Skip to content

Decides

ChatGPT Ads check

Decides whether ChatGPT Ads fit your B2B audience yet, writes ads that follow on from the buyer's question, and checks reported clicks against real visits.

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 chatgpt-ads-check.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
ChatGPT Ads documentation
Needs
Where you sell, and the most a lead may cost · Which of your markets Ads Manager shows as open · Your tracking plan for the test
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedPaid media
SkillPaid media

ChatGPT Ads check

Example output

Check ChatGPT Ads before a test
CheckResultCall
Market openUS onlyPass
Tracking before spendNext weekPass, open item
Who sees the adsNot confirmedRisk

Decides whether ChatGPT Ads fit your B2B audience yet, writes ads that follow on from the buyer's question, and checks reported clicks against real visits.

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

chatgpt-ads-check/SKILL.md

In the folder

  • chatgpt-ads-check/SKILL.md
  • chatgpt-ads-check/references/chatgpt-ads-facts.md
  • chatgpt-ads-check/scripts/reconcile_clicks.py
---name: chatgpt-ads-checkdescription: >-  Decides whether ChatGPT Ads fit a B2B SaaS company yet (market open, buyers there, core channels measured, tracking  ready before spend, a capped test with a stop rule), writes chat-card ads and context hints matched to the buyer  questions the company isn't named in, and after a test reconciles the clicks Ads Manager reports with the visits  and leads that actually arrive. Every platform fact carries the date it was checked. Returns GO TEST, NOT YET,  NO FIT or BLOCKED before a test, and RECONCILED, INVESTIGATE or BROKEN after one. Use it before spending on  ChatGPT Ads and after each test. Not for measuring organic AI visibility (use ai-answer-panel).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Uses a connected ad account or analytics  if the agent has one, otherwise exports or pasted figures. Python 3 is optional, for the reconciliation.metadata:  author: SHAKEUP  version: "1.0.0"  stage: decide  plugin: paid-media  homepage: https://www.shakeuphq.com/skills/chatgpt-ads-check--- # ChatGPT Ads check Says whether ChatGPT Ads are worth a test yet, writes ads that follow on from the buyer's question, and checks that the clicks you pay for become real visits. ## Purpose - One job: a go or stop call on ChatGPT Ads for one company, with the ads for a test, and the reconciliation after it.- Stage: decide. It ends in a verdict for the mode it ran.- It marks every platform fact with the date it was checked (`references/chatgpt-ads-facts.md`). The platform is new and changes often.- It never launches, funds or edits a campaign, and never invents a reach, a CPC or a conversion rate. ## Use when / not | Situation | Use || --- | --- || Asking whether to test ChatGPT Ads | this skill, mode `fit` || A test is approved and needs its ads | this skill, mode `fit` (the ads come with `GO TEST`) || A test has run and the clicks don't match the visits | this skill, mode `reconcile` || Measuring how often AI answers name you for free | `ai-answer-panel` || Setting the cost per lead ceiling | `lead-value-math` || Checking conversion tracking across every platform | `tracking-audit` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || Selling regions | yes (fit) | `## Company` "Regions"; the reader | ask once || Max CPL | yes (fit) | `lead-value-math` handoff; the reader | `BLOCKED` until set; offer `lead-value-math` || Market open in the account | yes (fit) | the reader's Ads Manager location lookup; the reader | `NOT YET` until confirmed || Evidence buyers use ChatGPT to research the category | no | analytics sessions from chatgpt.com (last 90 days); "ChatGPT" in free-text "how did you hear about us" answers; `ai-answer-panel` results | the fit call is marked `partial`; never assume || Core channel measurement | yes (fit) | `## Targets` filled and a channel with CRM outcomes reconciled within 15%; the reader | `NOT YET` || Tracking plan: Pixel or Conversions API, `oppref` kept, UTMs on the landing URL | yes (fit) | the reader; `tracking-audit` | `NOT YET` until it is live before spend || Buyer questions where the company isn't named | no | `ai-answer-panel` handoff; `## Buying moments` | use the buying moments' questions, coverage `partial` || Test budget and length | yes (fit) | the reader | ask once || Reported