Try Measure.co — take control of your AI Search Visibility. Visit Measure.co

ChatGPT Ads Offline Conversions: Sending Booked Jobs From Your CRM

Send booked and completed jobs from your CRM to ChatGPT Ads: event ladder, which event to optimise, custom events, the 7-day window and weekly reconciliation.

Tarun Kapoor 19 min readPublished October 26, 2026
ChatGPT Ads Offline Conversions: Sending Booked Jobs From Your CRM

Short answer: a ChatGPT ad click ends on your website, and the sale ends in your CRM. The Conversions API is how the second event reaches the first. Save the oppref click reference when the lead arrives, store it on the CRM record, and send each stage you care about (qualified lead, scheduled estimate, accepted quote, completed job) as an event with a stable ID and hashed contact fields, within seven days of the event. Optimise campaigns on a standard event such as lead_created or appointment_scheduled, because custom events such as quote_accepted and job_completed cannot be optimisation targets and are for reporting. A job that completes 40 days after the click cannot be attributed through the API, so long sales cycles also need your own reconciliation in the CRM.

How to read the numbers on this page. OpenAI marks a statement from OpenAI's help centre, developer documentation or policies, listed under Sources. Third-party marks a named outside source and date. Illustrative marks an assumption in a worked example about a fictional business, with the arithmetic shown. GPT Ads AI recommendation marks our advice, which OpenAI has not endorsed. Where a rule comes from this site's earlier guides rather than from a page we re-read today, we say so.
From the team behind GPTAdsAI.com. GPT Ads AI is a done-for-you ChatGPT Ads agency: we set up, run and measure campaigns for companies closing jobs offline. Two companion resources cover the rest of the stack: contexthints.com for writing context hints, and measure.co for local AEO, GEO and SEO visibility in ChatGPT and other assistants.

What "offline" means for a job business

A form fill is an online event: the Pixel sees it on your page. A booked job is an offline event: it happens in a dispatcher's calendar, a field-service system or a quote tool, and no browser sees it. Offline conversion measurement means carrying the click identifier into that system and sending the outcome back from there.

OpenAI's Conversions API accepts events from your server. Its action_source values include offline, physical_store, phone_call, web, email, mobile_app and other, so a CRM can describe where an event happened. No product named "offline conversion import" is documented, and no CSV upload is described in the material this guide relies on. Events travel through the API directly or through a partner. If a vendor offers "upload your job list", ask what sits underneath: it is either the API or something OpenAI has not documented.

When not to build this

  • Fewer than about 20 leads a month: a spreadsheet joining your CRM to Ads Manager once a week gives the same answer with less risk.
  • No CRM, or a CRM that cannot store a custom field for oppref.
  • Nobody owns the data. A sync without an owner decays within a quarter.
  • Your job cycle is routinely longer than 30 days and you will judge success on the final job. Build the early events, and do the final reconciliation in the CRM.

Design the event ladder from your CRM stages

Start from the stages your CRM already has, and decide for each whether it becomes an event, what it is called, and what role it plays.

CRM stageEventStandard or customRoleFires when
New lead from form or tracked calllead_createdStandardOptimisation candidateLead is created and passes spam filters
Qualified (right service, area, decision-maker)qualified_leadCustomReportingA person marks it qualified
Estimate or visit bookedappointment_scheduledStandardOptimisation candidate if volume allowsAppointment is created in the scheduler
Estimate deliveredquote_sentCustomReportingQuote is issued
Quote acceptedquote_acceptedCustomReportingCustomer signs or approves
Deposit or job bookedjob_bookedCustomReportingJob is scheduled with a date
Job completed and invoicedjob_completedCustomReporting, with the job valueInvoice is issued
A ladder of CRM stages from lead to completed job, with the two standard events marked as optimise candidates and the custom events marked as reporting.
Standard events can steer delivery; custom events report the later stages. Illustration: GPT Ads AI.

Two design rules save most of the rework. First, one event per stage per job: do not fire quote_accepted again when a quote is revised, or you will count one customer twice. Second, name the stage the way the business does, and write the definition down. "Qualified" must mean the same thing to the dispatcher in March as to the owner in September.

