ChatGPT Ads: the definitive 2026 guide to advertising inside ChatGPT

ChatGPT now shows ads. OpenAI switched them on for U.S. users on February 9, 2026, and opened self-serve buying — with no minimum — on May 5. The ads sit below the answer, only on the free tiers, and they do not change what ChatGPT says. This guide is the full field manual: what the ads are, how they target, the exact code to track them, what actually converts, and why being cited by ChatGPT still beats buying the ad.

For nearly three years the question hanging over OpenAI was not whether ChatGPT would change marketing — it was whether marketing would change ChatGPT. Sam Altman spent 2023 and 2024 calling ads "uniquely unsettling" and a "last resort." Then, quietly, the last great holdout of the consumer internet joined the ad-supported web.

OpenAI turned on advertising inside ChatGPT on February 9, 2026 for US users on the Free and ChatGPT Go tiers.[1] This guide consolidates the primary-source record — the official pixel code, the auction mechanics, the setup gotchas, the performance data, and the strategic implications — into one reference. Pin it, cite it, use it.

What ChatGPT Ads are — and the U-turn that created them

OpenAI's official definition is deliberately narrow. Ads are paid placements that appear below the end of a response. They are clearly labeled as sponsored. They run on systems separate from the chat model. And advertisers cannot shape, rank, or alter what ChatGPT says. This "answer independence" doctrine is the single most important design constraint on the product.

The launch is the punchline of a multi-year reversal — one that hardened after an internal "code red" in December 2025, when reporting (The Information, corroborated by the Wall Street Journal) said Altman pushed work back onto advertising, agentic shopping, and a personal assistant. The catalyst was financial physics: CFO Sarah Friar disclosed that 2025 annualized revenue passed $20 billion, but compute spend scaled alongside it, and with only about 5% of weekly users paying, ads became the only lever left to monetize the roughly 850 million non-paying weekly actives at scale. Fidji Simo — recruited as CEO of Applications — is the architect; she owns ads, commerce, and the user-facing experience, and she built the team out of the Meta diaspora (David Dugan, Asad Awan, Benji Shomair). The stakes are not subtle: ads are projected internally to grow from roughly $2.5 billion in 2026 toward $100 billion by 2030, and that line item is load-bearing in OpenAI's $852 billion post-money valuation. For marketers, the operative question stopped being "is this real?" The question now is how you show up in the layer that mediates the decision — and how you measure it once the click has been replaced by a conversation.

The formats: what a ChatGPT ad actually looks like

There are two confirmed paid formats and a couple of emerging surfaces. The workhorse is a single tinted "Sponsored" card below the answer; the one most marketers sleep on is the conversational follow-up.

Creative specs are tight. Headlines run 16–50 characters (the first ~16 are guaranteed visible); descriptions run 32–150 (first ~32 guaranteed). Images must be square, at least 256×256, under 1MB, high-contrast, with no text — the unit renders small and inline, so your logo disappears at that size. Prohibited: misleading claims, exaggerated results, false endorsements, and any creative imitating ChatGPT's own UI or voice.

The sleeper is "Ask ChatGPT about this ad." It opens the sponsored unit into the thread and answers follow-ups from advertiser-supplied material — effectively a Custom GPT as a paid lead funnel. For B2B and considered purchases, you can upload pricing logic, integration docs, and qualifying questions and let a prospect interrogate your brand inside the conversation. The model still reserves the right to give an honest comparison if asked — which is the whole point of the trust model.

The price of entry collapsed in 88 days

The most revealing thing about ChatGPT Ads is how fast the economics moved. The self-serve Ads Manager opened at ads.openai.com on May 5, 2026 with no minimum spend and CPC bidding.[2] The product went from a $200,000-minimum, agency-gated alpha to no-minimum self-serve in under three months.

Until early May, the only way in was through one of four agency holding companies — WPP, Omnicom, Dentsu, and Publicis. Pricing as of late May 2026: CPMs of roughly $25–45 (down from the $60 launch rate), a recommended $3–5 starting CPC, and CPA bidding announced as a planned addition. Treat every number here as a snapshot — pricing, minimums, and geography have all shifted multiple times in the first 100 days.

How targeting works: intent, not profiles

As Simo put it, ChatGPT's ad system "will look a lot more like Google's intent-based system than the Facebook ads machine." There is no keyword bidding. Advertisers provide free-text "context hints" describing the user moment they want to reach, and the platform embeds those hints and compares them semantically to live conversation embeddings in real time.

