Most small advertisers waste their first month on ChatGPT ads by testing five things at once. I used to do the same on Google: three offers, two landing pages, four ad angles, $400 gone, and no idea what failed. This list is for lean teams spending their own money and evenings. On a small budget, each test gets one question, a spend cap and a kill rule written before launch. Nothing scales without deliberate approval. Treat the caps below as ceilings you monitor and enforce, not permission to let an account run unattended.

1. Run one offer to one page

Question: Can a click from a chat become a qualified conversion? Start with your best offer and one landing page. Do not add a second offer because you want “coverage.” I used to tell clients to start with three ad groups for that reason. I was wrong. Coverage on a small budget buys three inconclusive tests when you could have bought one answer.

The audience is narrower than Google Search: only logged-in adults on Free and Go plans see ads; Plus, Pro, Business, Enterprise and Edu accounts do not see the audience breakdown. OpenAI publishes no cross-advertiser benchmarks, and Ads Manager is still beta. A recommended starting max bid of $3 to $5 is not a click price; observed CPCs run from $3 to $22 see the reported range. Set your ceiling before a fast delivery day sets it for you.

One coaching firm spent $500 in just over an hour at about $3.50 CPC for roughly 140 reported clicks. About 13% appeared in analytics, and the firm got zero leads read the small-budget test. That result is painful, but it answers more than a month of unexplained tweaks would.

First step: Choose one offer, one page and a $250 to $500 spending ceiling. Kill rule: Stop at the ceiling with zero qualified conversions. Do not add copy variants to rescue the test.

2. Score ten leads before you count them

Question: Are these leads worth following up? A cheap lead that needs six follow-ups and goes nowhere is not a bargain. Someone chatting may be earlier in a decision than someone who searches emergency plumber near me. Count names in a form if you want; make the budget decision on lead quality.

Write your definition before launch: right service, right geography, a budget or timeline, and a reachable phone number or real business email. Score the first ten leads blind, without looking at their source. That keeps you from grading the new channel on a curve because you want the experiment to work.

Fix attribution before the first click. ChatGPT ads append only oppref by default, so without your own UTMs, clicks can blend into organic chatgpt.com referrals or Direct. An October 2026 bug that stripped utm_source pushed 22 to 57% of ChatGPT sessions into Direct; ads tagged utm_source=chatgpt_ads kept attribution see the tracking account. Use your Pixel plus Conversions API with the same event ID if you can. You cannot score a source you cannot identify.

First step: Write the qualification checklist and add utm_source=chatgpt_ads to every ad URL. Cap the test at ten scored leads or $350, whichever comes first. Kill rule: If fewer than two of ten are qualified, stop. Do not argue that more volume will fix poor fit.

3. Build one page that answers the question

Question: Does a question-matched page convert better than your existing page? Your Google page may assume the visitor has already decided what they need. It opens with your brand, your headline and your proof. A chat visitor can arrive mid-thought and need the answer before the pitch. Same offer, different point in the decision.

I learned how much that fit matters watching Quality Score tank after a client swapped a landing page without telling me. The ad did not change. The destination did. For this test, leave your original page alone and build one variant. Mirror the question in the H1, answer it in 60 words, then state price, location and the next step in plain text. Send comparable traffic to both pages; otherwise you are testing who arrived, not what they read.

First step: Draft that single question-matched variant and keep your existing page as the control. Set a $250 spending ceiling and plan to compare lead rate after 50 clicks to each page. Kill rule: If the variant does not beat the existing page on lead rate at that point, delete it rather than rebuilding the whole site around a hunch. If the ceiling arrives before both pages reach 50 clicks, stop and mark the result inconclusive. Do not buy extra clicks just to make the spreadsheet look complete.

4. Test conversational copy against your best Google headline

Question: Does answering the conversation beat the headline that already works on Search? ChatGPT does not match keywords the way Google Search does. Delivery considers the current conversation, your landing page, your ad copy and context hints at ad-group level. Those hints guide delivery; they do not guarantee it see how targeting works. That gives conversational copy a reason to work, not a free pass.

Run two ads to the same page with the same budget split. Ad A uses the exact Google headline that already converts. Ad B answers a question directly and states the offer in its first line. Change nothing else. In one B2B test, keyword-style hints beat conversational phrasing 1.2% CTR to 0.9% CTR see the test details. Phrasing can change delivery as well as click appeal, so do not declare a winner from CTR alone.

First step: Put the proven headline beside one conversational rewrite. Cap the pair at $200 total and 2,000 impressions per side; stop when the spending ceiling arrives even if impressions fall short. Kill rule: If Ad B cannot beat the control on qualified click-through by the cap, keep the control and stop rewriting. A clever line that attracts the wrong click is still the wrong line.

