Forms for Research-Stage Clicks
By Keshav Parsai 9 min read
In brief
Use lower-friction conversion events for research-stage ChatGPT Ads traffic without starving the oCPC optimiser of signals.
Last verified: 12 September 2026 | Version: 1.0 | Next scheduled review: 12 December 2026
Ask for less than a sales call. Someone who clicked an ad underneath a model's answer is comparing options, not booking. The standard B2B response, a "Request a demo" form with nine fields including phone number and company size, is asking a researcher to behave like a buyer, and it fails twice: once on the visitor, and once on the bidding system.
The default error, stated precisely
The mistake is not that demo forms are bad. It is applying one offer to traffic whose arrival state you know is earlier than that offer assumes.
Your context hint describes a situation. Read your own hints back and ask what the person in them has decided. If the hint says "an IT manager being asked by their insurer to evidence endpoint protection before renewal", that person has a deadline and a requirement and has not chosen a vendor. A demo form asks them to convert a research task into a calendar commitment with a salesperson, three steps ahead of where they are.
The offer ladder
Every rung is a real conversion event. The question is which one your traffic can reach today.
| Rung | What you ask for | What the visitor gets | Fields | Fits |
|---|---|---|---|---|
| 1. No ask | Nothing | The full answer, on the page | 0 | Very early themes, content-led hints, brand building |
| 2. Self-serve proof | Email only | Interactive tool, calculator, scorecard, live sandbox | 1 | Products a person can evaluate alone |
| 3. Documented answer | Email, sometimes company | Pricing detail, implementation scope, security or compliance pack, comparison written straight | 1 to 2 | Committee purchases, regulated categories |
| 4. Free tier or trial | Email, password | Access to the product | 2 | Product-led B2B |
| 5. Scoped consultation | Email, company, one qualifying question | A named 20 minute session with a stated agenda, not "a demo" | 3 to 4 | High consideration services |
| 6. Demo request | Full qualification set | A sales meeting | 5 or more | Late-stage, branded or high-intent themes only |
Two rules for using it. Each ad group gets one primary rung, chosen from its hint theme rather than from company preference. And every page above rung 1 carries a quiet secondary path one rung down, so the visitor who is ready to talk can, and the visitor who is not still has somewhere to go.
Picking the rung: two tests
The decision test. Write the single decision the visitor has already made, based on your hint. If the answer is "that this problem is worth solving", you are at rungs 2 to 3. If it is "that a tool like this is the answer", rungs 3 to 4. If it is "that it is between you and two competitors", rungs 5 to 6.
The stranger test. Would you make this request of someone who has read one page of your site and had one conversation with a model? A phone number fails this test in almost every B2B category. An email address passes it if something real arrives in return.
The signal floor: why this is a bidding problem, not a copy problem
This is the argument that changes the decision, and it has nothing to do with conversion rate.
ChatGPT Ads offers three objectives: Reach on CPM, Clicks on CPC, and Conversions on oCPC. The Conversions objective optimises toward the event you fire through the OpenAI pixel or the Conversions API. Optimisation is learning, and learning needs examples.
Now count. At a reported three to five dollar CPC, a 1,500 USD month buys roughly 300 to 500 clicks. If the only conversion event on your site is a demo request, and demo requests are the rare thing they usually are in B2B, that month produces a single-digit number of conversion events, sometimes zero. An optimiser handed four positive examples across a month is not optimising. It is guessing, expensively, and you are paying oCPC prices for the privilege.
Call that the signal floor: the minimum event frequency below which choosing the Conversions objective is worse than choosing Clicks. Nobody outside OpenAI can state where the floor sits, because OpenAI publishes no learning-phase threshold, no minimum conversion volume and no cross-advertiser benchmarks. What is safe to say is that a rung 5 or 6 event alone will sit below it for most B2B accounts at starting budgets, and a rung 2 or 3 event will not.
The practical response is to instrument both. Fire a mid-ladder event as the optimisation target and keep the demo request as a separate reported event you judge the channel on. Reported standard conversion events include registration, trial start, lead and purchase types, per integration documentation updated September 2026, so the mid-ladder event usually has a standard slot already.
Field count, and who is telling you about it
The circulating numbers are consistent and their provenance is uniform.
Lapis, in a guide updated August 2026, reports three to five fields as optimal, a 7 to 10% conversion reduction per additional field, roughly 5% recovered by removing the phone number alone, and a table running from 10 to 15% conversion at three fields down to 2 to 4% at ten or more. That is a company selling ChatGPT ads landing page services, describing results on that service, and no equivalent figures have been published by a party without a commercial interest in short forms.