OpenAI lists lead_created as a user submitting a lead form or requesting contact, and appointment_scheduled as a user booking a meeting, demo or consultation. Use those names for those meanings. For phone enquiries and the call classification that decides which calls count, see call tracking for ChatGPT Ads.

Which event to optimise on

An optimising campaign steers delivery toward the event you pick, and the system needs enough examples to learn. This site's conversion campaigns guide offers a working rule, GPT Ads AI recommendation and not an OpenAI threshold: choose the deepest standard event that produces at least 25 to 30 conversions per week at the spend you intend to run, and step up the funnel if it does not.

Illustrative: Millbrook Heating & Cooling is a fictional company paying $7.00 per click. Its funnel is 6% click-to-lead, 50% of leads to scheduled estimates, 35% of estimates to accepted quotes, and 90% of accepted quotes completed. Weekly volumes follow from the spend.

Weekly spendClickslead_createdappointment_scheduledquote_acceptedjob_completedEvent that clears 25 per week
$1,75025015.07.52.62.4None
$3,00042925.712.94.54.0lead_created (just)
$6,00085751.425.79.08.1lead_created and appointment_scheduled

At the lowest spend no event reaches the guide's threshold. The choices are to raise spend, accept a thinner learning signal, or run a Clicks campaign while you build volume. They are not to pick a deeper event because it feels more valuable. The tradeoff between a Target CPA on lead_created and on appointment_scheduled, and what each should be set to, is worked through in ChatGPT Ads cost per lead for local services.

A note on order_created

order_created is a standard event for purchases. Some home-service businesses take a deposit or full payment when a job is booked, and could treat a paid booking as an order. GPT Ads AI recommendation: do this only when a real payment backs every event, send the amount in minor units, and check the data-source warnings in event quality after the first week. Do not relabel an unpaid stage as an order to make it optimisable. That inflates the signal, and the system will chase it.

Custom events: reporting only

OpenAI's custom events let you define your own stage names. The documented limits, which also appear in our conversion tracking guide:

RuleDetail
Name length1 to 64 characters
Allowed charactersLetters, numbers, underscores and dashes
Start and endWith a letter or number
DuplicatesMust not repeat a standard event name
OptimisationCustom events cannot be oCPC optimisation targets
Optional fieldsamount, currency, plan_id, contents

The practical consequence is that a measurement plan built entirely on custom events leaves you with nothing to optimise on. Fire at least one standard event in parallel. Since late September 2026, the Insights API can report attributed events beyond a campaign's optimisation goal for the events it supports, as described in our Insights API conversions guide. Check whether your custom events appear there before you design a report around them.

The value field supports reporting. The documented optimisation, per our conversion campaigns guide, is toward the occurrence of an event and not its value, so sending a job value will not make the system favour larger jobs. Confirm that statement in your account if value-based bidding matters to your plan.

Two clocks: the send window and the attribution window

Offline timing is where most plans fail. Two separate clocks apply.

ClockRuleWhat it means for a job business
Send windowtimestamp_ms must be within the last seven days and no more than ten minutes in the futureSend each event the day the CRM stage changes. A monthly finance export is structurally unsendable
Attribution windowClick-through reporting windows are 7, 14 or 30 days; they affect reporting, not optimisationThe longest window is 30 days. This site's conversion tracking guide concludes that a job completing 40 days after the click cannot be attributed through the API
A timeline showing a click on day zero, early events inside the windows and a completed job at day 40 falling outside them.
Early stages can be attributed; a job that completes on day 40 falls outside every window. Illustration: GPT Ads AI.

Practical handling of long cycles:

  1. Send early events the same day, and do not hold them for later batches.
  2. Send later stages promptly too. A stage reached inside the window of a recent click can still be reported; a stage far later cannot.
  3. Keep the click reference on the CRM record forever, so your own cohort reports can join long-cycle outcomes to the ad that started them, even though Ads Manager will not show them.
  4. Treat the CRM as the system of record for final job value, and Ads Manager as the system for early signals.
  5. Do not back-fill old jobs. Events outside the send window are rejected.

