ChatGPT Ads for Lead Generation — The 6-Month Reality Check

six months in, chatgpt ads have crossed a $1 billion annualized revenue run rate, opened self-serve access in 52 countries, and launched sponsored agents. openai still hasn’t published a single performance benchmark. cpcs range from $3 to $22 depending on who you ask. one study found ads in 25% of commercial prompts. another tracker swung from 0.05% to 51% ad load in the same month.

nobody knows what “good” looks like yet. that’s the problem.

i manage $2.4m a year in paid acquisition across google and meta for lead-gen businesses — 28 active clients, 3.8x average roas, five-plus years doing this. my job is cost per qualified lead, not cost per click. so when a new channel shows up, i don’t ask “is it revolutionary.” i ask “what does a lead actually cost here, and how would i know.”

the short version: chatgpt ads can generate qualified leads, but month one will cost 3x+ your google search benchmark and the channel will not tell you the truth about itself. you have to instrument it before you spend a dollar. this is the playbook — the 90-day curve from real campaign data, the setup that matters, and the exact kill criteria.

in this guide:

• what chatgpt ads actually are (and aren’t) for lead gen
• the 6-month landscape: cpcs, ctr, and the measurement gap
• the 90-day curve: what to expect month by month
• the step-by-step setup: pixel, utms, context hints, crm
• why last-click measurement will lie to you
• when to test, when to wait, and what budget it takes
• the media buyer’s verdict

what chatgpt ads actually are (and aren’t) for lead gen

chatgpt ads are in-conversation placements. a user asks something, chatgpt answers, and a sponsored result appears inside or adjacent to the response. ads show on the free and go tiers — that’s where the inventory lives. there is no separate ad network, no display placement, no in-app banner network. one surface, one format family.

here’s what doesn’t exist, six months in: no retargeting. no abm. no lookalike audiences. no frequency controls worth the name. your targeting is context hints — short descriptions of who you want to reach — and openai’s model deciding when your offer is relevant to the conversation. that’s it.

if your media plan depends on remarketing, this channel isn’t ready for you. a huge share of lead-gen performance on meta and google comes from retargeting people who visited but didn’t convert. that entire layer is absent here.

so why bother? because of what i call problem-stage intent.

on google search, people type keywords they already know: “google ads agency,” “crm for hvac,” “media buyer for hire.” the keyword is the product of research they’ve already done. in chatgpt, people describe symptoms before they know the keywords: “why did my cost per lead double last month,” “is $80 a good price per lead for a roofing company,” “how do i stop wasting money on facebook ads.”

that’s a lead-gen goldmine if you sell the thing that solves the symptom. the person asking “is $80 a good price per lead for roofing” doesn’t need a keyword — they need a media buyer. the question is the intent, expressed earlier in the journey than search ever captures.

the 6-month landscape: what we know now

let’s separate what’s actually measured from what’s being repeated on linkedin.

scale is real. $1 billion annualized run rate, self-serve in 52 countries, and sponsored agents launched september 16 — post-click conversational agents that can qualify leads inside chatgpt instead of sending them to a landing page. advertisers are spending real money. this is not a beta science project anymore.

cpcs are all over the map. openai suggests $3-5 as a starting bid. actual cpcs run $2-22 depending on market, vertical, and who’s reporting. b2b saas and professional services sit at the high end; consumer-ish categories at the low end. plan your math at the high end of your range, not the low end, or month one will shock you.

ctr is 0.91% on average, and it’s structural. this is the number most advertisers misread. on google search, a 0.91% ctr means your ad copy is broken. here, it’s the format: chatgpt just answered the user’s question. clicking an ad is an interruption, not a continuation. you cannot copywrite your way out of this. don’t burn two weeks on ad text iterations expecting ctr to double — spend those two weeks on landing pages and measurement.

the measurement gap is the real story. cleverly ran a test and reported the numbers everyone should memorize: openai’s dashboard claimed 57 clicks. google analytics recorded fewer than 20 visits. zero conversions. they paused the test at month one and wrote it off.

why the gap? click definitions differ between platforms, attribution windows don’t line up, and some browsers and privacy setups strip referrer data before analytics ever sees the visit. every ad platform overcounts to some degree — but a 3x gap between reported clicks and recorded visits means you cannot run this channel on openai’s dashboard alone. your own tracking is not optional.

inventory is volatile. one study found ads appearing in 25% of commercial prompts. another tracker measured ad load swinging from 0.05% to 51% within a single month. openai is still tuning when and where ads surface. your volume will lurch. that’s the channel being six months old, not your campaign being broken.

the 90-day curve: what to expect month by month

the best public data point so far comes from intelegencia’s six-month case study on a financial advisor matching platform. it’s the closest thing we have to a real cpl curve for lead gen, and every media buyer testing this channel should know it cold.