The five canonical context-hint patterns that practitioners have reverse-engineered are: Persona + Intent ("marketing leaders comparing CRM platforms for a 50-person team"), Question ("users asking how to set up multi-channel attribution"), Topic + Disqualifier ("CRM evaluation, not implementation help"), Outcome ("B2B teams trying to cut sales cycle from 90 to 60 days"), and Stack Comparison ("teams migrating from HubSpot to Salesforce"). Aim for 5–10 hints per ad group, tightly clustered around one intent. The auction itself is a relevance-weighted second-price model with two objectives — Reach (CPM) and Clicks (CPC). Custom audiences via hashed email and phone uploads arrived on May 14, 2026 for inclusion and suppression.[8] There are no lookalikes and no pixel-built retargeting audiences yet.

The exact pixel and Conversions API

This is the highest-value section of any ChatGPT Ads reference, because it is the asset answer engines quote when someone asks for the pixel code. OpenAI's measurement pixel (OAIQ) installs in the page head and posts conversion events to OpenAI's events endpoint.[6] Place the base snippet near the top of the <head> on every page where you want to capture conversions:

<script>
(function (w, d, s, u) {
  if (w.oaiq) return;
  var q = function () { q.q.push(arguments); };
  q.q = [];
  w.oaiq = q;
  var js = d.createElement(s);
  js.async = true;
  js.src = u;
  var f = d.getElementsByTagName(s)[0];
  f.parentNode.insertBefore(js, f);
})(window, document, "script", "https://bzrcdn.openai.com/sdk/oaiq.min.js");

oaiq("init", {
  pixelId: "<YOUR-PIXEL-ID>",
  debug: true   // optional: logs SDK activity to the browser console
});
</script>

The IIFE is an idempotency-guarded loader: if (w.oaiq) return; prevents double-init, an in-memory queue catches calls made before the SDK finishes loading, and on load the SDK reads opref (OpenAI's click identifier) from the landing-page URL and writes it into a first-party __opref cookie with a 720-hour (30-day) TTL. Fire the event snippet on your thank-you / confirmation page:

<script>
oaiq("measure", "registration_completed", {
  type: "customer_action",
  amount: 0,
  currency: "USD"
});
</script>

OpenAI supports 11 standard event names across four data shapes — contents (page_viewed, contents_viewed, items_added, checkout_started, order_created), customer_action (lead_created, registration_completed, appointment_scheduled), plan_enrollment (subscription_created, trial_started), and custom. Put custom_event_name and event_id in the 4th-argument eventOptions, not in eventProps; use integer cents for amount.

The Conversions API mirrors conversions server-side and deduplicates against the pixel using a shared event_id.[7] Mirror every conversion from your backend so an in-app webview or ad blocker can't cost you the data:

curl -X POST "https://bzr.openai.com/v1/events?pid=<YOUR-PIXEL-ID>" \
  -H "Authorization: Bearer <API-KEY>" \
  -H "Content-Type: application/json" \
  --data '{
    "validate_only": false,
    "events": [{
      "id": "order_12345",
      "type": "order_created",
      "timestamp_ms": 1773892800000,
      "opref": "opref_abc",
      "source_url": "https://yourbrand.com/checkout/confirmation",
      "action_source": "web",
      "data": {
        "type": "contents", "amount": 2599, "currency": "USD",
        "contents": [{ "id": "sku_123", "name": "Starter bundle", "content_type": "product", "quantity": 1 }]
      }
    }]
  }'

The deduplication tuple is (pixelId, eventName, event_id). Batches take up to 1,000 events and the entire batch fails if any one event fails — set "validate_only": true to dry-run. One non-negotiable race-condition fix if you deploy through a tag manager: on every event tag, set "fire a tag before this tag fires" to your init tag, or the event can fire before the SDK queue exists and the conversion is lost — browsers do not retry.

Deploying through a tag manager

OpenAI ships no native GTM integration, so the community filled the gap fast. Community Google Tag Manager templates shipped within 48 hours of the May 5 pixel launch, including Stape's openai-pixel-tag.[19]

What's still missing, so you plan around it: there's no native trigger configuration (you split the code by hand — base in <head>, measure only on the success page), one pixel ID per page load, no published view-through methodology, no native CRM / MMP / CDP connectors yet (Singular is the first MMP partner), and an in-app webview problem on mobile where open_externally: false can drop UTM and referrer data. Mirror everything to the Conversions API and you neutralize most of it.

