Attribution

How do I tell if ChatGPT ads are driving demo requests that show up as direct traffic or branded search?

By The Tofu team ·

Run a holdout test, since OpenAI's docs offer no lift study and ChatGPT Ads counts conversions only up to 30 days after a click or 1 day after a view. Later direct visits and brand searches go uncounted. So run ads in some countries or US markets and keep similar ones dark, or switch ads on and off by period. Then compare how branded searches, direct visits and demo requests change in each group. Extra demo requests where the ads ran are the lift, whatever OpenAI reported.

Ask Tofu next

Pick a question and Tofu Ads Agent answers it in chat.

As of November 2026.

Why would ChatGPT ad leads show up as direct traffic or branded search?

Because many buyers see an ad, don't click, and come back later another way: they search your name on Google or type your web address. OpenAI credits a conversion only inside its attribution windows: 7, 14 or 30 days after a click (30 is the default) and 0 or 1 day after a view (1 is the default). A buyer who saw your ad on Monday and searched your brand the next week isn't in OpenAI's numbers.

Your analytics won't connect them either. Google Analytics 4 says "All attribution models exclude direct visits from receiving attribution credit," unless every visit on the path was direct. A branded search gets credited to Google. An ad click that arrives with no tracking tags and no referrer can land in Direct too; to keep tracked clicks out of it, see how to tell ChatGPT ad clicks from organic ChatGPT traffic in GA4.

Does OpenAI offer a lift study for ChatGPT ads?

Not that we could find. OpenAI's reporting docs cover impressions, clicks, spend and conversions inside those windows, with breakdowns by country, device, platform or product. They describe no lift study, holdout or incrementality report. We couldn't read OpenAI's help center to check whether Ads Manager offers one, so ask your OpenAI contact if you have one. Google runs lift studies for its own ads; see Google's Conversion Lift for Search and Performance Max for how those work.

What should I measure?

Pick signals you can split the same way you split the test:

SignalWhere to find itSplits below country?
Branded searches (clicks and impressions)Search Console, Performance report: click the filter bar, pick Query, then BrandedNo. Google groups it by "the country where the search originated"
Branded search ad impressionsGoogle Ads, your brand campaign's Matched locations reportYes, but Google says "Metrics listed below country level may not add up to campaign totals"
Direct visitsGA4, the Direct channel (source "(direct)", medium "(none)" or "(not set)")Yes, by region or city
Demo requestsYour CRM, by the country or state on the formYes

Search Console's branded filter has limits. Google sorts queries into branded and non-branded with its own AI-assisted system, not your keyword list. It works only on top-level properties with enough impressions, and its data starts in March 2025. If you don't see it, use a custom (regex) query filter with your brand's spellings.

How do I set up a geo holdout?

These steps are our suggested method; OpenAI doesn't publish one.

  1. Pick test and control areas that looked alike over the past few months in branded searches, direct visits and demo requests. Use countries if you sell in several. If you sell in one, use smaller areas: OpenAI's location targeting covers countries, regions (such as California) and markets (metro areas such as "San Francisco - Oakland - San Jose").
  2. Target only the test areas. OpenAI's API reference says a campaign's ads show "only in the locations specified in campaign targeting." To run in a country minus your control regions, include the country and exclude those regions.
  3. Record a baseline of several weeks before the ads start.
  4. Hold everything else steady in both groups: Google, LinkedIn and Meta budgets, events and outbound pushes.
  5. Run it longer than your usual gap between a buyer's first visit and a demo request, so delayed demand has time to show.
  6. Compare the change, not the totals. The lift is how much the test areas rose over their baseline, minus how much the control areas rose over theirs.

Search Console splits only by country. In a test across US markets, use GA4 regions, your brand campaign's matched locations and your CRM's states instead.

OpenAI's docs don't say how it decides where someone is, so some people in control areas may still see your ads. That shrinks the gap between the groups, so a test like this tends to understate the lift rather than overstate it.

Can I switch ads on and off instead of splitting by area?

Yes, if your market is too small to split. Run ads for a few weeks, pause them for the same length, and repeat. A campaign's start and end times set when its ads can show. It's a weaker test: seasonality, other campaigns and holidays move your numbers between periods, and buyers who saw ads in an "on" period can still book in the "off" one. Alternate several times and compare each "on" period with the "off" periods on either side.

How big does the lift need to be before I can see it?

Bigger than your normal week-to-week swings. Look at last quarter's weekly demo requests in each area. If they move by 10 either way and you expect ChatGPT ads to add 3, a test of a few weeks won't show it. Branded searches come in far larger numbers than demo requests, so watch them as the early signal. For the analysis, GeoLift is an open-source tool from Meta Open Source that calculates lift at a geo level; Meta says it isn't a Meta product and is for research only.

How do I turn the result into a cost per demo request?

Divide the ChatGPT spend in the test areas by the extra demo requests the test found. Then compare that with OpenAI's own count for the same dates:

  • Lift well above OpenAI's count: the ads drive demand OpenAI can't see, through direct visits and branded search.
  • Lift near zero while OpenAI reports conversions: those leads would probably have come anyway.

Pull OpenAI's numbers twice, with the view window at 1 and at 0, to see how much of its count rests on views alone. Our page on tracking conversions from ChatGPT ads shows those settings and how to compare them with your CRM.

Can a "How did you hear about us?" field replace a holdout?

No, but run one alongside it. Buyers who name ChatGPT give you a quick read. Many won't remember an ad from weeks ago, though, and some who say "ChatGPT" will mean its answers, not your ad. For the field and other checks after the click, see whether you can see which ChatGPT conversations your ads appeared in.

Will this show whether ads get my brand into ChatGPT's answers?

No. A holdout measures demand on your site and on Google. Whether ads change what ChatGPT says is a separate question, covered in whether running ChatGPT ads gets your brand mentioned in its answers.

Sources

Let an agent build your next campaign.

Tofu Ads Agent reads your CRM, plans the funnel, builds the audience and creative, and stages full campaigns across LinkedIn, Google and Meta for your approval.