month one at 3.2x your search cpl is normal. let that sink in, because it’s the number that kills most tests. if your google search cpl is $60, chatgpt ads will open at roughly $190 per lead. your ceo will see that number in week two and want to pull the plug. agree on the curve before you spend, not after.

months two and three are the hard part. improvement is real but slow, and there’s no lever you can pull to speed it up. cleverly paused at month one — before the curve had any chance to bend. most advertisers will quit in this window. that’s not a criticism of them; it’s a budget reality. if you can’t fund three ugly months, you can’t afford this channel at all.

but “don’t touch it” doesn’t mean “do nothing.” the waiting zone has work, it’s just not bid management:

• rewrite context hints that aren’t matching. the delivery data tells you which ad groups got impressions and which starved. starved ad groups have a hint-language problem — rewrite them around the symptom phrasing you see in your search term reports, not your positioning doc.
• fix the landing page for the leads you did get. even at 3x cpl, you’re buying the cheapest user research available: what did people who clicked actually do? if form starts are high and completions are low, the form is the problem, not the channel.
• interview the leads. call every chatgpt-attributed lead in month two and ask what they typed. their answers are your next round of context hints, written in the exact language the model needs.

none of that speeds the model’s learning. it makes sure that when the curve bends — and in the intelegencia data it bent at month four — your funnel is ready to convert what arrives.

month four is the crossover, at least in the intelegencia data: cpl approaches the search benchmark. by month six, cost per lead ran 35% below the search benchmark. that’s the prize — not a viral channel, just cheaper qualified leads than search, from inventory your competitors mostly aren’t buying yet.

the decision framework: pull the plug if cpl isn’t falling meaningfully by day 90. “meaningfully” needs a number, so use mine: trending under 2x your search benchmark by day 60, and under 1.5x by day 90. if you’re still at 3x on day 90 with clean tracking and decent context hints, the channel doesn’t have your audience yet. kill it, keep the notes, retest in six months. this is a kill decision, not a failure — the channel is six months old and your category may simply not have volume here yet.

how to set up chatgpt ads for lead generation (step by step)

the setup is short. doing it in order is what most advertisers skip.

step 1: install the pixel and conversions api before launch

non-negotiable. remember the 57 reported clicks versus fewer than 20 analytics visits. the openai pixel plus a conversions api connection is what lets your own data — not openai’s dashboard — be the source of truth. fire conversion events server-side where possible; browser-side tracking alone will undercount in exactly the privacy setups that cause the gap.

if you skip this step, do not spend money. you’ll be in cleverly’s position: a month of spend, a dashboard full of numbers you can’t verify, and no choice but to pause.

step 2: set up utm discipline on day one

every chatgpt ads url carries utm_source=chatgpt&utm_medium=cpc plus a consistent campaign name. one naming convention, written down, enforced on every ad group. six months from now, when you’re trying to figure out whether chatgpt influenced a deal that closed via branded search, clean utms are the difference between an answer and a shrug.

this is boring and it takes twenty minutes and almost nobody does it properly. be the exception.

step 3: write context hints like a media buyer, not a marketer

context hints are 280 characters that tell openai who your ad is for. the rules that matter:

• one theme per ad group. don’t write “business owners and marketers and founders” — write one problem per hint. the model matches conversations, and conversations are specific.
• mirror the language of the prompt, not your positioning. users type “why is my cost per lead going up.” they don’t type “unlock scalable acquisition.” your hint should contain the words real people use when describing the symptom.
• write hints for the question, not the answer. “people asking how to lower their google ads cost per lead” beats “companies seeking performance marketing solutions.”

a good hint for my own business: founders and marketing leads asking why their google ads cost per lead is rising, whether to hire an agency, or how to fix underperforming campaigns. that’s 170 characters, one theme, symptom language. then the ad copy mirrors it — same words, same problem, same promise.

step 4: make the landing page continue the conversation

the user was mid-conversation with chatgpt about their problem. your landing page opens with that problem — the exact one from your context hint — not with your brand promise. no bait-and-switch: if the hint said “cost per lead going up,” the headline addresses cost per lead going up.

the page has one job: answer the question well enough that the next step is obvious. for lead gen, that next step is usually a short form, a quiz-style qualifier, or a calendar link. keep it vanilla. this is not the channel for clever interactive experiences — the user was reading a text answer thirty seconds ago.

step 5: pipe chatgpt leads into your crm, separately

tag every chatgpt-attributed lead in your crm as its own source — hubspot, pipedrive, whatever you run — and measure cost per opportunity, not cost per click. the whole value of this channel shows up downstream: a chatgpt lead that closes at 2x the rate of a meta lead is worth paying 2x the cpl for. you will never see that in the ad dashboard. if you use hubspot, the source tagging takes an afternoon; if you’re spreadsheet-native, a mandatory “source” column works until it doesn’t.

measurement: why last-click will lie to you

here’s the pattern that will burn you. a founder asks chatgpt about fixing their lead flow. your ad appears. they click, read your page, don’t convert — it’s early, they’re researching. three days later they google your brand name and convert through branded search.

last-click attribution assigns that conversion to google branded search. chatgpt gets nothing. run that loop for a quarter and chatgpt ads look like a money pit while your branded search campaigns take the credit. this is exactly how new channels get killed inside companies that measure last-click only.

without multi-touch attribution, chatgpt will look worse than it is — structurally, not marginally. the fix isn’t a better dashboard; it’s measuring what actually moved.

