The Conversions ChatGPT Ads Will Never Show You
By Keshav Parsai 9 min read
In brief
Measure ChatGPT Ads demand that never clicks by tracking branded search lift, direct traffic and controlled campaign timing.
Last verified: 12 September 2026 | Version: 1.0 | Next scheduled review: 12 October 2026
A user asks ChatGPT how to handle multi-entity consolidation. They get an answer. Under it sits your ad. They read the brand name, do not click, finish the conversation, and three days later type your name into Google and book a demo.
Ads Manager records an impression. Your branded search campaign records a conversion. Nothing in either system connects them, and nothing ever will, because the connecting event happened inside a conversation OpenAI does not report on.
This is not a tracking failure you can fix with better tagging. It is a structural property of a placement that sits inside a task the user does not want to abandon.
Why this channel produces more of it than search does
Call the behaviour remembered demand: an ad response that consists of retaining a name rather than following a link.
Every channel produces some. This one produces more, for a reason specific to the placement. On a search results page, clicking is the purpose of the visit. The user came to leave. In ChatGPT, the user is mid-task, often several exchanges into building an understanding, and leaving costs them their place. The ad is the interruption, not the destination.
The view-through window is fixed at one day and cannot be extended, which means the platform's own view credit expires long before a considered B2B decision resolves. Anything remembered on Tuesday and acted on the following Monday is invisible to Ads Manager by design, not by misconfiguration.
The one published estimate, and what is wrong with it
The figure circulating in this category comes from Seresa, a server-side tracking vendor, which stated in 2026 that only about 40 percent of conversions attributable to ChatGPT Ads happen in the immediate post-click session, with the remaining 60 percent occurring hours, days or weeks later.
Treat that as Reported, permanently, and read it with three things in mind.
The source sells the remedy. Seresa's published pitch on the same topic ends by inviting readers to talk to them about building server-side oppref capture. A vendor diagnosing an attribution gap that its product closes is not disqualified, but it is not disinterested either.
The methodology is absent. The claim is sourced to "early advertiser data" with no advertiser count, no date range, no vertical breakdown, no statement of whether conversions were counted through the pixel, the Conversions API or both, and no definition of "attributable". Without those, 40 percent is a number shaped like evidence rather than evidence.
It is also measuring a different thing than the one people quote it for. As written, the claim is about timing, that conversions happen after the click session, which a longer attribution window partly recovers. It is widely paraphrased as a claim about visibility, that Ads Manager sees only 40 percent of true ROI, which is a stronger and less supported statement. The paraphrase is doing work the original sentence does not do.
Our honest position: the direction is almost certainly right and the magnitude is unverified. Use the figure to argue that the gap exists. Do not use it to size a budget.
What the pixel structurally cannot reach
Independent of any vendor's number, four populations are outside the pixel's reach at any window length.
The impression-only response. Saw the ad, never clicked, converted later. Recoverable for one day through view-through credit and not at all after that.
The cross-device response. Saw the ad on the ChatGPT iOS app, converted on a work laptop. oppref is stored in a first-party cookie on the device that received the click, so there is no join.
The delayed branded response. Clicked, did not convert, returned weeks later through branded search or direct. Recoverable only if it falls inside your configured click window, which for a 74-day sales cycle it usually will not.
The referred response. Saw the ad, mentioned you to a colleague, the colleague bought. Not measurable on any channel and not a ChatGPT Ads problem, but it is disproportionately present in B2B, where the person in the conversation is often not the person who signs.
Branded search lift, which is the practical proxy
You cannot recover the individual conversions. You can measure whether the population moved. The mechanism is branded search lift, and it works because remembered demand has to resurface somewhere, and for B2B it resurfaces as someone typing your name.
The procedure, in order:
Establish the baseline before you spend. Pull at least eight weeks of pre-launch branded search volume. Two sources, not one: Google Search Console impressions and clicks on queries containing your brand, and Google Ads impressions on your exact-match brand campaign. Also pull GA4 Direct sessions to the homepage and to any page a person would guess at. You need the baseline to have a shape, including its weekly seasonality, not just a mean.
Hold everything else still. No PR, no launches, no email sends, no conference, no organic content push in the test window. This is the part that fails in practice. A lift test contaminated by a webinar is not a lift test.
Run the spend in a step, not a ramp. A gradual increase is unreadable. Launch at the budget you intend, hold it flat for at least four weeks, and mark the start date.
Read with a lag. Expect nothing in week one. Remembered demand resurfaces on the buying cycle's clock, not the campaign's. If your median time to enquiry is three weeks, week four is the earliest honest read.
Compare on rate, not volume. Branded query volume is contaminated by everything. The more stable read is the ratio of branded search impressions to your ChatGPT Ads impressions, tracked weekly. If that ratio moves against a flat baseline, you have a signal.
