The platform reports impressions, clicks, click-through rate and cost. Everything past the click, conversions and revenue, comes from your own analytics through tagged URLs and your pixel or Conversions API. But there is a catch that makes ChatGPT harder to measure than any other channel: a large share of AI-assistant referrals arrive with no referrer and get logged as Direct, so your analytics shows a fraction of the traffic you actually earned. Measuring this channel well means correcting for what it hides.
What the platform shows you
In-platform reporting is deliberately simple: impressions, clicks, CTR, CPM and spend, joined on 21 August 2026 by one-day view-through conversions at the campaign, ad group and ad levels. A view-through conversion counts when someone converts within a day of an eligible impression without a qualifying ad click taking credit, which matters on a channel where many users read the ad inside the conversation and visit later. Product feed campaigns that serve in the carousel unit also get card-level impressions and clicks since 4 September 2026, reported separately from the billable impression. Since mid-September 2026 the conversion columns take a reporting window: a 7, 14 or 30-day click window and a 0 or 1-day view window, applied to every conversion column at once. Changing the window changes the report only, never bidding or billing. On a CPC or oCPC campaign a view-through conversion stays supplemental reporting. The exception is oCPM, the conversion-optimized campaign billed on impressions, which OpenAI's Help Center describes as optimizing on actions that follow ad clicks and ad views. Each pixel or data source also gets an event quality assessment under Tools, refreshed daily over seven complete days, with warnings for gaps such as missing emails or click references. Since 30 September 2026 the same windows, and a good deal more, are readable through the Advertiser API, covered under the events you are not bidding on below. Beyond that, the platform tells you whether your ads are serving and being clicked. Whether they made money is a question it cannot answer. For that you connect the click to your own analytics, which is why the tracking setup in the conversion-tracking guide is the foundation everything else sits on.
Since September 2026 the reporting also splits by platform, and so does targeting. A campaign can be pointed at any of five: Android app, Android web, desktop web, iOS app and iOS web. Insights segments by the same five, which is the first time this channel has been able to answer whether the phone or the desktop is doing the work. Read it before excluding anything: on a channel where people read an ad mid-conversation and act later, a platform with poor same-session conversion is not necessarily a platform to switch off.
The first published benchmarks, and their small print
Until September 2026 there was no first-party number to measure against. OpenAI has now published three, each with an eligibility filter and a window that travel with it. Pixel plus Conversions API measured 57% more attributed conversions and a 28% lower measured CPA than the pixel alone, median advertiser, more than 1,000 advertisers at $500 or more of spend, late May to late June 2026. Conversion Optimization showed a median 44% lower CPA than click bidding within the same advertiser, across 27 advertisers who ran both in July. Product feed campaigns showed a 74% higher observed click-through rate than accounts without a feed, across more than 10,000 advertisers above $100 of spend and 100 clicks in a month.
All three are observational comparisons, and OpenAI says so in its own footnotes. Read the 44% with its sample of 27, read the 74% as a retail figure that a services advertiser cannot reproduce, and read the CAPI result as a measurement effect first and an efficiency effect second. The one number to keep for your own planning is the quietest: more than 10,000 advertisers cleared $100 of spend in a single month, which tells you how busy the auction already is. The requirements behind the 44% are in the Conversion Optimization guide.
The part of the journey search never saw
The reason this channel is worth measuring properly is that it reaches buyers where search never could. According to OpenAI's own breakdown of the shopping journey, only about a third of buying activity is the evaluation-and-purchase moment that classic search was built to capture. The other two-thirds is problem framing, product discovery and comparison, with product discovery the single largest slice at roughly a third of all activity, and comparison and single-product evaluation each around a fifth to a quarter.
The public figures point the same way. OpenAI has said that roughly 83% of the queries that trigger ads inside ChatGPT would never have triggered a Google Shopping ad, that a large share of ad moments are purely research-oriented, and that many users who start a conversation with no commercial intent develop buying signals before it ends (reported by PYMNTS). Keyword search only ever monetised the last stretch of the journey. ChatGPT reaches the problem-framing, discovery and comparison stages that come before it, which is why measuring it as if it were a bottom-of-funnel search channel undercounts what it actually does.
