Skip to content

Checks

AI answer panel

Measures how often ChatGPT, Perplexity, Gemini and Google's AI answers name and recommend your brand, run like a poll on a fixed prompt panel, never as a rank.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick ai-answer-panel.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
AI recommendation consistency study, The consensus gap
Needs
Your brand, category, buyers and buying moments
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedOrganic visibility
SkillSEO and AI search

AI answer panel

Example output

Measure AI visibility
EngineNamedRecommended
Perplexity62%23%
ChatGPT56%18%
Comparison prompts0%0%

Measures how often ChatGPT, Perplexity, Gemini and Google's AI answers name and recommend your brand, run like a poll on a fixed prompt panel, never as a rank.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

ai-answer-panel/SKILL.md

In the folder

  • ai-answer-panel/SKILL.md
  • ai-answer-panel/references/panel.md
  • ai-answer-panel/scripts/panel_rates.py
---name: ai-answer-paneldescription: >-  Measures a B2B SaaS brand's visibility in AI answers like a poll: builds a fixed prompt panel from the company's  buying moments, runs each prompt 5 or more times in clean sessions across ChatGPT, Perplexity, Gemini and Google AI  Overviews, and reports how often the brand appears and how often it is recommended, per engine, with confidence  intervals and never a rank. Returns PANEL READY, VISIBLE AND TRUSTED, VISIBLE NOT TRUSTED, TRUSTED NOT VISIBLE,  NOT VISIBLE, NOT MEASURABLE or BLOCKED. Use it monthly, or before and after AI visibility work. Not for choosing  which cited pages to pursue (use cited-pages) or checking crawler access (use agent-readiness).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Uses an AI visibility tool's run-level  export if the reader has one, otherwise runs the reader pastes. Python 3 is optional, for the rate math.metadata:  author: SHAKEUP  version: "1.0.0"  stage: check  plugin: seo-ai-search  homepage: https://www.shakeuphq.com/skills/ai-answer-panel--- # AI answer panel Tells a company how often AI answers name it for the questions its buyers ask, and whether they recommend it, with enough runs to trust the figure. ## Purpose - One job: build or run a fixed prompt panel and report appearance and recommendation rates per engine, with a diagnosis.- Stage: check. It ends in a verdict with the rule that fired, and a severity on each gap.- It measures rates, never ranks: answer lists change on almost every run, so a position in one list means nothing.- It never reports a rate from fewer runs than the panel rules allow, and never blends engines into one score without showing each. ## Use when / not | Situation | Use || --- | --- || "Does ChatGPT recommend us?" or a monthly AI visibility report | this skill || Before and after AI visibility work, to see if it moved | this skill, same panel both times || Deciding which third-party pages to get onto | `cited-pages`, on this skill's cited sources || AI crawlers may not reach the site at all | `agent-readiness` first; a blocked site can't be cited || Checking a tactic for spam risk | `ai-search-risk` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || Brand name and aliases | yes | `## Company` brand terms; the reader's request | ask once; else `BLOCKED` || Category, buyers, buying moments, competitors | yes, to build the panel | the profile; the reader's answers | ask for them in one message; a panel without buying moments covers category and comparison only, coverage `partial` || The panel: fixed prompts with ids | no | `## AI answer panel` in the profile; a saved panel file from a past run; built in step 2 | build it; verdict `PANEL READY` until runs exist || Runs: per run or per prompt and engine | no | a connected AI visibility tool's run-level export; runs the reader pastes; runs the agent makes in clean sessions, if it can | `PANEL READY`: hand over the panel and the protocol || Last period's runs on the same panel | no | the saved results table | trend `not available` | ## Profile read From `company-profile.md`: `## Company` (name, brand terms, category, pricing), `## Product`, `## Buyers` and `### Buying group`, `## Buying moments`, `## Competitors`, `## Accounts and data` (connected tools), `## AI answer panel` (the saved panel, engines, runs per prompt, last run), and every `## Learnings` entry for `ai-answer-panel` with status `noted`, `fact` or `accepted` (including the saved panel version). No profile: ask for the brand and its aliases, the category in the buyer's words, the buyer, 3 competitors and the moments that start a search, in one message. Say that `company-profile` and `buying-moments` save this next time. ## Steps 1. **Apply learnings.