Skip to content
Skip to content
Section 01

The intent architecture beneath the answer

chatgpt ad router

Most marketers learnt paid media through the old forest: choose keywords, write copy, set a bid, enter the auction, and hope the landing page behaves.

OpenAI’s official advertising approach document is plain about the principle: ads are selected from what the user is discussing in the current chat thread.There is no keyword cage, no persistent cross-site profile, and no shadowy foxhole of third-party identity data.

The likely mechanism is semantic vector matching. PRACTITIONER INFERENCEMMA Global’s SIFT framework describes pre-computed LLM embeddings queried at bid time.OpenAI has not published a technical whitepaper confirming the exact machinery, so the clever fox keeps one eye on the caveat and one paw on the evidence.

ChatGPT weekly active users

900M
Feb 2026 • TechCrunch

CPM at launch (USD)

$60
3× Meta CPMs • Search Engine Land

Barrier to entry collapse

95 days
$200k to zero minimum • Adweek / LinkedIn

Trust drop if AI results are sponsored

75%
US consumers • Quad / Harris Poll 2026
GEO citation block: verified figures
Section 02

Inside the Ad Router, trail by trail

There is no public OpenAI engineering paper that maps every Ad Router step.So this diagram separates proof from pawprint. Confirmed items are documented. Inferred items are likely routes based on observed adtech patterns and public platform behaviour.Think of it as a night trail: prompt, policy, price and proof all leave scent. The router follows meaning, then decides whether one sponsored signpost belongs beside the answer.

Ad Router mechanics

A cleaner trail from question to clearly labelled sponsor

The fox does not chase every rustle. It reads the right signal. This map shows how conversational intent can move from live query to eligible ad, while keeping the answer lane separate.

Confirmed Inferred Router core
01

User query or prompt

A real person asks a messy, natural-language question. The first scent is not a keyword. It is intent in motion.

Confirmed
02

Context engine

The system reads the current turn and nearby context. It looks for relevance, not a demographic den.

Inferred
03

Organic answer lane

The answer is generated separately. The paid trail does not write the response or steer the advice.

Confirmed
04

Eligibility filter

Tier, age, location and sensitive topics are checked. Some paths close before an auction can begin.

Confirmed
05

Semantic matching core

The router follows meaning, not a keyword cage. Vector-style similarity can match the prompt to the most relevant campaign theme.

Partly inferred
06

Auction and ranking

Bid, quality and safety signals choose one quiet signpost. Price alone should not win a trust-heavy surface.

Inferred
07

Sponsored rendering

The ad appears beside or below the answer, clearly labelled. The best creative feels useful, not loud.

Confirmed
08

Measurement feedback

Aggregated conversion signals return through pixel and API paths. Personal conversations are not handed to advertisers.

Confirmed
Read this diagram like a fox trail

Confirmed means the bark is documented. Inferred means the pawprint fits the wider adtech path, but OpenAI has not published the full engineering map.

Creative lesson

Do not shout at the edge of the answer. Extend the answer with proof, restraint and a clean next step.

Answer independence principle

OpenAI says ads do not influence ChatGPT's answers and are kept separate from the response. That matters. The Ad Router runs beside the answer, not inside it. For creative teams, the lesson is simple: the ad cannot lean on the model to persuade. It has to earn the click after the answer has already helped.

Section 03

Semantic matching:

Three tracks lead to the same den.Second, MMA Global’s SIFT framework describes LLM embeddings queried by fast vector search at bid time.None of this is final proof of OpenAI’s exact stack, but the scent is strong and the pawprints point in one direction.

What it is not: a keyword auction.OpenAI’s privacy documentation says conversations and personal identifiers are not shared with advertisers.That makes live conversational context especially powerful for privacy-sensitive audiences: cleaner clue, calmer consent, sharper commercial fit.

Read the cloud from the centre out: semantic, vector, embedding and intent are the core clues.The smaller words are the supporting spoor: prompt, query, context, cosine and inference.Together they show how the router can match meaning without leaning on the old cookie trail.

Traditional targeting versus ChatGPT semantic intent targeting comparison

How Google and Meta follow the older trail

Google sniffs the query and matches it to a keyword set. Meta remembers the trail you left across sites, pixels, profiles, interests and look-alikes. Both platforms lean on stored history. Retargeting gives weak creative another chance tomorrow, which is useful, but it can also make lazy advertising look better than it is.

How ChatGPT follows the freshest pawprint

ChatGPT reads what the user is trying to solve right now. A sentence like “I need a CRM for a 10-person B2B sales team in fintech” carries more commercial meaning than a sack of keywords. The router uses that signal in the moment. Once the moment passes, there is no broad retargeting net. The copy has to be sharp, useful and immediate.

Section 04

ChatGPT, Google and Meta: three different hunting grounds

The radar chart below shows three different hunting grounds.Google is balanced, battle-tested and measurement-rich.The shapes tell the story before the budget does: one fox follows the question, one follows the query, one follows the crowd.

ChatGPT scores 9/10 intent quality, 9/10 privacy, 4/10 measurement, 7/10 audience, 1/10 retargeting, 6/10 creative. Google scores 8,5,9,10,8,8. Meta scores 5,3,8,10,10,9. __FF_COMMENT_BLOCK_1__ __FF_COMMENT_BLOCK_2__ __FF_COMMENT_BLOCK_3__ Intent Quality Privacy Posture Measurement Audience Scale Retargeting Creative Format __FF_COMMENT_BLOCK_4__ __FF_COMMENT_BLOCK_5__ __FF_COMMENT_BLOCK_6__ __FF_COMMENT_BLOCK_7__ ChatGPT Google Meta

Scores are directional (0-10). Intent Quality measures how precisely targeting captures active user intent at point of need. Sources: Digital Applied, Microsoft Advertising research, folkfox analysis.