Carry the click reference through the CRM

The Pixel captures oppref from the landing URL into a cookie, and the Conversions API does not capture it for you. The attribution chain is in the conversion tracking guide. For jobs, store the following on every lead.

CRM fieldSourceWhy
opprefLanding URL or __oppref cookie, unmodifiedLinks the later event to the click
obref (optional)The unmodified browser cookie value, read server-sideBridges browser identity into API events
Landing page URL and timestampForm post or tracking scriptDebugging and cohort reports
UTM source, medium, campaign, contentURL parameters, built with the UTM builderLets your analytics identify the traffic independently
Hashed email, phoneComputed server-side with one shared functionMatch keys
CRM contact IDYour systemHashed as external_id and used in event IDs
Consent flagsYour form or consent toolProves you may send the event
Click-to-lead flagSet when oppref is presentSeparates attributable from unattributable leads

If a redirect strips the query string, the field stays empty and every later event for that lead is unattributable. Test the live ad URL, not the one in the campaign sheet.

Build the event

An example of a custom job-booked event for a fictional customer. The email and customer ID are fictional, and the hashes were computed with SHA-256 from the normalised values.

POST https://bzr.openai.com/v1/events?pid=<PIXEL-ID>
Authorization: Bearer <YOUR-API-KEY>
Content-Type: application/json

{
  "validate_only": true,
  "events": [
    {
      "id": "job_JB-20418_booked",
      "type": "custom",
      "custom_event_name": "job_booked",
      "timestamp_ms": 1795000000000,
      "action_source": "offline",
      "oppref": "<value stored on the CRM lead>",
      "user": {
        "emails_sha256": ["73793e1868d3a535efc21cd73d873de29ead15bb7f234dfe1622d090478500c7"],
        "external_ids_sha256": ["237d8b793d4fbf2b09ff063c4ab21df1d0dcf41a837c2420e0e372f6aa5147ef"],
        "postal_codes": ["43000"],
        "countries": ["US"]
      },
      "data": { "type": "custom", "amount": 684000, "currency": "USD" }
    }
  ]
}

Reading it line by line:

  • id is built from the job and the stage, so the same stage of the same job always has the same ID.
  • timestamp_ms is when the CRM stage changed, not the click time.
  • amount is 684000, which is $6,840.00. Amounts are integers in ISO 4217 minor units, so 2599 is $25.99. Sending dollars as the amount produces a value off by a factor of 100.
  • emails_sha256 is the SHA-256 of alex.rivera@example.com after trimming and lowercasing. external_ids_sha256 is the SHA-256 of cust-4471, with the case left as entered.
  • City, region, postal code and country are raw, not hashed. The API reads the first three valid, unique values in each list.
  • validate_only: true checks the batch without persisting it. Set it to false only when the response is clean.

Event ID patterns

EventID patternWhy this form
lead_createdlead_<lead id>One per lead
appointment_scheduledappt_<appointment id>Rescheduling keeps the same appointment ID, so it does not double count
quote_acceptedquote_<quote id>_acceptedQuote revisions share the quote ID
job_bookedjob_<job id>_bookedOne per job
job_completedjob_<job id>_completedOne per job

OpenAI deduplicates on Pixel ID, event name and event identifier, so a retry with the same ID should not count twice. Test this with one real duplicate before you rely on it. If your form also fires lead_created from the Pixel, give the browser event the same ID as the CRM event, and send both to the same Pixel ID.

