Obenan

Signal · Advertising in ChatGPT

The ad got the credit. Did it bring you a new customer?

On 5 October 2026 OpenAI gave ChatGPT advertisers more ways to link their sales to their ads. The ad report gets more detailed. It still can’t tell you which customers would have come anyway, and OpenAI says its own work on that question is at an early stage.

Published

Illustrative example, not a real shop or result.

One shop, three customers

The same month of ads, seen three ways

An illustrative example written for this page, not a real shop or a measured result.

A bike shop runs a small ChatGPT ad for a month. Three customers buy a bike. Here is what the ad report shows, and what actually happened.

  • The new customer

    1. Saw the ad
    2. Clicked
    3. Bought

    Had never heard of the shop. Saw the ad, clicked, and bought a bike.

    In the ad report
    Yes
    Extra sale from the ad
    Yes

  • The regular

    1. Regular
    2. Clicked
    3. Bought

    Already knew the shop and was coming that week anyway. Clicked the ad on the way.

    In the ad report
    Yes
    Extra sale from the ad
    No

  • The walk-in

    1. Saw the ad
    2. No click
    3. Came in later

    Saw the ad without clicking, then walked into the shop the following week.

    In the ad report
    No
    Extra sale from the ad
    Possibly

The report credited two sales, and one of them would have happened anyway. The sale it missed may well have been the ad’s doing.

Three different numbers

One sale, three questions

Ad reports often call all of these “conversions”. They answer different questions, so keep them apart.

  1. The sale

    Did a sale happen?

    Who can answer
    Your till, booking system or online checkout.

    This is the plain fact. Every other number should match back to it.

  2. The credit

    Which ad gets the credit?

    Who can answer
    The ad report, following rules you choose.

    A rule matches the sale to an ad. It can give too much credit, like the regular who was coming anyway, or too little, like the walk-in who never clicked.

  3. The extra customer

    Would the sale have happened without the ad?

    Who can answer
    A fair comparison: similar people or areas, some shown the ad and some not.

    The difference between the two groups is what the ad added. No ad report can show this on its own.

A sale is not the same as credit, and credit is not the same as an extra customer.

What changed

What OpenAI announced on 5 October

Most of the news improves the second number, the credit. The work on the third number is still early.

Available now

  • Easier ways to send sales data to ChatGPT Ads from tools businesses already use: Hightouch, Tealium and LiveRamp.
  • Support for well-known measurement companies on websites and apps.
  • You choose how long after a click a sale can still be credited, and can also count sales within one day of someone seeing the ad without clicking. This changes the report, not how the ads are shown.
  • Negative Phrases, which let qualifying advertisers keep their ads away from topics their own policies rule out.

Early work or pilots

  • Tests with Haus, Measured and WorkMagic that compare similar geographic areas with and without ads, to measure what the ads actually caused.
  • Brand-safety pilots with DoubleVerify and Integral Ad Science, run in a controlled test setting without access to private conversations.
  • Early testing with Kantar and Cint on how ads affect brand awareness.

Not launched yet

  • A visual ad format shown while ChatGPT creates an image. Testing begins later in October 2026 in the US with an initial group of advertisers.

OpenAI doesn’t say in which countries each tool is offered. It also says ads don’t influence ChatGPT’s answers, so paying for an ad and being recommended in an answer stay two separate things.

A published result

One comparison test found the report undercounted

OpenAI published early results from its measurement partners. The one below used the kind of fair comparison this page describes.

  • WorkMagic Dose

    2.3×

    more extra orders than the last-click report credited

    Last-click report
    Comparison test

    Illustrative proportions: 10 credited orders to 23 estimated extra orders shows the reported 2.3× comparison; these are not trial counts.

    WorkMagic compared similar areas with and without Dose’s ads, and set the result against a last-click report, which gives all the credit to the last ad someone clicked.

  • WorkMagic Dose

    67%

    of the extra purchases came from new customers

    Illustrative approximation: about two of every three extra purchases came from new customers, representing the reported 67%.

    WorkMagic counted these as buyers new to Dose, not people who already bought from it.

Reported by WorkMagic, a measurement company OpenAI works with, for the wellness brand Dose, and published by OpenAI. We haven’t seen the underlying data. It is one brand’s result, not a forecast for a local shop.

Before you spend more

Four questions to ask of any ad report

They work for ChatGPT Ads and for any other paid campaign.

  1. What does my own till say?

    Start from your till, bookings and online checkout. The ad report is a second opinion about which of those sales an ad should get credit for.

  2. What rules did the report use?

    Write down how long after a click a sale can be credited, and whether sales after a view without a click are counted. Longer windows credit more sales to the same ad, never fewer.

  3. How many credited sales came from regulars?

    An ad that mostly reaches people who already buy from you can look strong in the report and add little.

  4. Can I compare fairly before a bigger budget?

    Run the ad in one area or for one group of customers, and not for a similar one, over the same weeks. For a single shop the numbers will be small, so read the result as a direction, not proof.

Sources

Where this comes from

We wrote this from OpenAI’s two posts below. We didn’t run a campaign or see any partner’s data, and nothing here describes an Obenan product, integration or customer result.

  1. Building advertising for the way people use AI

    OpenAI Published Checked

    The sales-data integrations, measurement partners (AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin, plus Fospha, Measured and INCRMNTAL), the early geo-based tests, Negative Phrases, the brand-safety pilots, the visual-format test, the statement that ads do not influence answers, and WorkMagic’s new-customer figure for Dose.

  2. More ways to measure ChatGPT Ads

    OpenAI Published Checked

    The click-through and one-day view-through reporting windows and that they change reporting rather than how campaigns are optimized, the Kantar and Cint brand testing, and WorkMagic’s geo-lift comparison with last-click attribution for Dose.