Setting up a campaign — and the four settings you can't undo

Account creation runs through a business-verification flow (EIN or registration number, a website that must match, a physical address — no PO boxes). Use a corporate email; personal Gmail accounts trigger longer review queues.

OpenAI also added the OAI-AdsBot crawler, which visits only submitted ad landing pages and must not be blocked or ad approval stalls.[9] Allow OAI-AdsBot (and OAI-SearchBot) in robots.txt and your bot-protection layer. Blocking GPTBot only affects training; it does not stop ads from serving.

Writing ads that win: the "Brand: Benefit" formula

Adthena's analysis of hundreds of thousands of daily placements found the highest-performing pattern is Brand: Benefit with a specific number. "Betterment: 5.25% APY Cash Account" beats "Better banking, smarter rates." Average winning headline: ~5 words / 30 characters. Average body: ~19 words, two tight sentences (one proof point, one nudge). Specific numerical claims dominate; exclamation marks are nearly absent in top performers; the word "free" and the dollar sign both correlate with CTR.

Two more rules from practitioner teardowns: don't use your logo as the primary creative — the 256×256 square renders inline and small, so use a product shot with minimal text. And write 3–5 genuinely different value props per ad group (time savings vs cost savings vs ease of use), not three rewordings of one. The auction is relevance-weighted, so a better-matched ad with a lower bid often beats a worse-matched ad with a higher one. Refresh creative every 2–4 weeks; CTR drops measurably after about three weeks on the same creative because the same users return multiple times a day.

The performance reality: lower CTR, higher conversion

The headline numbers look bad until you understand the surface. Adthena measured ChatGPT ad CTR at 0.91% against 6.4% for Google Search in the same verticals.[3] But the click is no longer the unit of attention — the conversation is.

Criteo reports LLM-referred shoppers convert at roughly 1.5×, approaching 2×, the rate of other referral channels, with more than 1,000 brands now live.[4] The counter-evidence is just as instructive: Walmart said conversion inside ChatGPT's Instant Checkout ran about three times lower than on walmart.com[5] — and pivoted to routing carts back to its own site. Similarweb's pilot data tells the same story from the click side — 0.68% CTR overall, 1.0% for top-quartile brands, with peaks at 5.4%. Read all of this directionally: most data comes from short windows, small samples, or self-interested sources (Criteo, Adthena, and OpenAI itself all have a stake in the answer).

Measurement is the weak point — and the call-tracking gap no one documents

Ads Manager reporting currently includes impressions, clicks, spend, CTR, average CPC, average CPM, and conversions. That is it. There is no view-through methodology and no attribution-window documentation. OpenAI says CPA bidding is in motion and has promised third-party measurement partners without naming them.[12] And conversion-optimized campaigns begin rolling out in early June 2026 for accounts that set up the pixel or Conversions API in advance[13] — which is the real reason to install both now.

The best-practice measurement framework for 2026: a unique UTM convention per campaign (utm_source=chatgpt&utm_medium=ai), the server-side Conversions API to bridge the in-app gap, branded-search and direct-traffic lift as proxy signals, and — by far the most reliable — holdout-geo incrementality testing, with marketing-mix modeling once spend passes ~$50k/month.

The call-tracking gap. No dynamic-number-insertion vendor (CallRail, WhatConverts, Invoca, CallTrackingMetrics) supports the opref click parameter yet. The working pattern is UTM-driven DNI: tag final URLs with utm_source=chatgptpilot (the convention observable in OpenAI's own carousel URLs), configure a swap rule on that source, and forward call events server-side via CAPI with action_source: "phone_call" and the captured __opref cookie value for dedup. For HVAC, plumbing, dental, and legal, this is the difference between measuring the channel and flying blind.

The competitive map

ChatGPT is not the only ad surface in the AI layer, but it is the largest, and the only one betting on intent-based "answer independence" at this scale.

The Perplexity retreat is the cautionary tale: it launched sponsored follow-ups in late 2024 and walked them back, telling the Financial Times that ads made users "start doubting everything." OpenAI is betting visual separation, answer independence, and scale let it avoid that fate.