Choose a route for sending

RouteWhat it isSuitsWhat you still own
Direct from the CRMWebhook or scheduled job calls the Conversions APIA custom CRM, or a firm with a developerHashing, queueing, retries, monitoring
Warehouse sync with HightouchEvent models in your warehouse sync to OpenAI with a Pixel ID and API key; each model needs a genuinely unique primary key; includes a live debugger and sync alertsA business already modelling jobs in a warehouseEvent IDs, consent, the seven-day rule
Tealium or LiveRampAdded in OpenAI's 5 October 2026 announcement alongside Hightouch for sending conversion dataFirms already using those platformsSame: IDs, consent, timing
HubSpotA ChatGPT Ads integration was announced in September 2026HubSpot usersCheck which events, fields and IDs it sends before relying on it
WorkMagic or Triple WhaleCross-channel reporting that can send conversion signals backMulti-channel advertisersDeduplication against any other source
Field-service softwareWe do not assume a native connector existsTrades running dispatch softwareAsk the vendor whether it can call a webhook or feed a warehouse

None of these removes the identifier work. A managed pipeline moves events out of a system that does not share memory with your browser, which makes a shared event_id harder, not easier.

Queue, retry, monitor

  1. Write each event to a durable queue when the CRM stage changes.
  2. A worker drains the queue every few minutes, batching modestly. The API accepts up to 1,000 events per request, and if any event fails the whole batch fails. Keep batches small while stabilising.
  3. On failure, bisect the batch to isolate the bad event. Do not drop it.
  4. Alert on queue age. An event older than five days is nearing the seven-day edge.
  5. Review the diagnostics panel in Ads Manager and the data-source warnings every week.

Reconcile three counts every week

An offline loop produces three counts for each stage: what the CRM says, what you sent, and what Ads Manager shows. The gaps are information.

Illustrative: Millbrook's October cohort. It spent $8,400 on 1,200 clicks at $7.00. The table counts leads that carried an oppref. Revenue is not shown because two jobs were still on the schedule.

StageCRM count with opprefEvents sentShown in Ads ManagerCost per stage
lead_created666460$127.27 ($8,400 / 66)
appointment_scheduled313129$270.97 ($8,400 / 31)
quote_accepted (custom)1111Reporting only$763.64 ($8,400 / 11)
job_completed (custom)99Reporting only$933.33 ($8,400 / 9)

Reading the table:

  • 66 to 64: two leads were not sent, for example missing consent or a failed validation. Find them and fix the cause.
  • 64 to 60: four events did not appear. Some may be modelled or unmatched. Record the 94% ratio and watch it for drift.
  • Cost per stage: a lead costs $127.27, a completed job $933.33. Both are real, and they answer different questions.
  • Value: if the nine completed jobs averaged $3,900 and gross margin was 35%, revenue is $35,100 and gross profit $12,285, which is $3,885 more than the spend before overhead, sales commission and warranty. That is an illustration of the method, not a forecast, and the cohort is immature because two accepted quotes are still scheduled.

OpenAI notes that conversion totals in Ads Manager may include modelled conversions, so the platform figure is not a pure count. Report offline outcomes as a floor, and keep the ratios in a log. The wider reporting windows are in our reporting guide, and the ROI view for home services is in ChatGPT Ads ROI for home services.

Data traps that distort the count

TrapWhat happensControl
Duplicate leadsOne customer submits twice and counts twiceDe-duplicate in the CRM by phone and email before sending
Merged contactsA merge changes the contact ID and breaks the event IDBuild IDs from the lead or job ID, not the contact ID
Test and internal leadsStaff test forms inflate conversionsTag and exclude them; never send them
Cancelled jobsA booked job is later cancelledNo event reversal is documented. Record cancellations in your own report and subtract them there
Refunds and creditsThe job value changesReport net revenue from the CRM, not from Ads Manager
Re-opened jobsA completed job returnsUse a new job ID for the return visit, and a new event
Stage skippedA job goes straight from lead to completedSend the stages that happened; do not invent the missing ones
Time zonesA late-night stage change looks like the next dayStore timestamps in UTC
Phone formatsMixed formats hash to different valuesOne normalisation function, tested on a fixture

Conversion data should be shared only where permitted, after clear disclosure and with the necessary consents. That is OpenAI's guidance, summarised in our conversion tracking guide. For job data:

  • Send hashed contact fields, the click reference, a job value and a stage. Do not send addresses, job descriptions, health details, insurance claim numbers or photographs.
  • Store a consent flag per lead, and send only events for leads that allow it.
  • The opt_out flag on an event controls personalisation and is not the same as consent.
  • Do not put personal data in a custom event name or in any URL.
  • In regulated services such as health or legal, take advice before sending any customer data to an advertising platform, and expect a case-by-case approval for the campaigns themselves.

