One Prompt, 42 Answers: What Prompt Insights Shows That a Single AI Answer Cannot

Short answer: Prompt Insights is a page for each tracked prompt that pools every successful stored answer in a date range, across the models you select, and counts each answer once. The Brands tab shows which brands recur, in what share of answers, and how early they are named. The Search queries tab lists the searches the models ran while answering, with a separate list of recurring phrases derived from them. The Sources & pages tab shows which sites the models read, how often your brand was named when a given site took part, and which of your own pages the models opened. The AI answers tab is the dated archive of every run, with Previous run and Next run navigation and links that reopen an exact answer. It shipped in Ayzeo 15.53.0, the Sources & pages tab followed in 15.54.0, and Pro and Enterprise customers can query the same data from an AI assistant through the get_prompt_insights MCP tool.

1. Why the Latest Answer Is a Sample of One

The numbers on your project dashboard (Mention Rate, Visibility, Citations) are built the way dashboard numbers should be: the latest run of each prompt on each model in the selected range, reduced to a figure you can track week over week. That is the right design for "how are we doing" and the wrong design for "why".

Ask an AI model the same question fourteen times in a month and you will not get the same answer fourteen times. The brand list shifts, the order shifts, and so do the searches and pages behind the answer. Read one answer and conclude "ChatGPT names us third", and you have learned about Tuesday. Read all fourteen and find you were named first in twelve, and you have learned something about the prompt.

Prompt Insights exists for the second kind of reading. The dashboard keeps its snapshot metrics; this is a second lens, pointed at one prompt.

2. The Worked Example: 42 Answers to One Prompt

Ayzeo's public demo project tracks HubSpot. One of its prompts is "What are the best CRM platforms for small businesses?". Over the last 30 days that prompt has 42 stored answers, 14 each from Google AI Overview, ChatGPT and Claude. Every number in this post is read from that page on 18 September 2026. HubSpot is our own demo project, not a customer, which is exactly why we can publish the figures.

The summary tiles read: Past AI answers 42, Your brand's Mention Rate 100% (42 of 42), Recurring brands 26, Search queries 37 distinct captured queries with query capture available in 25 of 42 answers, plus 128 recurring phrases derived from them. One line under the tiles, "Data coverage and model breakdown", spells out the denominators behind everything else on the page; section 7 comes back to it.

3. Brands: Who Recurs, and How Early

The Brands tab lists every brand named in the answers, counted once per answer, with its share, its Avg. mention position (the order it was named inside the answer, where 1 is first) and the models it appeared on. Your own brand is flagged "You". Context-only mentions, a brand referred to in passing rather than presented as an option, are excluded and stay visible in the full answer.

For this prompt, the top of the table reads:

Three things a single answer could not have told you.

The field is stable at the top and a coin flip below it. HubSpot, Pipedrive and Zoho CRM appear in almost every answer. Freshsales is in two thirds, monday CRM in exactly half. A single answer says either "monday CRM competes here" or "monday CRM does not", and both readings are wrong: it shows up every other time.

Share and position are different questions. Pipedrive appears in one more answer than Zoho CRM (40 against 39) but is named later when it does: Avg. mention position 3.3 against 2.7. HubSpot's 1.3 says it is usually the first brand named. Freshsales at 4.8 and monday CRM at 5.6 are on the list but well down it, and a buyer reads the top of a list differently from the bottom.

The models split Salesforce in two. "Salesforce Starter Suite" is named in 26 answers, plain "Salesforce" in 14. If you sell under a product name and a company name, this table shows which one the models use for your category.

Select any row and the matching answers open in a side drawer.

4. Search Queries: What the Models Looked For

While composing an answer, the models search the web, and this tab shows those searches in two lists.

Exact queries are the searches themselves, grouped ignoring case and spacing, each counted once per answer: 37 distinct queries across the 25 answers where query capture is available. The most frequent is "best CRM platforms small businesses 2026", in 11 answers, 44% of those 25. The prompt never mentions a year. The models add one.

