it. The details of every answer show these labels, the model that gave them, and the rubric version.

  • Checked before they count. The labels must follow the rubric's structure, and each excerpt is checked against the answer: one that does not appear in it word for word is flagged on the answer. A response that fails is asked for once more, and an answer still not labeled is counted as not read, never guessed.
  • Doubtful cases are marked. When the model marks a judgment as uncertain, the answer's details say so.
  • Prompts are read too. When you add a prompt, its intent is worked out, which decides whether it counts toward the recommendation rate. You can change it, and a model never overwrites your choice.
  • You can report a label. If a label looks wrong, you can report it from the answer's details. Reports go to a review queue, and the answers reviewed join a set of answers whose labels are settled.

The six metrics

Every metric pools the checks in the scope you select, by dates, AI search engines, markets, and prompts, and gives each check equal weight. An overall figure is therefore not an average of the AI search engines' figures, and the coverage shown beside each metric says how complete its scope is.

Visibility rate

The share of successful checks whose answer mentions your brand.

Visibility rate =answers that mention you ÷ answers read

  • Your brand counts once per check, however often the answer names it and whatever the tone. Recommendation and sentiment are separate metrics.
  • Answers that mention no brand at all count, as answers without yours.
  • Failed checks are left out. If no check has succeeded, the metric shows Not available rather than 0%.

For example, 250 scheduled checks, of which 200 succeeded and 40 mention your brand, give a visibility rate of 20%, with 200 of 250 checks complete.

Competitive share of voice

Your brand's appearances as a share of all appearances by your brand and the competitors you chose. The dashboard calls it Share of voice.

Competitive share of voice =your mentions ÷ mentions of every tracked brand

  • Each brand counts at most once per successful check, whatever the tone, and several brands can appear in the same answer.
  • Every brand is counted over the same checks, with the same filters, so the comparison is fair. Brands you do not track are not counted: this is share of voice among your selected competitors.
  • With no competitors configured, the metric asks you to add some. If no tracked brand appears at all, it shows Not available. If competitors appear and your brand does not, it shows 0%.

For example, 40 appearances for your brand, 35 for one competitor, and 25 for another give 40 ÷ (40 + 35 + 25), a share of voice of 40%.

Recommendation rate

The share of assessed checks for Choosing, Comparing and About your brand prompts in which the answer explicitly recommends your brand.

Recommendation rate =answers that recommend you ÷ answers read to buying and comparison prompts

  • Only prompts that ask for a choice count: buying, selecting, comparing, or finding alternatives. Each prompt's intent is set before any answer is read, and informational prompts are left out.
  • An answer recommends your brand when it endorses it or presents it as a suitable choice for the need the prompt states. A neutral mention, a description, criticism, or advice to avoid it is not a recommendation. A recommendation with conditions counts only when the conditions match the prompt.
  • Checks where your brand is absent, where only competitors are recommended, or where no brand is recommended count as not recommending you. Several brands can be recommended in one answer, so the rates across brands need not add up to 100%.
  • Assessments that could not be made are left out and shown in coverage. With no assessed checks for eligible prompts, the metric shows Not available.

For example, 20 recommending answers among 100 assessed checks for eligible prompts give a recommendation rate of 20%.

Average recommendation position

Your brand's average place in the ranked recommendation lists that include it.

Average recommendation position =sum of your list positions ÷ answers that rank you

  • Only a list whose order expresses preference counts, such as a top five or a ranked shortlist. A numbered list without ranking framing, an alphabetical list, steps, or a set of options is not a ranking.
  • Each check contributes at most one position: your brand's first place in the answer's main ranked list. Positions are kept as captured, including places held by brands you do not track.
  • Checks where your brand is absent or not ranked are left out. Absence is never counted as last place: visibility and recommendation rate measure it instead.
  • The number of ranked appearances behind the average is shown with it. With none, the metric shows Not available.

For example, positions 2, 4, and 3 in three checks give an average position of 3.0, based on three ranked appearances.

Website citation rate

The share of AI answers with observable citations that cite at least one page of your website.

Website citation rate =answers that cite your site ÷ answers with observable citations

  • An answer counts once, however many of your pages it cites, and a citation counts even when the answer does not name your brand.
  • Only the sources the answer itself cites count, not links mentioned in its text or ordinary search results beside it.
  • Answers that cite nothing count, as answers that did not cite you. Answers whose citations could not be observed are left out, and so is a Google search with no AI Overview, which has no AI answer to cite anything.
  • Only the website you configured counts as yours. Retailers, marketplaces, social profiles, and similar domains do not.

For example, 15 of 100 answers with observable citations citing your website give a website citation rate of 15%.