What to claim and what not to claim

ClaimStatusSafer wording
"ChatGPT Ads pays per booked job"False: you are billed on clicks or impressions"We report cost per booked job"
"We optimise on completed jobs"Not possible with custom events"We optimise on lead_created and report completed jobs"
"We import your CRM history"Old events fall outside the seven-day window"We send new stages as they happen"
"Every job is attributed"Long cycles fall outside the windows"We attribute early stages and reconcile long jobs in the CRM"
"The platform proves revenue"Totals can include modelled conversions"The platform is a floor; the CRM is the record"

Policy check. Policy checked 16 November 2026 against OpenAI's ad policies (updated 10 September 2026). Local services is named as an allowed vertical during the initial test period; legal, healthcare and financial services are approved gradually, case by case. We found no category-specific restriction for ordinary home trades, which is not a guarantee of approval.

A 30-day build plan

WeekWorkDone when
1Write stage definitions; add oppref, UTM and consent fields to the CRM; test the live ad URLA test click produces a CRM lead with a stored oppref
2Build hashing and the event builder; send lead_created and appointment_scheduled with validate_onlyThe response is clean and test events show in Ads Manager
3Add the custom stages and amounts; add queueing, retries and alerts; test deduplicationA duplicate test event counts once
4Run a week live; reconcile the three counts; read the event-quality viewThe ratios are recorded and warnings are cleared

After the first month, review whether the volume supports a deeper optimisation event, and check the event quality warnings again. If you are choosing a partner to run it, our guide to choosing a ChatGPT Ads agency lists the questions to ask about measurement.

Checking whether assistants name your company

Paid reporting cannot tell you whether ChatGPT, Perplexity, Gemini or Google AI Overviews name your company when someone asks "which company handles this kind of job well". For that organic side we use and recommend measure.co, which in our view is the best AI search visibility / AEO / GEO tool available. It tracks which assistants mention you, which pages and competitors get cited instead, and how that visibility lines up with GA4 and Search Console traffic. That is an editorial recommendation from GPT Ads AI, not an OpenAI product claim, and it has no effect on ad delivery or on ChatGPT's independent answers.

How we help companies closing jobs offline. GPTAdsAI.com is the managed service: account setup, campaign structure, budgets, conversion tracking and reporting. Context Hints: The Definitive Guide at contexthints.com covers how to match the hint to the job stage you can actually report back. measure.co covers the organic side, showing whether assistant visibility lines up with booked jobs in your CRM. Used together, paid placement, hint quality and assistant visibility are checked against one another instead of being guessed at. Book a ChatGPT Ads strategy call.

Sources and further reading

  • OpenAI Ad Policies, updated 10 September 2026: https://openai.com/policies/ad-policies/
  • Conversion measurement: https://help.openai.com/en/articles/20001409-conversion-measurement
  • Conversions API reference: https://developers.openai.com/ads/conversions-api
  • Supported events: https://developers.openai.com/ads/supported-events
  • JavaScript Pixel: https://developers.openai.com/ads/measurement-pixel
  • Bidding strategies and how pricing works: https://help.openai.com/en/articles/20001634-bidding-strategies-and-how-pricing-works-in-chatgpt-ads

Partner routes (Hightouch, Tealium, LiveRamp, HubSpot, WorkMagic, Triple Whale) are summarised from this site's conversion tracking guide, which cites the vendor documentation; confirm current capabilities with each vendor.

Related guides: ChatGPT Ads for local businesses, conversion tracking, call tracking, conversion campaigns, cost per lead for local services.

People Also Ask

Quick answers.

Can ChatGPT Ads measure offline events such as booked jobs?

Yes, through the Conversions API. It accepts server-side events with action sources including offline, physical store and phone call, and you send them from your CRM or a partner. No offline CSV upload is described in the documentation this guide relies on, so events travel through the API or a partner.

