ChatGPT Paid Ads: measuring ROI without the old signals

Why the old map no longer fits #

A fox does not hunt the same hedgerow twice and expect the same trail.

For two decades, performance marketers followed familiar scents: third-party cookies, cross-device identity graphs, tidy GA4 channels, keyword chains, and mature multi-touch attribution. Those signals shaped the strategy. They lit the path from query to click to conversion.

ChatGPT Paid Ads quietly removed most of those signals from the map.

ChatGPT runs as a logged-in, first-party conversational interface where ad selection happens inside OpenAI's own environment. Conversations and identities are never handed to advertisers, so third-party cookies cannot participate in targeting or attribution at all. There is no cookie to drop, no open-web breadcrumb to follow.

The keyword chain has gone quiet too. Similarweb's analysis finds ChatGPT ads are built on conversational intent, not discrete queries, and that 83 percent of ad-triggering prompts would never activate a classic Google Shopping ad. A single sponsored unit sits inside a conversation, not in a crowded results page.

In plain terms, three things you used to lean on have quietly disappeared:

• No cookies. You cannot follow a person around the open web, because the chat never hands one over.
• No keywords. Ads answer the mood of a conversation, not a tidy search term you can bid on.
• No clean path. The neat journey from click to sale has folded into a back-and-forth chat with very few clicks at all.

A capability gap, channel by channel #

Put ChatGPT next to Google and Meta and the difference is plain to see. The older platforms have spent years building a deep toolkit. ChatGPT is starting almost from scratch.

Here is the gap, side by side:

• Cross-device identity. Google and Meta can follow a person from phone to laptop. ChatGPT shows you no cross-device reports at all.
• Multi-touch paths. The older platforms map every touch on the way to a sale. ChatGPT gives you no path data, so the journey stays hidden.
• Audience reporting. Google and Meta hand you age, interests and lookalikes. ChatGPT offers conversational context and a little location data, and not much more.
• Retargeting pools. The classic networks rebuild warm audiences from years of behaviour. ChatGPT lets you match topic and context inside the chat, but you cannot build a pool to chase later.

In short, ChatGPT starts life cookieless from day one. That sounds like a loss, and in some ways it is. It is also a clean slate, and a fox reads a fresh field faster than a cluttered one. If you want a partner who plans for that clean slate, this is the heart of how folkfox approaches paid social measurement.

GA4 and the dark AI trail #

Most AI visits arrive at dusk, leaving no footprints.

Web analytics tools systematically undercount AI-originated sessions, and the reason is mechanical, not mysterious.

Referrer headers are stripped. In-app browsers stay sealed. Zero-click behaviour leaves no breadcrumb. So GA4 quietly files the visit under Direct and the trail goes cold.

Workshop Digital, studying 181.6 million GA4 sessions, watched ChatGPT and Perplexity traffic leak into "(not set)" and Direct across 22 sites. The signal is real, the volume is rising, and the standard report hides it.

That hidden volume matters. Practitioners estimate true AI-originated sessions may be two to three times higher than a standard GA4 report suggests.

The new AI Assistant channel, and its limits #

On 13 May 2026, Google shipped a new AI Assistant default channel group. When GA4 detects a recognised AI assistant referrer it sets the medium to ai-assistant and routes the session into the AI Assistant channel.

It helps. It does not heal.

The channel only works when referrer headers exist. Desktop clicks from chatgpt.com usually qualify, while the iOS and Android apps frequently strip referrers, so those sessions still slip into Direct. And AI Overviews remain filed under Organic Search, not AI Assistant, splitting the AI picture in two.

The OpenAI pixel, CAPI and Ads Manager #

OpenAI now hands advertisers a few honest instruments, and one honest limitation.

Within six months, ChatGPT Ads moved from impression-only pilots to a self-serve platform with an OpenAI pixel, a Conversions API, and cost-per-action bidding. The post-click picture is now possible. The full-journey picture is not.

The OpenAI pixel #

Digiday's review of the OpenAI pixel code shows support for standard commerce events: page viewed, lead created, order created, subscription created, trial started, and registration, fired by JavaScript on defined pages.

It listens for configured events and sends anonymised conversion data back to OpenAI. It cannot run where JavaScript is blocked, such as some in-app browsers or strict content security policies. And it sees post-click behaviour only: never the zero-click influence inside the conversation itself.

If the jargon is piling up, here are the three tools in one quick glance:

The OpenAI pixel

A snippet of code on your site that whispers back to OpenAI whenever someone takes an action, like a sign-up or a sale.

The Conversions API (CAPI)

The same idea, sent server to server, so it still works when a browser blocks the OpenAI pixel. Belt and braces.

Ads Manager

The dashboard where you set budgets and read the results. It only shows conversions once the OpenAI pixel or CAPI is feeding it.

The Conversions API #

OpenAI calls its server-side Conversions API a more reliable tracking source than the OpenAI pixel alone, and recommends running the two together, mirroring Meta's familiar pattern.

CAPI receives server-to-server posts from your backend or tag server, so conversions still register when browsers, tracking prevention, or adblockers smother the client-side call. It asks for the same discipline Meta and Google demand: authentication, event schemas, and deduplication.

What Ads Manager reports today #

Per OpenAI's own launch and Search Engine Land's coverage, Ads Manager today exposes impressions, clicks, CPC bidding, spend, and basic performance. Conversion counts, CPA and ROAS appear only once the OpenAI pixel or CAPI is installed.

What is missing is the part marketers miss most: audience breakdowns, demographic reporting, multi-touch attribution, assisted-conversion paths, view-through metrics, and cross-channel reporting. That gap is exactly why a strong organic presence pays off here, and why folkfox content marketing treats AI visibility as a measurement asset, not an afterthought.