Sentiment distribution

How the answers that mention your brand portray it: the shares that are positive, neutral, mixed, and negative.

Sentiment distribution =answers in each tone ÷ answers about you that could be read

  • Positive: a favorable evaluation. Neutral: a factual description, without evaluation. Mixed: substantive favorable and unfavorable points about your brand in the same answer. Negative: unfavorable on balance.
  • Each answer gets one category for your brand, judged on everything it says about your brand, not on its tone about the category or about other brands. Sentiment is separate from recommendation: an answer can recommend your brand and be mixed about it.
  • Answers that could not be read are counted apart, outside the four shares, and never as neutral.
  • The charts show three tones: positive, neutral, and negative, with the mixed answers in neutral. Each answer keeps its own category wherever it shows on its own, mixed included.

For example, 50 positive, 30 neutral, 10 mixed, and 10 negative answers give 50%, 30%, 10%, and 10%; the charts show 50% positive, 40% neutral, and 10% negative. If 10 more answers mentioning your brand could not be read, the metric shows 100 of 110 assessed.

Combined score and rank

One figure to rank brands by: your visibility rate, recommendation rate and website citation rate, added up and divided by three.

Combined score =(visibility + recommendation rate + citation rate) ÷ 3

  • Every brand you track is scored over the same answers, with the same filters, and over the same rates: the ones your brand has a figure for. If no answer in the scope has citation data, for example, every brand's score is the mean of the other two.
  • A brand without a figure for one of those rates gets no score and no place, rather than a score from fewer rates.
  • The rank orders your brand and the competitors you track now by combined score, highest first. Brands with the same score share a place: two brands level in second place are both second, and the next is fourth.
  • A change in places compares the previous equal period like for like: both periods over the AI search engines tracked in both, and over the rates both have.

For example, a visibility rate of 80%, a recommendation rate of 70% and a citation rate of 30% give a combined score of (80 + 70 + 30) ÷ 3 = 60.

Topics

Each prompt carries up to three topic tags, such as Pricing or Integrations. Choosing a topic in the filter bar narrows every view to that topic's checks: the metrics and their trends, the competitors, the sources, the prompts, and the answers.

  • A prompt in several topics counts fully in each. A check of a prompt tagged Pricing and Integrations counts in both, so the topics' figures don't add up to your brand's total.
  • Topics are as they were when the check ran. A check counts in the topics its prompt carried when it was asked. Changing a prompt's topics saves a new version of it, as changing its wording does: earlier answers keep the topics they were asked under, and no past day's figures change.
  • Coverage works the same way. Under a topic, as under any other filter, failed checks count in coverage and not in the metrics, and a topic with nothing to count says so rather than showing a zero.

Pages read but not cited

ChatGPT and ChatGPT with web search also report the pages they read for an answer, whether or not they cite them. Queryosity keeps that list beside the citations: the Sources view counts, for each page, the answers that read it and how many of them did not cite it, and lists the pages that were read in the selected dates but never cited. Answer details shows the pages an answer read and did not cite.

  • A page read but not cited is evidence only. It counts in no metric: the website citation rate counts only the pages an answer cites.
  • Gemini, Google AI Overviews and Google AI Mode report only the pages they cite, so for them this is not available, never zero. Each count says how many answers it rests on.

AI Search Audit

Every report checks your site for what lets AI search engines read and use it, on the pages it audits: the free report's ten, and on a paid plan up to your plan's limit every week. There are six checks, each Pass, Needs work, or Failing, with one fact behind it. The first five read the audited pages; the sixth reads the answers.

  • AI crawlers can read your site. robots.txt lets the answer bots of OpenAI, Anthropic, Perplexity, Google, and Microsoft fetch the home page and the audited pages. Training bots are shown and never counted. robots.txt is a request, not a lock: a firewall can still turn a bot away.
  • Pages readable without JavaScript. At least 30 percent of a page's visible text is in its HTML before any script runs, since AI crawlers rarely run scripts.
  • Structured data on key pages. A page's JSON-LD parses, declares a type suited to the page, and has the properties that type requires. A page whose JSON-LD only lacks properties needs work; it never makes the check fail.
  • Pages that answer buyer questions. How many of your tracked Choosing, Comparing and About your brand prompts an audited page answers, judged by a model that must quote the passage that does. Fewer than a third is failing. When you track no such prompt, the check reads whether your product, pricing, and help pages carry questions and answers for buyers.
  • Content kept fresh. Articles and help pages show a date within the last 12 months, or are evergreen.
  • Listed where AI looks. Whether your brand is on the review and forum sites AI cites most for your prompts: the five cited by the most answers in the 30 days before the audit, by the kind of source each site is. Your brand is on a site when a page cited from it names your brand in its title, or, as weaker evidence, when the answers citing it name your brand; each site's row says which. It passes on 4 or 5 of the 5, needs work on 2 or 3, and is failing on 0 or 1, in proportion when fewer qualify. With fewer than three review or forum sites cited, it is not applicable. It reads only answers already collected: nothing is fetched for it.

