How Parse measures
Parse asks AI what buyers ask, and counts who it mentions.
- 1Buyer questions
- 2Ask ChatGPT + Google AI Mode
ChatGPT
Google AI Mode - 3Read every answer
- 4Mentioned in =answers mentioning you / all answers
What Mentioned in means
The percentage of tracked answers that mention your brand. Your Overview includes questions about you and your rivals. Select a market to see its results.
How often
Pro updates daily. Free updates on Mondays. Workspace visibility covers 7 days. Public rankings and Mentioned in cover up to 180 days; graphs over time plot 30-day points. Each page shows its actual coverage.
What’s included—and what isn’t
Parse measures answers to the questions it asks. It does not access private chats, reproduce each buyer’s location-specific results, or cover tools it doesn’t query.
We ask questions in English with the location set to the United States.
What is counted, and what is written
Counted
Shares, ranks and the quotes from AI answers. Counted or copied from the answers Parse collected.
Written by AI
A brand’s description, a market’s opening lines, a topic’s verdict, a question’s summary, a comparison’s verdict. Written by an AI model from those answers and marked AI summary. A brand’s description is left out when it describes a different company.
More detail
Where every Parse number starts
Parse observes how public AI assistants answer tracked buyer questions, then extracts the brands, citations, and sources they use.
Questions
14K+Buyer questions indexed
Responses
4.1M+AI responses observed
Brands
800K+brands tracked
Citations
45M+citations analyzed
Where questions come from
Parse organizes buyer questions by market and topic, including generated wording. These questions define the coverage of the index.
- Questions about choosing, comparing and using products
- Tracked question coverage is not measured buyer demand. Parse does not read private AI conversations.
From an answer to a published measurement
Parse maintains a dataset of observed answers, the brands they name and the pages they cite. Each published result has a defined question set, collection window and counting rule. The details below explain what those choices mean for the numbers you see.
Collection, matching and quality checks
- Collection and scope
- Answers are associated with the question, engine and collection time. A research report states its own sample and dates; these can differ from a current brand or market page. A collection window does not mean every question was checked every day.
- Repeated observations and URLs
- Repeated answers are useful evidence of change over time. They are different from repeated mentions within one answer: a brand named twice still counts as one answer in its Mentioned in. Citation studies state whether they count links, distinct pages or websites. URL cleaning and hostname grouping are specified in each study; subdomains are not automatically one publisher.
- Brand matching
- Names, known naming variants, domains and answer context help connect a mention to a brand record. Similar names can refer to different companies or products, so a text match alone is not proof that two identities are the same. Matching errors can affect the counts and can be reported from the page.
- Exclusions and missing evidence
- Research describes its inclusion rules, such as organic questions, matched engines or citation-bearing answers. A study of cited answers cannot measure how often all answers contain citations. Missing observations are not zero mentions. Low-sample results show their percentage in a lighter shade, or withhold a comparison, and never print a count instead.
- Checks and interpretation
- Structured publication checks validate the shape of the data; metric rules define denominators and minimum evidence. Research methods explain additional checks, including alternative counting rules where tested. These checks do not make every AI statement true: a citation is not proof of support, and a mention is not a product endorsement. Research downloads provide aggregate figures where available so readers can inspect the reported results.
Models we query
| Model | Coverage | Refresh |
|---|---|---|
| ChatGPT Search | Every topic | Daily in markets a Pro brand is tracked in, and for a Pro brand’s Custom topics; about every 9 days in other markets; weekly for the topics a Free brand adds and, until AI first mentions the brand, the ones its check found |
| Google AI Mode | Every topic except the ones a Free brand adds or its check found | Daily in markets a Pro brand is tracked in, and for a Pro brand’s Custom topics; about every 9 days in other markets |
What we extract from each response
- Brand mentions
- Brand names identified in AI answers and matched to known brands. We can miss a name or match it to the wrong brand.
- Recommendation rank
- The order in which brands are listed when AI is asked who's best.
- Citations
- Outbound URLs the AI answer points to.
- Sources
- Root domains behind those citations.
What each number is measured over
Every figure on a Parse page has a denominator. This chapter names them, so a percentage on one page can be compared with a percentage on another.
The basis line, part by part
Example evidence line
Based on 12,431 AI responses from ChatGPT Search and Google AI Mode across 1,204 questions
- Based on 12,431 AI responses
- Every individual answer Parse collected for this market in the period. One question asked on two platforms produces two responses, so responses usually outnumber questions. This is the denominator for Mentioned in.