• branded search lift: watch your branded search volume and conversion rate before, during, and after chatgpt spend. if branded search conversions rise while chatgpt runs, that’s the channel working through the back door.
• assisted conversions: in ga4, look at conversions where chatgpt appears anywhere in the path, not just last. with clean utms from step 2, this is a saved report, not a project.
• pipeline revenue: the only number that survives a cfo review. cost per opportunity from chatgpt-tagged leads in the crm, compared against search and meta cohorts over 90 days, not 14.

if you want a harder answer, run an incrementality test: a 10-15% holdout — geographic or user-based — for two to three weeks, with chatgpt ads off for the holdout group. compare lead volume and pipeline in exposed versus holdout. it’s the same methodology you’d use for any channel that leaks credit, and it’s the only way to know what chatgpt ads caused versus what they coincided with. attribution tooling like prooflytics exists specifically for this problem if you’d rather buy than build.

and if you’re rebuilding measurement anyway, do it once: the same first-party data infrastructure that fixes chatgpt attribution also fixes the signal loss google’s side is creating. i cover that setup in the first-party data guide for lead gen.

when to test, when to wait

test now if:

• you’re b2b saas or professional services — the cpcs are high but the deal sizes carry them, and problem-stage questions in your category are common
• your purchase is high-consideration — long cycles mean the early-journey catch actually matters
• you have a 90-day runway — budget and patience, agreed in writing, before the first dollar
• crm attribution is in place — you can tag a lead’s source and follow it to pipeline

wait if:

• your paid search account isn’t fully optimized yet — fix the channel you can measure before opening one you can’t. (if google is forcing you through the ai max migration right now anyway, start with the ai max guide — that’s the bigger near-term swing for your search numbers.)
• you measure last-click only — the channel will read as a failure no matter what it does
• you can’t commit to a full quarter — three ugly months is the entry fee; a six-week test just buys you the expensive part of the curve and none of the payoff

budget: $5,000-$15,000 for a 2-4 week diagnostic phase, not $500 experiments. the math is merciless. at a $22 cpc, $500 buys about 23 clicks. that’s zero statistical signal — you can’t even establish a baseline cpl, let alone a curve. i’d rather see one client spend $10k properly instrumented than ten clients spend $500 each and all conclude “chatgpt ads don’t work.”

and if you’re currently splitting budget across google’s automated formats, the priority question is real: performance max and chatgpt ads compete for the same “let the model decide” dollars. i walked through that trade-off in the performance max for lead gen guide.

the media buyer’s verdict

chatgpt ads are not a replacement for google or meta. they’re a supplement — early-journey inventory that search can’t buy, at cpls that start terrible and, in the best public data we have, end 35% below your search benchmark by month six.

the early-mover advantage is real. competition is thin, cpcs in most lead-gen categories haven’t been bid up, and the learning you accumulate now — which hints work, which landing pages convert, what the curve looks like in your vertical — compounds while everyone else waits for a case study to tell them it’s safe.

sponsored agents, launched september 16, are the part i’m watching closest. instead of sending a click to your landing page, the ad opens a conversational agent inside chatgpt that can answer questions, handle objections, and qualify the lead before anyone fills out a form. for lead gen, that’s a structural change: it moves qualification from your sales team’s calendar to the moment of interest. a roofing company’s sponsored agent could ask roof age, location, and timeline, then book the estimate — all before the lead ever leaves the conversation. it’s early, the tooling is rough, and nobody has six months of data on it. but it’s the first ad format in years that changes where qualification happens, not just where the click lands. i wrote about the broader shift in ai agents for ads.

but the measurement gaps are just as real. a platform reporting 3x the clicks your analytics sees, no benchmarks, volatile inventory, and attribution that leaks into branded search. if you can’t instrument properly, you’re not early — you’re blind.

my recommendation, exactly as i’d run it for a client: small budget, one hypothesis, and a 90-day review date agreed in writing before the first dollar is spent. one icp, one problem theme, one context hint set. $5k-15k diagnostic. kill criteria set on day one — under 2x search cpl by day 60, under 1.5x by day 90, or it’s off. no mid-test budget surges, no emotional pauses at the month-one cpl, no “let’s just see” past day 90.

want a second opinion on whether chatgpt ads fit your media mix? book a strategy call — i’ll look at your current accounts and tell you straight what a 90-day test would look like, what it would cost, and whether i’d run it at all.

frequently asked questions