Running a ChatGPT ads pilot: test structure, budget, and success metrics

A ChatGPT ads pilot is a structured 30-day experiment that spends enough at the platform’s documented floors to see if this channel can produce repeatable, economically viable conversions for your business. This plan assumes you already run paid elsewhere and focuses on test design, not setup or tracking implementation, which are covered in a separate setup guide and a dedicated tracking guide.

Decisions a ChatGPT ads pilot must answer

By the end of the pilot you need to know whether ChatGPT ads can produce conversions at a return that fits your unit economics, and whether that performance is consistent enough to justify more spend. For ecommerce, Temper And Forge’s practice is to decide on conversion volume and the return on ad spend (ROAS) from that conversion volume: conversions must materialize and return against spend, and inconsistent conversion that cannot be extrapolated to full spend volume is a kill, not a scale.

You also want a read on fit by business type. As of October 2026, Temper And Forge’s early testing across clients points to local service, B2C, and general awareness as the most prominent fit, while ecommerce remains under active test. The pilot should therefore be framed to answer both “can this channel work on my economics” and “does my business type look like one of the early fits.”

Clicks and traffic quality are secondary questions. Per OpenAI’s documentation, you can buy traffic on a Clicks objective, but without conversion tracking your only readable metrics are traffic metrics, so you cannot answer the core business question. A pilot that stops at click metrics is a traffic test, not a channel viability test.

Pilot structure: campaigns, ad groups, and objectives

Per OpenAI’s help, campaigns must be created with 1 of 3 objectives: Views, Clicks, or Conversions. The objective controls how the system optimizes and bills (per 1,000 impressions for Views, per valid click for Clicks, and toward a conversion event while billing per click or per 1,000 impressions for Conversions), and it is locked at creation and cannot be changed afterward. This lock means you should design the pilot around separate campaigns if you need to test different objectives.

For a first pilot focused on performance, a single Conversions campaign is usually the cleanest structure, because it lets the system optimize toward your defined conversion event while still billing per click or per 1,000 impressions. If you also want to compare traffic quality or upper-funnel reach, you can add 1 additional campaign on Clicks or Views, but keep the bulk of spend on the objective that matches your primary decision metric.

OpenAI recommends multiple ad groups per campaign to test themes and messaging. A practical pilot structure is:

• 1–2 campaigns total, each with a single objective.
• 2–4 ad groups per campaign, each representing a distinct audience theme or message angle.
• Shared budgets at the campaign level, with ad groups competing for delivery based on performance.

This structure gives you enough variation to see what resonates without fragmenting a modest budget. Since objectives are locked, avoid creating many small campaigns that each struggle to reach the platform’s minimum daily budget.

Budget size and duration for a readable sample

Per OpenAI’s documentation, the minimum daily budget is $25 USD, and daily budgets are enforced over a 7-day window. A single day can spend up to 2x the daily budget, but the Sunday–Saturday week is capped at 7x the daily amount. OpenAI also recommends starting max cost-per-click (CPC) bids in the $3–5 range.

Temper And Forge’s pilot practice for small-to-medium businesses is a maximum test budget of $20–40 per day, but the platform’s $25 minimum makes the effective on-platform range $25–40 per day. At that range and OpenAI’s recommended $3–5 max CPC, a 30-day test spends roughly $750–1,200 and buys roughly 150–400 clicks. This is the scale Temper And Forge uses to get a first directional read on conversion behavior and ROAS.

The 30-day duration pairs with the platform’s 7-day budget mechanics and reporting lag. Per OpenAI’s help, metrics refresh about every 15 minutes, but spend can lag by 7–8 hours, and weekly totals are the honest read unit given the 7-day budget enforcement. A shorter test risks confusing pacing adjustments with performance shifts, which is why Temper And Forge’s practice aligns with a four-week read.

For deeper arithmetic on how this test size interacts with your economics and potential outcomes, see the dedicated cost and budget math reference.