clicks and tagged analytics sessions, by day | yes (reconcile) | Ads Manager or the Insights report; analytics export; pasted rows | `BLOCKED` for reconcile || Leads, spend, platform conversions | no (reconcile) | CRM; Ads Manager | visits reconciled, outcomes not | ## Profile read From `company-profile.md`: `## Company` (regions, category, business model), `## Buyers` and `### Buying group`, `## Buying moments`, `## Messaging house`, `## Voice`, `## Claims and proof`, `## Channels` (status of ChatGPT Ads and the others), `## Targets`, `## Accounts and data` (analytics, CRM, connected ad accounts), and every `## Learnings` entry for `chatgpt-ads-check` with status `noted`, `fact` or `accepted`. No profile: ask in one message for the selling regions, the most a lead may cost, the test budget, and which analytics and CRM hold the data. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings.** A past test's result, or a tracking fix, changes the prerequisites below. Name each one that fires. 2. **Pick the mode.** Reported clicks and sessions given: `reconcile` (step 9). Otherwise: `fit` (steps 3 to 8). 3. **Refresh the facts.** Read `references/chatgpt-ads-facts.md`. If its check date is more than 3 months old, open the OpenAI docs again and update what you repeat. Treat the "not confirmed" list as risks, never as facts. 4. **Run the prerequisites** (mode `fit`):    | Check | Passes when | Why |   | --- | --- | --- |   | Market | at least one selling region shows as available in the reader's account | a valid country code doesn't mean the country is open |   | Buyers there | analytics shows chatgpt.com sessions on buying pages, or buyers name ChatGPT in "how did you hear", or tracked buyer prompts exist for the category | ads under answers only matter where buyers ask |   | Core channels measured | `## Targets` is filled and one channel's platform and CRM counts agree within 15% | a new channel can't be read if the old ones can't |   | Tracking before spend | Pixel or Conversions API sends the lead event, `oppref` survives the landing page and reaches server events, `event_id` dedupes browser and server, UTMs on the final URL; live now, or scheduled before launch with a test lead confirmed before the first spend (an open item that gates the launch) | advertisers get aggregated reports, so tracking is the only way to count leads |   | Landing page | reachable by OpenAI's ad crawlers, no redirect to another address, picks up the buyer's question in its headline | a redirect can drop `oppref`, and the ad has to continue the conversation |   | Budget and stop rule | test budget at least 2 x max CPL, a fixed length, and a written rule agreed before launch | under 2 leads' worth of spend, a zero is noise |    The stop rule to write: after the test budget, continue only if cost per lead is at or under max CPL and the reconciliation is `RECONCILED`; stop if 0 leads at 2 x max CPL; repeat with a change if leads came in above the ceiling. 5. **Note the reach risk.** Which plans see ads is not confirmed in OpenAI's docs (see the reference). If the reader's buyers work at companies that give staff company plans, say the test may reach fewer of them, and read results by job role on the landing page form, not by clicks. 6. **Pick the questions.** From `ai-answer-panel` (prompts where the company isn't named and the buyer is close to choosing) or the buying moments' questions. One ad group per cluster of questions that one card and one page can answer honestly. 7. **Write the ads** (only on `GO TEST`): per ad group, 2 or 3 chat cards (title 3 to 50 characters, body up to 100) that follow on from the question without repeating the answer, plus context hints that describe the use cases and needs in plain words, not keywords. Proof only from `## Claims and proof`. Run `voice-check` and `claim-check` on every card (inline without them: no dashes, no profile words to avoid, every number in the claims). Fix and recheck at most twice. 8. **Write the test plan:** markets, platform groups, ad groups and their questions, budget, length, bid strategy (a click objective with a fixed bid or Maximize clicks while conversion volume is low; Maximize conversions once the lead event fires reliably), tracking checklist, stop rule, and when to reconcile (after the first 7 days and at the end). 