** Read them first: the panel version in use, prompts the reader retired, and any accepted change to the thresholds. 2. **Build or load the panel.** If `## AI answer panel` in the profile (or a saved panel file) has one, use it unchanged (a reworded prompt is a new prompt). Otherwise build one with `references/panel.md`: 20 to 40 prompts across the 6 families (category, comparison, budget, problem, use case, integration), each written as the buyer types it, seeded from `## Buying moments`, with the reader's brand in none of them. Give each an id and a family. 3. **Collect the runs** with the protocol in `references/panel.md`: 5 or more runs per prompt per engine, clean sessions, at least 2 engines, all within 7 days. Without runs, stop here: verdict `PANEL READY`, hand over the panel, the protocol and the results table format. 4. **Record each run:** every brand named, whether the answer recommends the reader's brand for the buyer described, and every cited URL. A Google AI Overview that didn't appear is `NO_ANSWER`, left out of the appearance rate and counted as a trigger rate. 5. **Compute the rates.** With Python:    ```   python3 scripts/panel_rates.py --csv runs.csv --brand "<brand>" --alias "<alias>" --competitor "<competitor>" --prior last_month.csv   ```    Without Python, per engine: appearance rate is answered runs that name the brand divided by answered runs; recommendation rate is runs that recommend it divided by runs that name it. For a 95% interval, use the Wilson formula, or as a rough guide take the rate plus and minus 1 divided by the square root of the runs. Say which way you ran it.    | Rule | Why |   | --- | --- |   | A prompt counts on an engine only with 5 or more answered runs | below 5, one odd answer swings the figure |   | An engine gets a rate only with 10 or more counted prompts | fewer prompts measure a few phrasings |   | The panel needs 2 or more engines with a rate | one engine's view is not the market's |   | The headline rate is the mean of per-engine rates | an engine with more runs shouldn't dominate |   | A change against last period is real only when the 2 intervals don't overlap | otherwise it is the sample moving, not the brand | 6. **Diagnose.** Visible: appearance 50% or more, the point where a buyer asking is more likely than not to see the brand. Trusted: recommended in 50% or more of the answers that name it. Apply per engine and to the headline. 7. **List the gaps** with severities: `critical` when the brand is named in 0 runs of a high-intent family (category, comparison, budget) on an engine with a rate; `warning` when a family's rate is under 20%; `suggestion` when one engine's interval sits wholly below another's (an engine gap). Add competitors' appearance rates on the same runs. 8. **Pass on the cited sources.** List the domains cited most, by how many prompts cite them, for `cited-pages`. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No brand name after asking || `PANEL READY` | A panel exists or was built, and there are no runs yet || `NOT MEASURABLE` | Fewer than 2 engines with 10 or more prompts at 5 or more runs each || `VISIBLE AND TRUSTED` | Appearance 50% or more, recommendation 50% or more || `VISIBLE NOT TRUSTED` | Appearance 50% or more, recommendation under 50%: work on independent proof for your claims || `TRUSTED NOT VISIBLE` | Appearance under 50%, recommendation 50% or more: work on the sources answers draw from || `NOT VISIBLE` | Both under 50%: fix how clearly the brand is described first, then the sources | Named stops: `PANEL READY` waits for the reader to run the prompts (or connect a