How each of the first five is judged:

  • Pass: every page it judged passes, or, for AI crawlers, every answer bot may read every page checked.
  • Failing: it fails on the home page, or on half or more of the pages it judged.
  • Needs work: anything in between.
  • Not measured: fewer than half of the pages it applies to could be judged, for example when robots.txt or the pages could not be read, unless the home page already fails. A page whose result is unknown is left out of the judgement and never counts as a pass.
  • Not applicable: nothing it judges is among the audited pages, such as a site without articles or help pages. It is left out of the summary, which then reads, for example, 3 of 5 passed.

Actions

Actions turns what Queryosity observes into a short, ranked list of things your brand can do to appear in more AI answers, on every paid plan. An action is an observation with its evidence: the answers it rests on, which you can open and count, and the audit checks behind it. It is ranked by how much evidence it has, never by a forecast.

The principle: Queryosity ranks and evidences; it never writes your pages and never publishes for you.

What an action is built from

Every action is assembled from data Queryosity already stores, in fixed wording around its counts. No model writes an action.

  • Technical: something keeps AI from finding or reading your site, from the latest AI Search Audit: AI crawlers kept out by robots.txt, pages that do not load or show their text only after JavaScript runs, structured data missing or incomplete, errors in robots.txt, or no sitemap.
  • Content: a tracked prompt you are losing. Its answers name a tracked competitor and not you, cite a competitor's page and none of yours, or stopped citing a page of yours. What the audit knows about whether one of your pages answers the prompt sets the first step.
  • Third party: a review site, publisher, reference or video site cited in answers that name a tracked competitor and not you.
  • Community: the same, on a forum or community site.
  • Prompts: the answers keep comparing you with a tracked competitor, and no prompt you track asks that comparison.

Company sites are left out: another vendor's site is evidence in Sources, not something you can act on. An action that rests on answers needs at least 3 of them, on at least 2 days, because one run is not a trend.

The free report's fixes come from its one day of answers and its 10-page audit, by the same rules. A fix there needs 2 answers, and a page no longer cited can't be one, since that takes two runs. The report shows three in full: the best Technical, Content, Third party and Community fix in turn, then the next best of those types. Prompts fixes are never among the three, and the others are listed by type and title.

How actions are ranked

The same stored data always gives the same list in the same order: nothing in the ranking is a model, random, or a forecast.

  1. What blocks AI from your site comes first, since nothing else works until it is fixed.
  2. Then the actions with answers behind them, by how many answers each touches, each answer weighed by its prompt's importance, from 1 to 5: in full at 4, a quarter more at 5, down to a quarter at 1. Until you set one, Choosing, Comparing and About your brand prompts are 4 and Learning prompts 2, so a Learning answer counts half: recommendations happen on the other three.
  3. Then prompts worth tracking, which extend what is measured rather than fix anything.
  4. Then site checks that no answer points to.

Within each group, ties go to the action whose answers name more of your tracked competitors, then to the one seen in more AI search engines and markets, then, for site checks, to the one whose failing checks weigh more across the pages. This is ranking version 2, which added importance; at the default importances it orders actions as version 1 did. A new version re-ranks the list at the next refresh and changes no status.

Reach

Each action shows its reach: the number of captured answers in its evidence, each counted once, or "Site check" when no answer points to it. Reach is a count of what was observed, never a forecast of what fixing it would bring.

When actions are refreshed

  • Every Monday, an hour after Sunday closes in your workspace's timezone, over the 28 days of answers ending that Sunday.
  • After every AI Search Audit, with the audit's new checks.

An action whose evidence is gone becomes No longer observed, and opens again if its evidence returns over the minimum.

Statuses

  • Open: found by a refresh. New marks what the latest refresh found first.
  • In progress: an Owner or Editor is working on it.
  • Done · measuring: marked done, and measured for 14 days.
  • Dismissed: set aside by an Owner or Editor, with an optional reason.
  • No longer observed: its evidence is gone.

A dismissed or done action opens again on its own only when its evidence has changed materially: at least twice the answers and at least 5 more, or a tracked competitor newly named in 3 or more of its answers; for a site check, at least twice the pages and at least 2 more, or it has started to block AI. Every change is kept in the action's history, with who made it, when, and why.