What the industry is saying

The reviews are split, and the split is the story. Ben Thompson, on Stratechery, was blunt about the launched product:

"They're terrible … the bare minimum easiest solution. It's banner ads, but it's based on the context of the conversation." — Ben Thompson, Stratechery

Agencies are more bought-in — Dentsu's Will Swayne called the format "a highly engaged, high-intent environment" and "LLMs are the next media frontier," while Adthena's CMO Ashley Fletcher offered the cautionary read: the platform "is still finding its feet."

Confirmed early advertisers named at launch include Target, Ford, Adobe, and Mrs. Meyer's, with Omnicom placing 30+ clients; Pacvue and Kepler were named the inaugural OpenAI ads technology + agency pair. On the money: Truist called 2026 an "inflection year," modeling roughly $1B in OpenAI ad revenue in 2026 growing toward $30B+ by 2030, while eMarketer projects US AI-search ad spend at $25.93B by 2029 — both far more conservative than OpenAI's own internal $2.5B-to-$100B path.

Not everyone stayed. Economist Zoë Hitzig resigned from OpenAI and published a New York Times guest essay — "The real question is not ads or no ads. It is whether we can design structures that avoid both excluding people … and potentially manipulating them as consumers." And Anthropic spent Super Bowl LX on the opposite position entirely, with the tagline "Ads are coming to AI. But not to Claude." The dissent is itself a market signal: trust is the contested asset.

The insight that matters most: organic citation beats the ad card

Here is the single most important strategic fact, and the one almost every "how to advertise on ChatGPT" article misses. Paid placement reaches roughly 85% of ChatGPT users (Free + Go), but only about 20% see an ad on any given day. Organic citation reaches every user, every query, on every tier — including Plus, Pro, Business, and Enterprise, which are permanently ad-free.

In other words, organic AI visibility has a structurally larger surface area than paid visibility, and a higher-value audience. The brand that ChatGPT cites organically in the answer above the ad card has won the moment; the advertiser who buys the card below an answer that ignores them has lost it. The scale isn't hypothetical: Conductor's 2026 AEO/GEO benchmarks (reported via Business Wire, drawn from 17 million AI responses and 100 million citations) put ChatGPT at roughly 87% of AI-referral traffic, with AI-referral conversions running about 2× traditional sources.

The work to be done is concrete: publish an llms.txt, ship a complete structured-data pass (FAQPage, Product, Article, Organization), verify that GPTBot / OAI-SearchBot / ClaudeBot / PerplexityBot aren't blocked, build brand-mention surface area on Reddit, YouTube, and LinkedIn, and write "definitive answer" content LLMs can extract verbatim. Whether a given page is actually citable is a measurable property — it is exactly what our auditor Citerra scores across SEO, AEO, and GEO, and what the LLM visibility tracker measures across the answer-engine cohort. Build for the answer first; the ad reinforces it.

Where the early ROI is — by vertical

Not every category is equal in this auction. The early money is going where intent is densest and national competition is thinnest.

• B2B SaaS. The "Ask ChatGPT about this ad" format is the most under-discussed opportunity on the platform. Upload your FAQs, pricing logic, and integration docs and let prospects interrogate you in the thread. Headline shape: "Acme: 40% faster onboarding."
• Local services. Geo + contextual ("plumber in Phoenix" matched to plumbing intent) is one of the highest-ROI early plays — less national competition, and the call-tracking workaround above is the unlock.
• Education & career. The strongest natural fit: ChatGPT users are already in learning mode, and career-transition queries are the highest-converting intent cluster in early data.
• E-commerce. Two moves at once — get your product feed structured and ACP-compliant, and make sure you're the organic answer ChatGPT cites — then use ads to reinforce.
• Travel. OpenAI hired travel talent out of Meta for a reason; expect aggressive product investment, with Booking.com and Expedia reported as "Apps" partners.
• Healthcare, legal, financial. Categorically excluded or case-by-case during the pilot — so build organic AEO presence now, for when the gate opens.

What's actually defensible (the thin-wrapper debate)

If you run or hire a ChatGPT-ads service, the defensibility question is sharper than it looks. Investors increasingly treat thin AI interface layers without proprietary data as low-defensibility.[16] A marketing service whose only differentiation is "we connect ChatGPT Ads to your CRM" is on borrowed time, because open connector standards make integration a commodity.