What this gives you is a directional answer for the account as a whole. What it does not give you is anything campaign-level or ad-group-level, so it cannot be used to allocate between ad groups. For causal claims at a level you can act on, the method is an incrementality design, covered in our post on measuring incrementality without a search term report.
What to do with the gap in practice
Three decisions change once you accept that a fraction of the return is unobservable.
Do not optimise toward the observable fraction alone. If you bid only on conversions the pixel sees, you will systematically favour ad groups that produce fast clickers over ad groups that produce considered buyers. The bias is invisible because the metric that would reveal it is the one you are missing.
Put the gap in the reporting, not in the footnotes. A monthly report with one line reading "reported conversions: 34; branded search index versus pre-launch baseline: +11 percent, four-week lag applied" is an honest document. One that reports 34 and says nothing else implies a precision that does not exist.
Set a floor you would accept on reported numbers alone. Because the unobserved portion cannot be sized, the defensible discipline is to require the channel to pay back on what you can see, and treat everything else as upside you are not budgeting against. Any other rule lets an unverifiable multiplier justify any amount of spend.
What we cannot tell you
- The size of the unmeasured fraction. The only published figure is a vendor estimate with no disclosed methodology. Nobody has published a replicable measurement.
- Whether view-through credit at one day captures a meaningful share of impression-only response. OpenAI publishes no distribution of time-to-conversion.
- How often an ad is seen but not clicked. Impressions and clicks are reported, so you can compute a rate, but you cannot know how many impressions were actually read inside the conversation.
- Whether cross-device recovery is on OpenAI's roadmap. Nothing announced. Absent.
- Any first-party lift measurement of our own. InPromptAds runs no campaigns and has no branded search baseline to test against.
Quick answers
What is the dark funnel in ChatGPT Ads? The conversions caused by an ad that never touch the tracked path: seen and not clicked, converted on another device, or converted weeks later through branded search or direct. Ads Manager cannot see any of them.
Is ChatGPT Ads really only 40 percent visible? One vendor, Seresa, published in 2026 that roughly 40 percent of attributable conversions happen in the immediate post-click session. No methodology was disclosed and the vendor sells server-side tracking. Treat the direction as plausible and the number as unverified.
Why is this worse on ChatGPT than on Google? Because the user is mid-task. On a search results page, clicking is the point of the visit. In a conversation, leaving costs the user their place, so retaining the brand name is a more common response than following the link.
Can a longer attribution window fix it? Partly. A longer click window recovers delayed conversions from people who did click. It recovers nothing from impression-only response after day one, and nothing cross-device.
How do I measure branded search lift? Baseline eight weeks of branded impressions in Search Console and Google Ads, hold all other marketing still, launch spend as a flat step, and read the branded-to-ChatGPT impression ratio from week four onward.
Should I add an uplift multiplier to reported ROI? No. The multiplier would be unverifiable, and an unverifiable multiplier can justify any budget. Require payback on observed conversions and treat the rest as unbudgeted upside.
Sources
| Claim | Source | Tier |
|---|---|---|
| Roughly 40 percent of attributable conversions occur in the immediate post-click session, 60 percent later | Seresa, attribution write-up, 2026. Vendor sells server-side conversion tracking. No methodology disclosed | Reported |
| View-through window fixed at one day and not configurable | OpenAI documentation, August 2026, as reported by PPC Land | Confirmed, primary for the rule; Reported for the date |
oppref stored in a first-party cookie on the device receiving the click |
OpenAI Help Center, Conversion Measurement, 2026 | Confirmed, primary |
| No conversation data, no adjacency reporting, no cross-advertiser benchmarks | OpenAI Ads Manager documentation, by absence | Absent |
| Ads serve to Free and Go plan users across the reported serving markets | OpenAI Help Center, 2026 | Confirmed, primary |
| Branded search lift procedure and the impression-ratio read | Inference, ours | Inference, ours |
| Remembered demand as the dominant response mode in a conversational placement | Inference, ours | Inference, ours |
Related reading
- Why Conversational Ads Work Differently: The Psychology of the Placement
- Measuring Incrementality on a Channel With No Search Term Report
- What OpenAI's Attribution Windows Actually Count
- How to Measure ChatGPT Ads, and What You Cannot Measure Yet
- Do ChatGPT Ads Work?
Changelog
12 September 2026, v1.0. First publication. Names remembered demand as the placement's dominant response mode, audits the single published visibility estimate against its own missing methodology, and sets out branded search lift as the practical proxy.
Field kit
Tools
Site
OpenAI (primary)
Keshav studies how AI systems retrieve, verify, and cite brand information. At InPromptAds, he leads source research and turns platform documentation into practical guidance for advertisers.