5. Put real customer questions against generic prompts

Question: Do your customers’ words steer delivery toward better prospects? Pull ten exact questions customers already asked in sales emails, support tickets or calls before buying. Use those phrases as context hints in one ad group. In a second, use the generic category wording a stranger might type. Keep the offer, page and budget split the same.

The mechanism is worth understanding before you touch the settings. Matching reads the live conversation first, then your page and hints; the hints steer rather than dictate placement see the targeting explanation. Real questions may bring the ad closer to buying intent, but the test exists because “may” is not a budget decision.

Watch pacing. The US self-serve floor is $25 per day per campaign, and OpenAI can spend up to twice your daily budget in a day and seven times that budget across a week see the minimums and pacing. A daily setting is not a reason to check once a week.

First step: Collect the ten questions, then launch the two otherwise identical groups. Set a $250 total ceiling and check spend often enough to pause before it runs away. Kill rule: If the customer-question group has not produced a qualified lead at a lower cost than the generic group by the ceiling, keep neither running. Fix the offer before buying more traffic.

6. Pause a winner to see what disappears

Question: Did ChatGPT add qualified demand, or take credit for demand you already had? Run a winner for two full weeks, then pause it for one week while holding everything else constant. Count qualified leads and booked calls, not clicks. A dashboard conversion tells you where a buyer clicked; it does not tell you whether that buyer would have found you through Google anyway.

This is an incrementality check, not a magic proof of causation. A holdout region or a clean on-off read is more useful than treating every last-click conversion as new business see the measurement discussion. Keep the offer and other campaigns steady, or the comparison becomes another five-variable test wearing a lab coat.

First step: Write down what you will count, schedule two weeks on and one week off, and set a $350 spending ceiling for the on period. Monitor pacing during those two weeks; the off week costs no ChatGPT ad spend. Kill rule: If qualified volume does not drop when ChatGPT is off, do not scale it. If the ceiling cuts the on period short, record that limitation rather than pretending you ran the full comparison. A winner has to survive being switched off.

7. Put ChatGPT head to head with your Google Search control

Question: Does ChatGPT earn qualified leads more efficiently than the channel you already use? This is the test I would start with if you run Google Search. Use the same offer and landing page, separate budgets and the same two weeks. Do not force the same headlines onto both channels. Hold the offer steady so the comparison is about where the demand came from, not which version of your pitch happened to be live.

One B2B account spent $2,319 across three verticals for 69,630 impressions and 739 clicks on ChatGPT: a 1.1% CTR and $3.14 CPC, with zero MQLs. Google Search produced 2 MQLs in the same account at $6.76 CPC and 5.2% CTR read the independent test. Another account-level comparison puts ChatGPT clicks 54% cheaper but reports a 0.21% conversion rate against Google’s 3.71%. At about $3 per click, the ChatGPT figure works out near $1,429 per conversion, versus a $66 Google median see the cost math. Cheaper traffic is not the same thing as cheaper customers.

Tabletop model of two market stalls side by side, illustrating a controlled ad test

Before launch, give each ad its own UTMs and agree on what counts as a qualified lead. Reporting can lag by about a day, and billing may appear as partial charges: the $500 coaching-firm test showed five separate charges and a 24-hour delay read what happened. Do not double-count charges or panic over a first-day dashboard. Once the data settles, judge cost per qualified lead, not CTR or CPC. Hostinger CPMs above $65 may look alarming, but a cheap click that never buys is the most expensive click you own see the benchmark discussion.

First step: Set up the same-offer comparison and a one-line log: question, cap, result, decision. Cap ChatGPT at $300 or 100 clicks, whichever comes first. Kill rule: If cost per qualified lead is three times your Google control with no path to close the gap, pause ChatGPT and keep Google. If the cap arrives before either channel produces a qualified lead, call it inconclusive, not a win.

If you have run two clean tests but cannot watch pacing daily, hand the work off only on written terms: a shared qualified-lead definition, UTMs on every ad, losers paused, and your approval required before any winner scales. A percentage-of-spend freelancer earns more when you spend more; that is not an incentive to enforce your kill rule. groas runs Search Ads across Google, Bing and ChatGPT for a flat $1,499 a month, month to month, with no cut of spend. A dedicated manager sets goals with you, bids and budgets are worked around the clock, and actions are logged with reasoning. If a shop will not agree to your stop conditions, keep your $300 and run the control yourself.

If I had $500 and had to pick one test, I would start here. I used to start with copy because copy felt productive. I was wrong. Rewriting an ad on an unproven channel is polishing the wrapping before you know whether anyone wants the box. Put the same offer against Google, score the leads blind, and let cost per qualified lead decide what gets another dollar.