tool) and paste the results. `NOT MEASURABLE` reports the raw counts, labelled as counts and not rates, and asks for the missing runs. No other stops. ## Learn Append a learning to the profile when: - the panel is built or changed: a `fact` with the panel version, the prompt count per family and where the file is kept (the panel must not drift);- the reader overrides a recommended or not-recommended call on a run (a `proposal` to tighten what counts as recommended);- a period's rates are measured: a `fact` with the headline rates, per-engine rates and the date, so the next run reads the trend;- a threshold keeps misclassifying for this market (a `proposal`, applied only once accepted). Example: `- 2026-10-24 | ai-answer-panel | fact | Panel v1: 24 prompts (category 5, comparison 5, budget 3, problem 4, use case 4, integration 3), kept in the profile's AI answer panel section | Why: built from buying moments | Status: noted` After a run, update `## AI answer panel`: panel version, engines, runs per prompt, last run date and each prompt's appearance rate. Without file access, end with that section as a paste-ready block. ## Output ```Verdict: <verdict> · Coverage: <complete | partial (what's missing)> · Rates: <script | by hand>Panel: v<n>, <n> prompts · Engines: <list> · Runs: <n> answered, <n> with no AI answer · Window: <dates> Headline: appearance <n>% · recommendation <n>% (mean of per-engine rates) | Engine | Prompts counted | Runs | Appearance (95% interval) | Recommendation (95% interval) | Diagnosis | Change since last period | | Family | <engine 1> appearance | <engine 2> appearance | ... | Gaps:| Severity | Engine | Family | Finding | Competitors on the same runs: <name: appearance rate per engine>Cited most (for cited-pages): <domain: prompts, engines>Not counted: <thin cells, prompts with fewer than 5 runs>Learnings applied: <list, or none>``` For `PANEL READY`, the output is the panel (id, family, buying moment, persona, prompt), the protocol and the results table format. Then the handoff block. ## Hand off ```handoffskill: ai-answer-panelversion: 1.0.0verdict: VISIBLE NOT TRUSTEDcoverage: completefor: <brand>, panel v1, <month>outputs:  appearance_rate: 0.63  recommendation_rate: 0.19  engines: {chatgpt: 0.60, perplexity: 0.66}  critical_gaps: [comparison on chatgpt, comparison on perplexity]  cited_domains: [g2.com, review-site.example, vendor-a.example]next: [cited-pages, claim-check]open: [add Gemini and AI Overviews runs next month]``` `next` by verdict: `VISIBLE NOT TRUSTED` goes to `[claim-check, cited-pages]` (proof first); `TRUSTED NOT VISIBLE` and `NOT VISIBLE` go to `[cited-pages, agent-readiness]`; `PANEL READY` waits on the reader. ## Failure modes | Problem | Do this || --- | --- || The reader's tool gives one visibility score, no runs | Don't use the score as a rate; ask for its run-level export || A prompt was reworded mid-month | Treat it as a new prompt; the old one has fewer runs and drops out || The brand name is a common word | Match it only as the brand (capitalized, or beside the category); spot-check 10 runs by reading them || Answers name the brand with the wrong category or price | Count the appearance, and log the error as a finding for `company-profile` and the brand's own pages || The reader wants a rank ("we're #2 in ChatGPT") | Explain that lists rarely repeat; report the appearance rate and the share of runs where it came first || AI Overviews rarely trigger for the panel | Report the trigger rate; don't fill the gap with another engine's figure | ## Influenced by - SparkToro's study of AI recommendation consistency, which found lists almost never repeat across runs and argued for appearance rate across many runs instead of rank: https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/- Kevin Indig's analysis of citations across ChatGPT, Perplexity and AI Overviews, which found most cited pages appear in only one engine and argued for per-engine measurement: https://www.growth-memo.com/p/the-consensus-gap- SHAKEUP's AI visibility playbook: the 6 prompt families, clean-session runs, and the split between being named and being recommended.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