So use the direction, which is old, stable and consistent with decades of form research outside this channel, and discard the decimals. The operating rule that survives: no field on a research-stage form that a human will not use within 48 hours. If nobody looks at company size before the first call, it is not a qualifier, it is friction you are paying three to five dollars a click to add.
The reckoning: a softer offer can look like progress and not be
Moving down the ladder reliably raises measured conversion rate. It does not reliably raise pipeline, and on this channel you will find out late.
Conversions take 24 to 48 hours to appear in ChatGPT Ads reporting, per agency documentation from August 2026. Add a B2B sales cycle on top and you can spend six weeks optimising toward an ebook download that converts to revenue at a rate nobody has measured yet. There are no cross-advertiser benchmarks to sanity check against, and no multi-day attribution detail that would let you reconstruct the path afterwards.
The correction is not to avoid soft offers. It is to refuse to call the soft event a result. Pick the mid-ladder event as the optimisation target because the optimiser needs volume, and set your judgement threshold on the downstream event, in your CRM, on a schedule long enough to actually see it. If your CRM cannot tell you which ChatGPT ad group produced a qualified opportunity, fix that before increasing the budget, because nothing in Ads Manager will do it for you.
What we cannot tell you
- Where the signal floor actually is. OpenAI publishes no learning-phase requirement or minimum conversion volume for the Conversions objective.
- Whether ladder position changes lead quality on this channel specifically. Not published by any advertiser, and InPromptAds has no first-party campaign data.
- Whether the field count figures hold here. They come from a vendor selling landing page services, and no independent test on ChatGPT Ads traffic has been published.
- What a typical conversion rate is. No cross-advertiser benchmarks exist. Every number in circulation is one company describing its own accounts.
- Which context hint produced a given lead. There is no hint-level reporting, so ladder decisions can be evaluated at ad group level and no finer.
Quick answers
Should I put a demo request form on ChatGPT Ads traffic? Only on late-stage or branded ad groups. For consideration-stage themes, ask for something a researcher can say yes to today, and keep the demo as a visible secondary path.
How many fields should a ChatGPT Ads lead form have? Three to five on research-stage offers, one or two on a pure information exchange. The working rule is to cut any field a human will not use within 48 hours.
Should I ask for a phone number? Not on research-stage traffic. One vendor reports around 5% conversion recovered by removing it alone, and a researcher reads it as a commitment to being called.
What conversion event should I optimise toward? A mid-ladder event that fires often enough for the oCPC optimiser to learn from, not a rare demo request. Keep the demo request as a separate reported event for judging the channel.
Why does a rare conversion event matter beyond lost leads? The Conversions objective learns from the events you fire. A handful of events a month gives it almost nothing to learn from, so you pay optimised pricing for unoptimised delivery.
How fast will I see conversions in reporting? Agency documentation from August 2026 reports 24 to 48 hours. Plan any judgement on a downstream event around your sales cycle, not around the dashboard.
Sources
| Claim | Source | Tier |
|---|---|---|
| Objectives are Reach (CPM), Clicks (CPC) and Conversions (oCPC) | OpenAI Ads Manager documentation, 2026 | Confirmed, primary |
| Conversion measurement via the OpenAI pixel and Conversions API | OpenAI Ads Manager documentation, 2026 | Confirmed, primary |
| Standard conversion event set including registration, trial start, lead and purchase types | AppsFlyer, ChatGPT Ads integration setup, updated 10 September 2026 | Reported |
| Conversions appear in reporting after 24 to 48 hours | Choice OMG, ChatGPT Ads field guide, August 2026 | Reported |
| Three to five optimal fields, 7 to 10% loss per added field, around 5% recovered by removing phone number | Lapis, ChatGPT ads landing page guide, updated August 2026 | Vendor |
| Three to five dollar CPC bid guidance | Multiple agency write-ups citing OpenAI guidance, 2026 | Reported |
| No learning-phase threshold, minimum conversion volume or cross-advertiser benchmarks | OpenAI Ads Manager documentation, by absence | Absent |
| The offer ladder and the signal floor | InPromptAds method | Inference, ours |
Related reading
- The ChatGPT Click Arrives Mid-Conversation. Most Landing Pages Restart It
- Continuity: Making the Page Answer the Question the Ad Interrupted
- How to Measure ChatGPT Ads, and What You Cannot Measure Yet
- How the ChatGPT Ads Auction Works
- Build Context Hints From Sales Calls, Not From Brainstorms
Changelog
12 September 2026, v1.0. First publication. Introduces the offer ladder and the signal floor, reframing form choice as an optimiser input rather than a conversion rate tactic.
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.