Further down it gets specific: "site:hubspot.com/pricing sales hub pricing free CRM" and "site:zoho.com/crm/zohocrm-pricing.html Zoho CRM pricing 2026". A model asked for the best CRM for a small business goes and reads pricing pages, on the vendor's own domain, with a site: operator. If you wanted one page to keep accurate and current for this prompt, the list has just told you which one.

Recurring phrases are two and three word phrases derived from those queries that appear in at least two answers: 128 here. They are analysis output for spotting themes, not searches anyone ran, and they overlap. Do not add them to the query count.

Both lists export as CSV, every row in the selected period, model and search, not only the 100 shown. Selecting a query or phrase opens the answers it appeared in.

5. Sources & Pages: What They Read, and Where You Were Named

The Sources & pages tab, added in 15.54.0, holds three tables.

Consulted sources lists the domains that took part in the answers: 105 distinct domains across the 39 answers with stored search results. The top rows:

The most-read site for "best CRM platforms for small businesses" is a competitor's own domain, salesforce.com, in 87% of the answers that stored their sources. techradar.com is in 72%, reddit.com in half. HubSpot's own site is in 14 of 39, roughly a third, and HubSpot is still named in every one of the 42 answers. Being read and being named are different events, and this table shows them side by side. A row like slack.com, in 17 answers to a question that is not about Slack, deserves a read of the answers before you decide what it means.

The "You named" column shows, for each domain, the share of that domain's answers in which your brand is named, against the prompt's own rate in the line above the table. On this prompt it is flat: 100% on every row, because HubSpot is named in all 42 answers. That is what a dominant brand looks like, and here the column has nothing to add. It earns its place for a brand that appears in some answers and not others: if your prompt-wide rate is 40% and answers where reddit.com took part name you 70% of the time, that row is worth investigating. It is co-occurrence, not causation: the two happen together, which does not mean the source produced the mention. It stays blank for a domain seen in a single answer.

Your pages the models opened lists your own URLs the models opened. For HubSpot the most-opened page is hubspot.com/pricing/crm, in 4 answers. Set that next to the pricing searches in section 4 and the brief writes itself.

Pages opened is the full list: 58 distinct URLs, with opened pages available in 25 of 42 answers. Google AI Overview never reports opened pages, so its 14 answers count towards consulted sources only. "Available in 25 of 42" describes what the data can show, not a count of answers that opened nothing.

6. AI Answers: The Archive Behind Every Number

The last tab is the archive: every run in the selected range and models, newest first, 20 to a page, with the run time in UTC, the model, whether your brand was named, the brands, the captured queries and the answer. Open a run and you are reading that exact stored answer; Previous run and Next run walk the archive in date order.

Two details matter in practice. Share copies a link that reopens that exact run, with the model filter written into it, so a colleague sees the same answer in the same scope. And selecting a brand, query or phrase on the other tabs never filters this archive: matching answers open in a separate drawer, and Back to matching answers returns you to that list. The archive follows the date and model filters at the top of the page and nothing else.

7. Reading the Coverage Line

"Query capture available in 25 of 42 answers." Lines like this appear on every tab, and they are the reason the page's percentages can be trusted.

Not every answer carries every field. Some answers cannot expose the searches behind them, and Google AI Overview never reports the pages it opened. When a field is unavailable for an answer, that answer is left out of the denominator for that field, and the page says so. The top query at 44% means 11 of the 25 answers with query capture, not 11 of 42. Computed over all 42 it would read 26%, look just as real, and be wrong.

The rule: a missing field is unknown, never zero. "Data coverage and model breakdown" under the summary tiles has the exact figures: answer text, extracted brands and query capture, each as a count of runs, plus runs per model.