We read AI answers as a poll, not a ranking, because the same prompt returns a different list on most runs. Rates are reported per engine with an interval, and a change only counts when the intervals stop overlapping.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Build a fixed panel from your buying moments

    20 to 40 prompts across category, comparison, budget, problem, use case and integration questions. Your brand appears in none of them.

  2. 02

    Run each prompt 5 times per engine

    Clean sessions, logged out with memory off, at least 2 engines, all in one week. One answer is one draw, not a measure.

  3. 03

    Record what each answer did

    Every brand it named, whether it recommended you for the buyer described, and every page it cited.

  4. 04

    Compute rates per engine

    Its script turns runs into appearance and recommendation rates with a 95% interval, and refuses a rate from too few runs.

  5. 05

    Diagnose and list the gaps

    Named in half the runs or more means visible. Recommended in half of those means trusted. A zero on a comparison prompt is a critical gap.

  6. 06

    Pass on the cited pages

    The pages answers cite most go to the cited pages skill, ranked by how many prompts cite them.

What you need

  • Required

    Your brand, category, buyers and buying moments

    Your company profile → Your answers to one set of questions

  • Optional

    A saved prompt panel

    The panel file from your last run → Built in this run from your profile

  • Optional

    The runs, per prompt and engine

    A run-level export from an AI visibility tool → Results you paste → Runs your agent makes in clean sessions, if it can

  • Optional

    Last period's runs on the same panel

    Your saved results table

What it can return

  • VISIBLE AND TRUSTED

    You appear in half the runs or more, and are recommended in half of those or more.

  • VISIBLE NOT TRUSTED

    You appear often but are rarely recommended. Work on independent proof for your claims.

  • TRUSTED NOT VISIBLE

    You're recommended when named but rarely named. Work on the sources answers draw from.

  • NOT VISIBLE

    Rarely named and rarely recommended. Fix how clearly your brand is described first.

  • PANEL READY

    The panel is built and no runs exist yet.

  • NOT MEASURABLE

    Too few runs, prompts or engines for a rate. It reports the raw counts as counts.

  • BLOCKED

    There's no brand name.

Where it stops for you

  • When the panel is ready and the prompts need running
  • When there are too few runs to report a rate

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells attribution software. It ran 10 prompts 5 times each in ChatGPT and Perplexity in one week.

Verdict

VISIBLE NOT TRUSTED

Coverage: partial (Gemini and AI Overviews not run)

Rates: script

Headline

named in 59% of answers, recommended in 20% of those. Both are the mean of the 2 engines.

EngineRunsNamed (95% interval)Recommended (95% interval)Diagnosis
ChatGPT5056% (42% to 69%)18% (8% to 36%)Visible, not trusted
Perplexity5062% (48% to 74%)23% (11% to 40%)Visible, not trusted
Critical gap

named in 0 of 10 runs on each engine for the 2 comparison prompts, where buyers name a competitor and ask for alternatives.

Next

the answers name this company often but rarely pick it. That points to proof first: independent sources that back its claims, and the comparison pages answers cite.

The handoff block the next skill reads
skill: ai-answer-panel
version: 1.0.0
verdict: VISIBLE NOT TRUSTED
coverage: partial (Gemini and AI Overviews not run)
for: panel v1, October 2026
outputs:
  appearance_rate: 0.59
  recommendation_rate: 0.20
  critical_gaps: [comparison on chatgpt, comparison on perplexity]
next: [claim-check, cited-pages]
open: [add Gemini and AI Overviews runs next month]

What good looks like

  • A rate per engine, each with its interval
  • No rank and no single blended score without the engines behind it
  • Gaps named by prompt family, with comparison prompts first
  • A change called real only when the intervals don't overlap
  • The same panel wording as last period

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Why not report a rank?

The same prompt rarely returns the same list twice, so your position in one answer says little. How often you appear across many runs is the figure that holds.

How many runs do I need?

At least 5 per prompt per engine, at least 10 prompts and at least 2 engines. Below that, it reports what you found as counts and asks for the missing runs.

Does it run the prompts for me?

It can if your agent can open each engine in a clean session. Most teams run them manually or export them from an AI visibility tool, then paste the results.

What does the script do?

It turns your runs into appearance and recommendation rates per engine, with intervals, flags gaps by prompt family and lists the most cited pages. Without Python, the skill makes the same count itself and says so.