Measured impact

When an action is marked done, Queryosity compares the 14 days before that day with the 14 days after it:

  • Content: your visibility and website citation rate on its prompt.
  • Third party and Community: of the answers citing the site, how many name you.
  • Technical: your website citation rate in the AI search engines concerned, or in all of them for a site-wide check.

A change counts as clear only when the 14 days after are complete, at least 30 answers stand behind each figure, and a rate moves by at least 5 points and by at least a fifth of where it started, or an average position by at least half a place. It is an observed change, not proof of cause: AI answers change for other reasons too.

Content briefs

From a Content action, an Owner or Editor can draft a content brief: the page's goal, the questions it must answer, and an outline of headings with a note of one line each, with the answers and sources it was drafted from. A language model drafts it from the captured answers to the action's prompt and to the tracked prompts that share its topics. A brief never contains the page's text: every part has a length that leaves no room for it, and a draft that oversteps is thrown away. You write the page.

Implement

Implement drafts two technical files from data Queryosity already holds, for you to copy into your site:

  • Head tags for an audited page: its title and meta description as the audit read them, its canonical address and og:title, and an FAQ schema outline (FAQPage JSON-LD) holding your most important tracked questions, each answer left for you to write. A page with no description gets a comment asking for one, never words written for it.
  • An llms.txt for your site: your brand's name and description, the audited pages under Key pages, and your tracked questions.

No model writes either file, so the same data always gives the same files, and nothing is written into your site.

What Queryosity never does

  • Write finished page copy.
  • Publish anything, or change your site, its robots.txt, or any third-party site.
  • Contact anyone on your behalf.
  • Promise or forecast uplift: no action carries an expected-impact figure.

Ask Queryosity

Ask Queryosity answers questions about your brand in your own words, from Queryosity's data and nothing else.

How a reply is made

A language model reads your brand's data through the same views your screens use, for your plan's AI search engines and the history it keeps: visibility, prompts, competitors, the answers and their text, citations, sentiment, changes, actions, the AI Search Audit and the files Implement drafts. It can't read another brand's data, browse the web, or change anything. It writes the reply from what it read for that question.

How a reply is checked

Before a reply shows, every number, name, website and quote in it is checked against the data read for that question. A reply that holds anything the data doesn't is refused and written again once; if it still doesn't pass, you see "I can't answer that from your data" instead. Each figure carries a chip with the number of answers behind it, which opens the screen it comes from with the same filters.

Searching the answers

It searches the text of your brand's answers in two ways: by your words, and by their meaning, so an answer that says the same thing in other words is found too. A count of answers is always of the answers that use your words. An answer found by meaning alone is shown as related, read before anything is said about it, and never counted. When no answer uses your words, the reply says so, never that no answer says it.

General advice

A reply may end with advice under the label "General advice, not from your data". It holds no numbers, websites, brand names or quotes, so it can't pass for a finding.

What it can't answer

Website traffic, revenue, sales and search volume, which Queryosity doesn't measure; brands you don't track; AI search engines outside your plan; and dates outside the history your plan keeps. It says so and links to the nearest screen. It hands over your llms.txt, head tags and an action's content brief as they are, and never writes page copy.

Your conversations

A conversation is yours alone: teammates and clients don't see it. It is kept 90 days and can be deleted at any time. Text inside answers and cited pages is read as data, never as instructions.

Coverage and missing data

Every screen says how complete its data is, and never shows a zero where the truth is that something has not been collected.

  • The coverage line under each screen's title gives the last day included and how many checks are complete, for example "Data through 28 Sep 2026. 250 of 250 checks complete." It adds "still collecting" while checks in the range are running, and while a first report is being collected it counts the checks and answers as they arrive.
  • Partial: when the dates you selected include checks that have not completed, a Partial badge appears and every metric shows its own coverage.
  • Not available: a metric with nothing to count says so, with the reason, such as no checks having completed yet, no tracked prompt with buying or comparison intent, or no mention that could be judged for tone.
  • Failed checks are counted in coverage and left out of the metrics, so a failure never reads as your brand being absent.

What Queryosity does not measure

  • Traffic. It does not count visitors, clicks, impressions, or how often people ask a question. It measures what the AI search engines answered.
  • Personalized answers. Answers are captured for a fresh question from the market you choose, not as a signed-in person with a history, so someone's own answers can differ. Perplexity's are asked through its API, not its app.
  • Other assistants, languages, and markets. It measures the four AI search engines above, with prompts in English, from the United States, Sweden, and India; Claude is in development.
  • Why an answer says what it says. It reports what the answers contain and cite, not what caused them. Actions suggest what to change from that evidence, and measured impact compares before and after without claiming a cause.
  • Classic search rankings. A position is a place in an AI answer's ranked list, not a ranking in ordinary search results.