8. A Ten-Minute Routine per Prompt

For the prompts that matter, once a week:

1. Brands: look for movement, a competitor crossing 50% or your own Avg. mention position drifting later.
2. Search queries: read the top five exact queries and ask whether a page of yours answers each one. Note any site: search that points at a competitor.
3. Sources & pages: check whether your domain is read at all, look for a "You named" row above your prompt-wide rate, and confirm which of your pages the models open. Read the answers behind any row that surprises you.

If you change a page, come back after the next runs and read the same three tables. The archive keeps every run, so the before and the after are both there.

9. The Same Data for Your AI Assistant

On Pro and Enterprise, the get_prompt_insights MCP tool returns the same aggregation for one prompt: search queries, recurring phrases, recurring brands with your own brand flagged, the coverage summary, runs per model and the paged answer archive. Ask your assistant "what do the models search for when they answer our small-business CRM prompt?" and it reads the answer from Ayzeo instead of guessing. The payload names its denominators (shareOfAnswersWithQueryCapture), so a share of captured answers cannot be mistaken for a share of every run. Setup: the MCP help article.

10. Frequently Asked Questions

Q: How is Prompt Insights different from the prompt's numbers on the dashboard?

A: The dashboard reads the latest run of each prompt on each model in the selected range and reduces it to Mention Rate, Visibility and Citations. Prompt Insights reads every successful stored run of one prompt in the range, pools them across the models you select and counts each answer once. The two answer different questions, "how are we doing" and "why", and neither changes the other.

Q: Does opening Prompt Insights run the prompt again or use up AI queries?

A: No. The page reads answers that are already stored. Nothing runs, no AI query is spent, and the Refresh button only reloads the stored data. New answers appear as your scheduled runs add them.

Q: Who can open Prompt Insights?

A: Every member of a project, viewers included, from the Insights button on a prompt card, the same button on the Prompt Performance cards of the analytics page, or the Prompt insights link inside any answer. The get_prompt_insights MCP tool follows the MCP server and needs a Pro or Enterprise plan.

Q: What is a recurring phrase, and why is it not a search?

A: Exact queries are the searches the models ran while answering. Recurring phrases are two and three word phrases derived from those queries that appear in at least two answers. They are analysis output for spotting themes, they overlap with each other, and nobody typed them, so they must not be added to the query count.

Q: What does the "You named" column mean?

A: For each consulted domain it is the share of that domain's answers in which your brand is named, shown against your prompt-wide rate. It records that the two occur together; it does not say the source caused the mention. It stays blank for a domain seen in a single answer.

Q: Why does the Sources & pages tab count fewer answers than the Brands tab?

A: Because not every answer stores every field. Sources are shown for the answers that stored search results, opened pages for the answers that reported them, and each share uses its own denominator. An answer without a field is treated as unknown, not as zero.

Q: Why are there no opened pages for Google AI Overview?

A: Google AI Overview never reports the pages it opened, so its answers count towards consulted sources only. The page says so wherever it matters. Treat the absence as unavailable data, not as "opened nothing".

Q: Why is Avg. mention position missing for some brands, or for the whole prompt?

A: A position is only computed for answers with a stored mention rank, and a prompt that names your brand in its own text is excluded from position metrics entirely, because the order of mention says nothing when the question already contains the brand. The table shows how many ranked answers each position is based on.

Try It on Your Own Prompts

Prompt Insights is on every project today: open any prompt card and select Insights. If a prompt matters to your business, the 42 answers behind it will tell you more than the one on top.

Key Takeaways

• One answer is a sample of one: Prompt Insights pools every stored answer to a prompt in the period and counts each once; the dashboard keeps its snapshot metrics.
• Share and position are different questions: a brand can be in 95% of answers and still be named third; Avg. mention position shows where in the answer it sits.
• Exact queries are real searches; recurring phrases are derived: export the queries as CSV and check that a page of yours answers each of the top ones.
• "You named" is co-occurrence, not causation: read it against your prompt-wide rate, and expect a flat column for a brand that is named everywhere.
• A missing field is unknown, never zero: every share uses the answers that actually carry that field, and Google AI Overview never reports opened pages.