The high-intent minority who click #

A fox does not chase every rustle. It reads the right one.

Few users click out of ChatGPT, but those who do tend to be hunting with purpose. The minority who step onto the path arrive further along their decision than the typical search visitor.

Criteo's March 2026 release, drawn from 500 US retailers, put AI-referred conversion at about 1.5 times other referrals. A May follow-up found categories such as consumer electronics, lifestyle and home edging toward two times the rate of traditional search.

Independent data echoes the pattern. Verityscore, using Adobe Analytics, reports AI-referred retail traffic converting 42 percent better, spending 48 percent more time on site, and generating 37 percent more revenue per visit. Loamly tracked AI referrals climbing from 49 percent worse than non-AI traffic to 31 percent better by the 2025 holidays: a rapid maturation of visitor quality.

The zero-click influence problem #

The largest part of the trail is the part no OpenAI pixel can see.

Here lies the deepest dark. Most people who meet a brand inside ChatGPT never click at all. They read the answer, note the name, and move on, carrying the impression with them.

The same controlled experiment found ChatGPT users made correct choices more often and faster than Google users, especially on complex tasks. The implication is sharp: the few who do click are likely further down the funnel, and 88 percent finish their research without ever leaving the interface.

This is not a ChatGPT quirk. It is the wider weather of search. Bain reports around 60 percent of traditional searches now end without a click, and Forrester notes AI summaries reducing organic traffic by 15 to 25 percent.

Attribution beyond last-click #

Last-click is a floodlight pointed at the wrong end of the path.

Last-click attribution gives all the credit to the final click before a sale. It has a known flaw: it under-rewards everything that happens earlier and over-rewards branded search and Direct. Picture a relay race where only the last runner gets a medal. For ChatGPT, the low click rate makes this worse, because the channel does its real work early in the race, then quietly hands the baton to whatever clicked last.

Models that give the mid-journey its due #

Position-based and time-decay models hold up better than pure last-click when identities are fragmented, because they grant partial credit to mid-journey touches. Some experts propose treating AI as a mid-funnel introducer or closer, testing models that give 20 to 40 percent credit to AI touches even when they are not the final click.

GA4's data-driven attribution can, in principle, credit AI Assistant for assists, but only for the AI touches it can actually see. Dark traffic and Direct contamination mean most AI influence never enters the path model at all.

Incrementality and marketing-mix modelling #

When click data is sparse, the proven answer is to measure lift, not clicks. Geo holdouts, lift studies and marketing-mix modelling have measured TV, out-of-home and podcasts for years. They suit ChatGPT well.

The catch is scale. ChatGPT inventory may be too small in many markets for clean geo-split tests today, so early incrementality work is likely to fold ChatGPT into a broader AI search or conversational bucket within the model.

AI visibility tools #

When you cannot count the clicks, count the citations.

A new class of tool measures how often a brand is named, ranked and recommended inside AI answers. These do not replace conversion tracking; they complement it, quantifying the share of voice that pixels cannot.

Across the category, the pattern holds. Tools like LLMrefs, Profound and xSeek quantify AI share of voice and crawler behaviour. They sit beside GA4 and the platform pixels, not in place of them. Less noise, more nuance.

Used well, these tools are the bridge between paid and organic. They show whether your ads, your articles, and your reputation are all pulling in the same direction inside the answer engines. That joined-up view is the backbone of folkfox SEO and GEO services, and it matters most in trust-heavy markets like iGaming marketing where every citation is scrutinised.

Setting targets before benchmarks exist #

A fox does not need a floodlight when the scent is strong.

As of June 2026, no independent, verticalised CPA or ROAS benchmarks exist for ChatGPT Paid Ads. The figures floating in early case studies are anecdotal, retail-skewed and statistically thin. First Page Sage estimates CPCs around three to five US dollars, but publishes no clear median CPA by sector.

So set a ceiling, then test downward.

Borrow your existing Google and Meta CPAs as the ceiling for early ChatGPT campaigns, then iterate down. Treat the first spend as learning budget for channel fit, not performance spend. Frame Q2 and Q3 2026 campaigns, as agencies like Opollo and Launchcodex do, as structured experiments: the goal is to learn the cost to incremental branded search, the AI citation uplift, and the conversion quality relative to search and social.

Seven moves for a measurement scaffold #

Build the framework before the first campaign, not after the first confusion.

A resilient setup marries the OpenAI pixel and CAPI with GA4 custom channels, UTMed landing pages and AI visibility tools, anchored in proxy benchmarks from contextual and native channels. Here is the order folkfox follows.

Frequently asked questions #

The final takeaway #

Only a small, bright wedge of ChatGPT's commercial impact appears as tracked last-click conversions.

A much larger, darker ring of AI-assisted and zero-click influence surrounds it, and it must be inferred from branded search, AI visibility and incrementality rather than from pixels alone.

The fox does not mourn the cooling cookie trail. It learns the new one. It pads past pixel panic, follows faint footfall, and reads the revenue rustle before the rest of the market hears it. Behind the data is a human asking ChatGPT a real question, and the brands that show up early, get cited often, and measure honestly will read that trail before the trap snaps shut.

Less guessing, more knowing. That is the whole game now.

Ready to outfox soulless measurement?

folkfox builds measurement scaffolds for ChatGPT and AI-first channels that read the whole trail, not just the last click. The fox guides. You stay the hero.

Start the conversation Explore paid social

Keep exploring the den #

If this trail led somewhere useful, here are the next paths worth padding down:

Want the bigger picture first? Start with folkfox brand strategy, learn how we run app marketing for retention-led products, or see our work across FinTech marketing and music industry marketing. When you are ready, get in touch with folkfox.