Prerequisites before you start spending

Conversion measurement must be in place before launch if you want to judge the channel on business outcomes. Per OpenAI’s documentation, this means configuring the pixel and/or Conversions API before you start the campaign. UTM parameters persist on clicks, which lets you validate traffic and conversions in your own analytics in parallel.

A common misconception is that a pilot can skip conversion tracking and judge success on clicks alone. On this platform, without the pixel or Conversions API you are limited to traffic metrics, so you cannot evaluate conversion volume or ROAS. For a pilot whose decision metric is conversion volume plus ROAS, skipping tracking turns the test into a traffic experiment that cannot answer whether month two should happen.

Beyond tracking, you should have your conversion events defined and prioritized. For ecommerce, that is typically a purchase event. For lead or local service flows, it might be a form submit or booked appointment. The key is that the same event definition is used in the platform and in your own analytics so that ROAS and conversion volume line up with what you see in other channels.

Week-by-week reading without misinterpreting lagged data

As of October 2026, OpenAI’s help states that metrics update roughly every 15 minutes, but spend reporting can lag 7–8 hours. Combined with the 7-day budget averaging, this makes daily views noisy and misleading. Temper And Forge’s practice is to read performance on weekly totals, not daily, which is consistent with the cost doc’s four-week read.

In week 1, the goal is to confirm delivery and basic click behavior, not to call the test. Check that campaigns are spending close to their 7-day budget cap, that you are seeing clicks at roughly the expected volume from your bids, and that conversions are starting to register. Because a single week is influenced by the platform smoothing budget across the 7-day window, a week of data is not enough to judge the channel.

In weeks 2 and 3, you are looking for patterns in weekly conversion volume and ROAS. Compare each week’s totals to the prior week, knowing that intra-week swings can be pacing rather than performance. If you see conversions but the weekly ROAS and conversion counts jump around so much that you cannot extrapolate them to your full test budget, Temper And Forge’s practice is to treat that inconsistency as a negative signal rather than a reason to scale.

By week 4, you should have 4 weekly data points on spend, clicks, conversions, and ROAS. Read them as a sequence: is conversion volume present, and is ROAS within a band that your economics can support, with enough consistency that you can reasonably project performance if you maintained or modestly increased spend? This weekly lens avoids overreacting to the 7–8 hour spend lag and the 7-day budget smoothing that can distort daily views.

Scale or stop: how to decide after the pilot

The pilot’s primary outcome is a go or no-go decision for continued spend at similar or higher levels. For ecommerce tests, Temper And Forge’s practice is explicit: TF kills based on conversion volume and the ROAS from that conversion volume. Conversions must materialize and return against spend, and inconsistent conversion that cannot be extrapolated to full spend volume is a kill, not a scale.

Outcomes that justify scaling are those where weekly conversion volume is present and ROAS from that volume is within a range that your economics can support, with enough week-to-week stability that you can project performance to your ongoing budget. In this case, “scale” usually means maintaining the pilot budget for another period or modestly increasing it while keeping the same objective and structure.

Outcomes that end the test are those where conversions fail to appear at all over the 30-day window, or where conversions appear but ROAS and conversion counts are so inconsistent that you cannot responsibly extrapolate them to your full spend. In those cases, Temper And Forge’s practice is to stop or significantly rethink the channel rather than chase better results with more budget.

Temper And Forge’s current fit prior, based on early testing, is that local service, B2C, and general awareness use cases show the most promising fit, while ecommerce remains under active test. That prior shapes expectations going into the pilot, but the decision to scale or stop is still made on your specific conversion volume and ROAS behavior over the four-week read.

Temper And Forge’s practice on ChatGPT ads pilots is to size SMB tests at $25–40 per day for 30 days, read performance on weekly totals because of the platform’s 7-day budget mechanics and 7–8 hour spend lag, and make ecommerce decisions strictly on conversion volume plus ROAS, treating inconsistent, non-extrapolatable conversion as a reason to kill rather than scale, while continuing to test fit across local service, B2C, general awareness, and ecommerce.