- from ChatGPT Search and Google AI Mode
- The answer platforms behind those responses, on a public page and on Pro alike. Retrieval by AI search APIs is reported separately and is not in this count.
- across 1,204 questions
- The distinct buyer questions behind those responses. Evidence floors are checked against question count, not response count, so one question answered many times does not on its own clear a floor.
Which numbers are shares, and of what
| Number | Counts | Out of | Sums to 100% |
|---|---|---|---|
| Mentioned in | Responses that mention the brand | Every response collected for that market in the period | No |
| Recommendation share | Recommendation slots the brand occupies | Every recommendation slot in that market | Yes |
Recommendation share is the only figure above that is a share of a fixed total. Mentioned in and share of voice describe the same measured quantity, so several brands in one market can each be mentioned in a large fraction of answers and the column will not add to 100%.
How often numbers update
Collection, public pages and Pro run on different clocks. The refresh column in chapter 01 reports the first of the three.
- Collected Daily
- Answers are collected every day, but not every topic is asked every day. A topic in a market a Pro brand is tracked in is asked daily; other topics come round about every 9 days, and topics that have gone quiet less often. The refresh column above reports this.
- Public pages
- Each public page shows the dates its numbers cover. Rank and Mentioned in cover up to 180 days, and a change compares each with its value 30 days earlier. Read each page’s actual coverage dates; newer questions may have less history.
- Pro Daily
- Pro asks its questions and updates its numbers every day: Mentioned in over 7 days, and its markets’ rank over the same window of up to 180 days as the public pages.
How the evidence connects
We are building a single connected map of buyer topics, questions, brands and web pages across AI platforms and response types. It brings together what assistants say, which pages they cite and which pages separate web-search providers return. These connections help us study the same brand or topic from different sources of evidence.
Each observation keeps its question, source and collection date. Our regular collection measures ChatGPT and Google AI Mode. Historical datasets can add answers from other engines; their model labels and dates are reported by the original publisher. Search-provider results are a separate kind of observation, with their own coverage and denominator. Connecting them does not turn a search result into an assistant mention or citation.
Historical source data: Bikoosh Answer Watch and Product.ai AI Shopping Divergence Study, used under CC BY 4.0. Parse selects eligible questions and derives its own brand and citation measurements from the original answer text. Product.ai dates identify collection batches, not independently verified times for each answer.
Our collection system supports Parallel Search, You.com Search, Tavily Search, Exa Search, Linkup Search. A page returned by one of these providers does not show that ChatGPT or Google AI consulted it. Published pages identify the engines, dates and sample behind each measurement; the map is an ongoing development, not a claim of complete coverage of every AI system.
Pages found
Retrieval is the share of a brand’s measured questions for which at least one third-party web-search API returned a page on that brand’s domain in its top 10 results, for the question asked as written. Real AI assistants rarely search a question verbatim — they rewrite and decompose it — so retrieval is not an answer, and it is not a reading of what any assistant retrieves.
These are search APIs sold to agent builders (Parallel Search, You.com Search, Tavily Search, Exa Search, Linkup Search), not the retrieval stacks inside ChatGPT or Gemini. Retrieval is not part of any figure here and never moves a rank.
We report the share and collection date, but not a trend: we haven’t established how much these results vary between checks. Each eligible question comes round once every 30 days, and a Pro brand’s Custom topics once every 7.
As of Oct 27, 2026.
Evidence floors and precision
We show a comparison result only when it has at least 3 claims across at least 2 distinct questions.
- Head-to-head side scores
- Rounded to a whole number. The comparison card deliberately shows a coarser figure than a brand page, because far fewer claims sit behind it.
- Large counts on summary cards
- Abbreviated to one decimal in thousands or millions with a plus, for example 12.4K+. Exact counts appear in the basis line instead.
- Shares of fewer than 20 answers
- A public page prints the percentage in a lighter shade, with no bar or plot. The workspace shows no number until there are enough answers. A share never prints as a count of answers.
- Suppressed values
- A number that fails an evidence floor is not rounded to zero. It is not rendered at all.
How to read a head-to-head comparison
See which brand AI favours overall, where each stands out and how the comparison differs by AI engine.
What we count
A claim is one comparison statement in an AI answer. An answer can contain several claims. Comparison pages count claims; market pages count whole answers.
The overall bar groups claims into those that favour the first brand, those that favour the second and those that favour neither.