9. **Reconcile** (mode `reconcile`). Run `python3 scripts/reconcile_clicks.py --csv data.csv`. Without Python: gap = 1 minus tagged sessions / reported clicks, by campaign and by day; list days with a gap over 50%; say you ran it by hand.    | Gap | Verdict | Why |   | --- | --- | --- |   | at or under 15% | `RECONCILED` | the same tolerance platform and CRM counts are held to |   | over 15% to 50%, or sessions more than 15% above clicks | `INVESTIGATE` | something loses or mixes visits; work the cause table in the reference, cheapest check first |   | over 50%, or 0 sessions on 20 or more clicks | `BROKEN` | most paid clicks never reach analytics: stop reading performance until it's fixed |    With leads: cost per lead against max CPL. With platform conversions: compare click-through conversions only, on the same date basis, before calling a gap. ## Check and verdict Mode `fit`, first match wins: | Verdict | Condition || --- | --- || `BLOCKED` | No selling region, or no max CPL after one question || `NO FIT` | None of the selling regions is open to ChatGPT Ads in the reader's account || `NOT YET` | Any other prerequisite in step 4 fails: name each one and what fixes it || `GO TEST` | Every prerequisite passes: the ads and the test plan follow | Mode `reconcile`, first match wins: `BLOCKED` (no clicks or no sessions), `BROKEN`, `INVESTIGATE`, `RECONCILED`, by the gap table in step 9. Named stops: `BLOCKED` asks for the missing input. `NOT YET` lists the failing prerequisites in the order to fix them. `BROKEN` stops any reading of the test's performance. No other stops. ## Learn Append to `## Learnings` when: - a test ends: a `fact` with spend, leads, cost per lead, the reconciliation gap and the stop rule's call;- a reconciliation finds the cause of a gap (a redirect, a consent setting): a `fact`, so the next test checks it first;- a platform fact in the reference turns out to have changed: a `fact` with the date and the page, and the reference updated;- the reader goes ahead despite `NOT YET`: a `proposal` with their reason. Example: `- 2026-10-24 | chatgpt-ads-check | fact | The demo page redirect dropped oppref; 31% of clicks had no session until it was removed | Why: reconciliation, 2026-10-01 to 10-14 | Status: noted` ## Output ```Mode: <fit | reconcile> · Verdict: <GO TEST | NOT YET | NO FIT | BLOCKED | RECONCILED | INVESTIGATE | BROKEN> · Coverage: <complete | partial (what's missing)>Facts as of: <check date of the reference, and what was rechecked> fit:| Prerequisite | Result | Evidence | Fix |Reach risk: <one line>Questions and ad groups: <list>Ads: | Ad group | Title (chars) | Body (chars) | Proof | Context hints |Test plan: <markets, platforms, budget, length, bidding, tracking, stop rule, reconcile dates> reconcile:| Campaign | Clicks | Sessions | Gap | Status |Days over 50%: <dates> · Cost per lead: <value or not available> · Likely causes, in order: <list> Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: chatgpt-ads-checkversion: 1.0.0verdict: GO TESTcoverage: partial (reach by plan not confirmed)for: <company> ChatGPT Ads testoutputs:  mode: fit  markets: [US]  test_budget: 3000  stop_rule: continue if CPL at or under 285 and reconciled; stop at 0 leads after 570  ad_groups: 2  facts_checked: 2026-10-24next: [tracking-audit]open: [confirm the pixel fires on the demo form before spend]``` After a test, `mode: reconcile` with `gap` and `cost_per_lead` in `outputs`. ## Failure modes | Problem | Do this || --- | --- || The reader quotes a platform fact newer than the reference | Use it if they can show the source, then update the reference date || Analytics can't separate paid from organic ChatGPT visits | Add the UTM to the final URL before the test; until then, reconciliation is `INVESTIGATE` || The test is too small to reach 2 x max CPL | `NOT YET`; offer a longer test at the same daily budget instead of a bigger one || Ads Manager isn't open to the reader's account yet | `NOT YET` or `NO FIT` by region; record the date checked || The reader asks for ads before the fit check | Run the fit check first; ads come only with `GO TEST` | ## Influenced by - OpenAI's ChatGPT Ads documentation for advertisers and developers: bidding, targeting, context hints, chat card limits, the Measurement Pixel, the Conversions API and attribution windows: https://developers.openai.com/ads- SHAKEUP's ChatGPT Ads practice: a capped test with one question and a stop rule agreed in advance, tracking before the first spend, and ads planned on the buyer prompts where the company isn't named yet.