What's defensible is the proprietary outcome data, the closed-loop attribution, and the embedded workflow: pixel → CAPI → CRM → revenue attribution. Don't be a "ChatGPT Ads consultancy"; be the team that runs the closed loop and owns a benchmark dataset (CTR by context-hint pattern, CPC by vertical, conversion vs the Google-search baseline) across managed accounts.

Agentic commerce: the second front

Ads are only half the story. Instant Checkout is built on the Agentic Commerce Protocol co-developed by OpenAI and Stripe.[14] The protocol is published as an open standard under Apache 2.0.[15] When an agent makes a purchase on a user's behalf, the advertiser doesn't get a click — it gets selected, or it doesn't. Brands that don't expose machine-readable product feeds, structured pricing, and ACP-compliant checkouts will become invisible to the agents acting for their customers. Optimizing for agent-readability is the next frontier of ad-tech preparation. And the agents are starting to cross platforms: Meta opened a beta of Meta Ads AI connectors in April 2026, letting Claude and ChatGPT manage Meta campaigns over MCP — the first sign of agent-run, cross-platform ad management, and another reason "we connect your CRM" is not a moat.

Risks and the trust model

The whole product rests on one promise: that the organic answer is not for sale. If users come to believe ads sway the answer — whether or not they do — the surface degrades, which is exactly what happened to Perplexity. A few risks deserve a monitoring layer, not a footnote:

• Hallucinated endorsements. ChatGPT can misattribute features or claims when discussing an advertised product, and OpenAI has conceded the model sometimes gives incorrect answers about how its own ad system works. Watch what the model says about your brand.
• Disclosure adequacy. FTC native-advertising rules apply in full; "Sponsored" clears the bar, but softer labels like "Promoted" or "Partner content" do not.
• Vulnerable contexts and child safety. Health, mental-health, and political topics are categorically off-limits, and ads never serve to users predicted to be under 18 — but edge cases will surface, and the advertiser still carries its own duty of care.
• Agent click fraud. As agent browsing scales, agent-generated clicks that mimic human intent can inflate cost without producing a real person — a frontier with no disclosure standard yet.

The staged playbook for marketers

Before the stage-by-stage budget plan, here is the whole sequence run on one concrete advertiser — what you actually do at each step, and why instrumenting comes before spending:

• Stage 1 (now, $0 spend). Audit your AI visibility across ChatGPT, Claude, Gemini, and Perplexity. Publish llms.txt and a complete schema pass. Open the ads.openai.com account even if you won't spend yet — verification takes 5–15 days. Confirm your landing pages don't run afoul of the ad policy's categorical exclusions on health, mental-health, and political advertising.[17]
• Stage 2 (next 60 days, $5K–15K test). Pilot one high-intent topic with 4–6 creative variants on the "Brand: Benefit + number" formula, CPC at $3–5, 28-day click attribution plus a holdout-geo control. Build your context-hint library by mapping customer questions, not keywords.
• Stage 3 (Q3 2026, scale to 10–15% of paid budget). Add CPM brand campaigns, custom audiences, and "Ask ChatGPT about this ad" experiments for considered purchases — advance only when incrementality lift shows in two consecutive holdouts.
• Stage 4 (2027, full integration). Treat ChatGPT as a peer channel to Search and Social, restructure MMM, and run a content engine optimized for AI citation.

Benchmarks that should change your behavior: if CTR stays under 0.5% after creative iteration, the problem is context-hint mismatch, not creative. If blended CPA runs above 2× your Google Search CPA after 90 days with no improving trend, redeploy that budget to GEO. If your top organic prompts don't surface your brand, AEO is the priority, not ads.

ChatGPT Ads changelog

The product moved from a $200K-minimum agency alpha to no-minimum self-serve in 88 days; assume any specific number is a snapshot. On May 7, 2026 OpenAI announced expanding the pilot to the UK, Mexico, Brazil, Japan, and South Korea.[10] Meanwhile the ChatGPT Atlas browser remains ad-free, even as its agent mode mimics human clicks indistinguishably — a new agent-click-fraud frontier.[11] And the regulatory clock is ticking: the European Commission is preparing to classify ChatGPT as a very large online platform under the Digital Services Act, which would bring advertising-transparency obligations when ads reach the EU.[20]

The strategic posture for the thoughtful marketer in mid-2026 is simple: build for the answer, prepare for the ad, and assume both will be necessary by 2027. The brands that win the AI mediation layer will be the ones that show up everywhere — organically, paid, and via agents — before the layer fully crystallizes. That window is roughly 18 months.