Where each brand stands out
For each attribute, the table names the brand that more claims favour and shows its share of all claims about that attribute. If both brands have the same count, it shows “Tied” and the share for each. A brand can lead with less than half the claims because some favour the other brand and some favour neither.
Comparing AI engines and sources
Engine rows show the claims from each AI engine. Source rows show claims that cite each source. They describe what AI said when citing that source, rather than the publisher’s own preference.
Percentages are rounded to whole numbers. Check the dates and claim count on the comparison page for its scope.
Core metrics
Definitions for the metrics you’ll see in Parse.
Brand metrics
Metrics that measure a brand's AI visibility and performance in AI answers.
Mentioned in
How often AI mentions your brand.
More detail
The share of analyzed AI answers where the brand is mentioned, over a defined set of questions, such as a market's buyer questions on a ranking page. It is a rate, not a split — every brand in a market can be mentioned often, and the rates do not sum to 100%. 15% means the brand appears in about 15 out of every 100 answers analyzed.
Metric type: Rate (percentage)
Market rank
Your rank in this market.
More detail
Where the brand sits in this market's published ranking, ordered by Mentioned in over the market's measured questions. Other surfaces can quote a family-level standing over the market's ranked brands, which counts a brand family as one entrant. The two answer slightly different questions.
Metric type: Rank (lower is better)
Rank change (30d)
How the rank moved over the last 30 days.
More detail
Compares the brand's 180-day rank with its 180-day rank 30 days ago. Useful for spotting longer-term trends beyond weekly noise.
Metric type: Delta (signed number)
Average position
Your average position in AI answers.
More detail
When the brand is mentioned, this shows where it typically appears in the response (1st, 2nd, 3rd, etc.). Lower numbers mean more prominent placement.
Metric type: Rank (lower is better)
Mentions
Number of AI responses that include this brand during the selected period.
More detail
How many times your brand appears in the AI answers we analyzed. Compare this with Mentioned in to account for changes in the number of answers.
Metric type: Count
Citation metrics
Metrics about how AI cites and references sources.
Citations
Number of times AI models cite the domain.
More detail
How often AI answers cite this domain.
Metric type: Count
Answer metrics
What Parse reads from one answer, and how a number moved between periods.
Mentioned
Whether the focus brand appears in an individual response.
More detail
Whether this AI answer mentions the selected brand.
Metric type: Boolean
Change
Change from the previous period.
More detail
Represents the delta between the current value and the prior comparable period. Positive change shows improvement, negative values indicate decline.
Metric type: Delta (signed number)
Perception metrics
Metrics about the language and tone AI uses when describing a brand.
Sentiment mix
Share of positive, neutral, and critical or mixed language in evidence about the brand.
More detail
Sentiment mix groups the evidence language around a brand into positive, neutral, and Critical or mixed buckets so framing changes are visible without hiding ambiguous criticism.
Metric type: Rate (percentage)
Descriptor frequency
How often a specific descriptor appears in evidence about the brand.
More detail
How often AI uses each word or phrase to describe a brand. Compare periods to see which descriptions are becoming more or less common.
Metric type: Count
Descriptor polarity
How AI describes your brand.
More detail
Whether AI’s descriptions of a brand sound positive, neutral or negative.
Metric type: Score (0-100)
Data quality and validation
The method defines a number. Validation tests where that method can go wrong.
A mention, a recommendation, a warning and a cited source are different kinds of evidence. Brand extraction can confuse a product with its parent company, miss a name or select the wrong entity. Read the answer behind a result when that distinction matters.
Counts describe the questions and answers collected. They are not weighted estimates of buyer demand. A longer window can include more answers and older behavior; it does not establish representativeness or independence.
Repeated answers share questions and collection dates. A large answer count alone does not establish statistical precision. Missing runs and extraction failures also need to be distinguished from answers that omit a brand.
When comparing periods, keep questions and engines consistent, inspect missing coverage and show the size of the change. Evidence floors and strict thresholds are useful controls, but are not a validated confidence level or a product-wide false-alert guarantee.
- How well does Parse identify and classify brands? Historical error counts and a blind review of public answers, with reference-label limitations.
- How many prompts do you need? Reproducible sampling comparisons, question selection, missing observations and reporting windows.
- When does a change deserve attention? Controlled simulations of question mix, genuine shifts and shared weekly effects.
Each study states its data version, method and limits, with downloadable evidence. These studies do not establish a single current-production accuracy percentage. See Corrections for revisions and reporting an error.