How it thinks

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

What we built

ChatGPT Ads are new and change month to month, so every platform fact here is dated and every test is capped with a stop rule written first. Tracking comes before the first spend, because the reports are aggregated and leads only count if you track them.

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

    Read your learnings first

    A past test's result or a tracking fix changes the checks before it starts.

  2. 02

    Refresh the platform facts

    Every fact it uses carries the date it was checked against OpenAI's docs. Older than 3 months, it checks again.

  3. 03

    Run the checks

    An open market, buyers who use ChatGPT, channels already measured, tracking live before spend and a test of at least 2 leads' worth.

  4. 04

    Write the stop rule first

    Before any spend, it writes what result would justify more, and what ends the test.

  5. 05

    Write the ads

    Only after a go. Each chat card follows on from the buyer's question within the platform's limits, with plain context hints.

  6. 06

    Reconcile clicks and visits

    After a test, it compares reported clicks with tagged visits by day. Over 15% apart, it works through the likely causes.

What you need

  • Required

    Where you sell, and the most a lead may cost

    Your company profile → The handoff from lead value math → Your answer

  • Required

    Which of your markets Ads Manager shows as open

    The location lookup in your own account → Your answer

  • Required

    Your tracking plan for the test

    Your answer → The handoff from a tracking audit

  • Optional

    Buyer questions where you aren't named

    The handoff from an AI answer panel → The buying moments in your profile

  • Optional

    Reported clicks and tagged visits, by day, after a test

    Ads Manager and your analytics, connected or exported → Rows you paste

What it can return

  • GO TEST

    Every check passes, and the ads and test plan follow.

  • NOT YET

    A check fails, such as tracking, measurement, budget or market. Each one comes with its fix.

  • NO FIT

    None of your markets is open to ChatGPT Ads in your account.

  • BLOCKED

    Your markets or your cost per lead ceiling are missing.

  • RECONCILED

    Visits are within 15% of reported clicks.

  • INVESTIGATE

    Visits and clicks are 15% to 50% apart, or paid and organic visits are mixed.

  • BROKEN

    More than half the clicks never arrive as visits, so no result can be read.

Where it stops for you

  • When your markets or your ceiling are missing
  • When any check fails before a test
  • When the tracking is broken after one

Example output

What you get back.

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

Here's a fit check for a made-up B2B SaaS company that sells attribution software in the US, UK and Netherlands. Platform facts were checked against OpenAI's docs on 24 October 2026.

Verdict

GO TEST

(US only)

Coverage: partial (which plans see ads isn't confirmed)

CheckResultEvidence
MarketPass, US onlyThe location lookup shows the US as open, with the UK and Netherlands not confirmed yet
Buyers therePass410 visits from ChatGPT in 90 days, and 9 demos that named it
Channels measuredPassGoogle Ads and the CRM agree within 5%
Tracking before spendPass, open itemPixel and Conversions API go live next week, before any spend
Budget and stop rulePass$3,000 is above $570, which is 2 leads at the $285 ceiling
TitleBody
Attribution that keeps your stagesReads your HubSpot lifecycle stages and never rewrites them. See it on your data.
Pipeline by channel for your boardSetup takes 14 days on average. Book a walkthrough.
Stop rule

continue only if leads cost $285 or less and visits reconcile within 15%. With 0 leads after $570, the test ends.

The handoff block the next skill reads
skill: chatgpt-ads-check
version: 1.0.0
verdict: GO TEST
coverage: partial (reach by plan not confirmed)
for: ChatGPT Ads test, US
outputs:
  mode: fit
  markets: [US]
  test_budget: 3000
  stop_rule: continue if CPL at or under 285 and reconciled; stop at 0 leads after 570
  ad_groups: 2
  facts_checked: 2026-10-24
next: [tracking-audit]
open: [confirm the pixel fires on the demo form before spend]

What good looks like

  • Every platform fact carries the date it was checked
  • A stop rule written before any spend
  • Tracking live before the first click
  • Ads that follow on from the question within the card limits
  • A reconciliation by day that points to the day something broke

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.

Are ChatGPT Ads worth testing for B2B?

Only once your other channels are measured and tracking is live, and only where your buyers already research in ChatGPT. The skill checks each condition and says which one fails.

Who sees the ads?

OpenAI's docs don't say which plans see ads, so the skill treats it as a risk. If your buyers use company plans, the test may reach fewer of them.

Why reconcile clicks with visits?

Because the clicks you pay for only matter if they arrive. A redirect, a consent banner or a slow page can lose visits without anything looking wrong in Ads Manager.

How big should a test be?

At least what 2 leads may cost at your ceiling, with a fixed length and a stop rule agreed first. Under that, a test with no leads tells you nothing.