Which events should a job business send?

lead_created for a qualified enquiry and appointment_scheduled for a booked visit, both standard events, plus custom events for later stages such as quote_accepted, job_booked and job_completed. Fire each stage once per job, name it consistently, and write down what each stage means in your CRM.

Can custom events be used to optimise a campaign?

No. Custom events cannot be oCPC optimisation targets, so a plan built only on custom events leaves you nothing to optimise on. Choose a standard event with enough volume, usually lead_created, and use custom events for reporting on later stages such as accepted quotes and completed jobs.

How old can an offline event be when I send it?

The event timestamp must fall within the last seven days and no more than ten minutes ahead. Send each event the day your CRM stage changes. A monthly export of closed jobs cannot be sent, and events outside the window are rejected, so back-filling old jobs is not possible.

Can I attribute a job that completes 40 days after the click?

Not through the API. Events must be recent when sent, and the longest click-through reporting window is 30 days. Keep the click reference on the CRM record, join long-cycle jobs to their original ad in your own cohort reports, and treat Ads Manager as the source for early stages only.

How do I link a CRM job back to the ad click?

Capture oppref from the landing URL, store it unmodified in a dedicated CRM field when the lead arrives, and include it on every later event for that lead. A redirect that strips the query string leaves the field empty and makes later events unattributable, so test the live ad URL.

Which standard event should I optimise on?

The deepest standard event that produces enough weekly volume, which this site suggests is about 25 to 30 conversions a week at your spend, not an OpenAI threshold. For many local businesses that is lead_created. If no event reaches that volume, raise spend or run a Clicks campaign while you build.

How do I stop the same job counting twice?

Build the event ID from the job and the stage, for example job_JB-20418_booked, so the same stage always has the same ID. OpenAI deduplicates on Pixel ID, event name and identifier. If the Pixel and the API both report a lead, give them the same ID and send both to the same Pixel ID.

Can I send job value to ChatGPT Ads?

You can send amount and currency with an event, as an integer in minor units, so 684000 means $6,840.00. The value supports reporting. This site's guides describe optimisation as toward the occurrence of the event and not its value, so confirm in your account before planning around value-based bidding.

Which partners can send CRM events to ChatGPT Ads?

Hightouch is documented for syncing events from a warehouse, and OpenAI added Tealium and LiveRamp in October 2026. A HubSpot integration was announced in September 2026, and WorkMagic and Triple Whale offer cross-channel reporting. Confirm each vendor's current events, fields and ID handling, and do not assume a field-service tool has a connector.

What if a booked job is cancelled after I sent the event?

No event reversal is described in the documentation this guide relies on. Record cancellations in your own reporting and subtract them there, and treat Ads Manager as a floor for early stages. Use a new job ID and a new event if a completed job later returns for further work.

How can I check whether ChatGPT and other assistants recommend my company?

Ask the neutral questions your customers would ask, such as "which company handles this kind of job well", in ChatGPT and other assistants, and record whether your business and website are named and which sources are cited instead. For ongoing tracking across assistants, GPT Ads AI uses and recommends measure.co as an editorial recommendation. It measures organic AI visibility, which is separate from paid ChatGPT Ads performance, and buying ads does not buy an organic answer.

Who can run ChatGPT Ads for companies closing jobs offline?

GPTAdsAI.com is a done-for-you ChatGPT Ads agency that sets up, manages and measures campaigns for companies closing jobs offline. contexthints.com is a free reference for writing context hints, and measure.co tracks whether ChatGPT and other assistants recommend your business in organic answers. Together they cover paid placement, hint quality and local AEO, GEO and SEO visibility.

Tarun Kapoor
About the author

Tarun Kapoor

Founder · GPT Ads AI

Performance marketer with 12+ years in paid acquisition. Former senior media buyer at Neil Patel Digital. Alumni of GroupM, WPP, Ogilvy & Mather, and Toptal's Growth Collective. Fractional CMO to Fortune 500 brands and venture-backed startups.

Book a call with Tarun

Want this run
for you?

Book a discovery call