The events you are not bidding on
A campaign optimises for one event. Everything else a visitor does used to be invisible in OpenAI's own reporting, which is a real gap when the event you can bid on is not the event that pays you. Since 30 September 2026 the conversions endpoint will return attributed events that are not the campaign goal, so you can read attributed purchases while the campaign optimises for something earlier in the funnel.
The effect is larger than it sounds. Running this against one of our own paused campaigns, goal-only reporting returned zero conversions, and the same request with the event expansion returned 63 attributed events on the campaign. Both numbers are correct. They answer different questions, and until now only the first one was available.
Two cautions worth carrying, because OpenAI states them and a summary would drop them. Reporting by conversion date has limited coverage of non-goal events, so these reports should be read on ad-event time. And an empty result does not mean an event never happened; it can also mean the account's event history has not finished publishing. The attribution windows apply here too: a 7, 14 or 30-day click window and a 0 or 1-day view window, changing the report and never the bidding.
What this changes in practice is which events are worth instrumenting. An event you cannot optimise for is now still worth sending, because it can be measured and reported against the campaigns that produced it. The conversion-tracking guide covers getting those events in.
Tagging and naming
Until August 2026 the platform accepted only static parameters, written into each landing URL by hand. It now supports macros that fill in the campaign, ad group, ad and account IDs automatically at delivery, which removes the most error-prone part of the job. The discipline still applies to everything the macros do not cover: one consistent convention for source, medium and campaign across every ad, so a click can always be traced back to what earned it. Inconsistent tags are the quiet way a channel looks like it is not working when really it is just not joinable. The setup is covered in the conversion-tracking guide.
Reporting it next to your other channels
For a long time ChatGPT ads sat outside the usual reporting stack, which made it easy to treat as a side experiment. On 5 October 2026 OpenAI named its first set of measurement partners, and most performance teams will find a tool they already pay for on the list:
- +Sending conversion data in: Hightouch, Tealium and LiveRamp.
- +Click attribution, reporting and the Conversions API: AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin.
- +Full-funnel measurement: Fospha, Measured and INCRMNTAL.
- +Geo-based incrementality experiments, which OpenAI describes as early-stage: Haus, Measured and WorkMagic.
- +Brand lift, in early testing: Kantar and Cint.
If you already run one of them, connecting this channel is the fastest way to stop judging it in isolation, which is usually how a top-of-funnel channel ends up unfairly compared to a bottom-of-funnel one.
OpenAI published early results with the list, each measured by the partner on a single advertiser. DV Rockerbox put WeightWatchers' attributed cost per acquisition 15.3% below its blended paid-search benchmark. WorkMagic's geo-lift study for the wellness brand Dose estimated 2.3 times more incremental orders than last-click attribution captured, with 67% of incremental purchases coming from new customers. Triple Whale found that 93% of Portland Leather's visitors from ChatGPT ads were new. OpenAI's own early analysis adds that 52.7% of eligible one-day view-through conversions happened within an hour of the impression, a good reason to keep the one-day view window on in your reports.
Finding the AI traffic that hides
Even with perfect UTMs, you will under-count. The mechanism is not in dispute: many AI assistants strip the referrer header, so the visit lands in Direct and never reaches an AI channel. The size of the gap is another matter. The most-cited measurement, a single 2026 study of roughly 446,000 visits, put the share of AI referrals arriving with no referrer near 70%, and it has not been independently reproduced. Other widely quoted figures have aged worse: one vendor's often-repeated claim that analytics under-counts AI traffic by a fixed percentage was removed from its own page in August 2026.
So treat every published multiplier as a measurement of someone else's site. What holds up is the method:
- +Recover the AI visits that do keep their referrer with custom channel rules for the main assistants.
- +Size the gap on your own site instead of importing a multiplier: compare server logs against analytics over the same window, which is the only measurement that is actually about you.
- +Treat the analytics number as a floor. Corroborate with landing-page and behavioural evidence before you put a number in a board deck.
This matters because the hidden traffic is the good traffic. AI-referred visitors arrive further along the decision and convert well above other sources, so under-counting them makes your best channel look like your weakest. That direction is well established even where the exact multiplier is not, which is the honest way to hold it: act on the direction, and measure the size yourself before you quote one. Correcting for it is exactly the AI-attribution work we already do for organic visibility, now pointed at your paid results. When you want the full picture built and reported for you, that is part of the service.