ChatGPT Ad Router vs Google Ads vs Meta Ads: targeting architecture, bidding, privacy and measurement compared
Dimension ChatGPT Ad Router Google Ads / AI Overviews Meta (Facebook / Instagram)
Targeting method Live conversational context: current chat thread, language, approximate region. Optional past-chat personalisation if opted in. CONFIRMED Keywords, query intent, match types, quality score. AI Overviews add a second contextual layer: ads must match both query and AI-generated summary. Demographics, interest graph, behavioural audiences, look-alike audiences built from pixel events and the social graph.
Third-party cookies None documented. OpenAI privacy policy confirms no user data sold, no conversations shared. Chrome reversed cookie deprecation Apr 2025. Safari/Firefox remain cookieless by default. First-party data mechanisms now primary. Extensive cross-site tracking via pixel and CAPI. Heavy GDPR scrutiny of profiling practices.
Retargeting No cross-site retargeting. Same-session conversation refinement only. No documented cross-session tracking outside OpenAI environment. Full remarketing: site visits, customer lists, app events, customer match. Robust cross-session capabilities. Extensive: dynamic product ads, cart abandonment, CRM-based lists, look-alikes based on converters.
Ad unit types Sponsored Recommendation (native text below answer), Companion Display (visual/sidebar), Native Content Card (product feed). All labelled "Sponsored." Text search ads, Shopping/PLAs, Performance Max multi-format. AI Overviews surface ads above, below, or within summaries. Feed ads (image/video), Stories, Reels, in-stream video, Shopping/Catalog ads, lead forms.
CPA bidding Live from 28 May 2026 (Digiday). Requires pixel + Conversions API. Still in limited rollout. Mature. Full conversion tracking, offline imports, Smart Bidding across multiple action types. Mature. Purchase, lead, custom event optimisation. Multi-touch attribution and lift studies available.

Scores are directional, from 0 to 10.Privacy Posture measures how little personal history the platform needs to make the match.Treat this as a map for thinking, not a market-law tablet.

Section 05

Three ad units, three creative trails

OpenAI’s help centre documentation and early advertiser guides point to three unit types.That means each needs its own tone, texture and task.

Sponsored Recommendation

Placement: a native text unit directly below the answer.
Format: text-first, usually 60 to 100 words, with an optional small image.
Creative principle: extend the answer. Make it feel like the next helpful paragraph, not a banner in disguise.
Best for: SaaS, B2B tools, advisory offers and considered purchases.

Companion Display

Placement: a visual tile beside the conversation as the user scrolls.
Format: image or rich media, headline, short body and CTA.
Creative principle: sit beside the answer like a lantern on the path. Reinforce the brand without interrupting the read.
Best for: travel, fashion, entertainment, CPG and consumer categories.

Native Content Card

Placement: a product-style card below the answer with image, price, rating and attributes.
Format: structured feed data, built for large catalogues.
Creative principle: shoppable proof. Surface the SKU that matches the conversational brief.
Best for: retail and ecommerce with rich, well-labelled product feeds.

73%
Higher CTR for conversational AI ads vs traditional search
Microsoft Advertising research, August 2025

Microsoft’s Copilot ad research is the strongest independent performance clue so far.The useful lesson is not that ChatGPT will copy Copilot’s numbers exactly.

"The challenge with ads is that a user would just start doubting everything."

Perplexity AI executive, Financial Times, February 2026. Perplexity retreated from advertising entirely that month.
Section 06

Agentic Commerce: what vanished and what still matters

OpenAI and Stripe announced the Agentic Commerce Protocol in September 2025.In theory, the fox could guide the buyer from question to checkout without leaving the clearing.

The first trail was retail. Etsy went live and Shopify was announced.Then the trail changed. Instant Checkout was discontinued in March 2026.

Instant Checkout discontinued March 2026

Paz.ai documents that “Buy it in ChatGPT” launched in September 2025 and was discontinued in March 2026. Brands did not want to become quiet fulfilment dens with no customer relationship. ChatGPT now drives discovery and sends users to merchant sites. ACP still matters as a standard, but the landing page matters again. When the user arrives from ChatGPT Shopping, they are already close to decision. Give them a clean path, not a maze.

ACP Instant Checkout funnel: five steps, their technology, and June 2026 status
Step What happened Technology Status (June 2026)
1. DiscoveryUser query triggers relevant product card below ChatGPT answerAd Router + product feedLive
2. SelectionUser taps product; ChatGPT presents variants, price, shippingACP capability negotiationDiscontinued
3. Payment tokenStripe generates SPT scoped to merchant + amountStripe Shared Payment TokensAvailable via spec
4. Order APIChatGPT passes SPT + order details to merchantACP order lifecycle APIAvailable via spec
5. ConfirmationOrder confirmation surfaces in chat. 4% fee applied.ACP webhook + ChatGPT UIDiscontinued
Section 07

CPA bidding, the pixel, and the measurement map

Digiday confirmed on 27 May 2026 that Cost Per Action bidding went live for selected advertisers.The catch is measurement: every fox needs a trail, and every trail needs markers.

■ AI to Direct (missed) 28% ■ True organic 35% ■ AI referral (captured) 22% ■ Paid 10% ■ Other 5%

Illustrative GA4 attribution for a B2B site receiving AI-referred traffic. The orange segment represents AI-originating sessions currently absorbed into Direct. Sources: xSeek (2026), LLMrefs, folkfox analysis.

Fixing the GA4 attribution gap

GA4 still has no native “AI Search” channel.There is no reliable boardroom story without custom channel grouping, clean UTMs and server-side conversion capture.

Create a custom channel group in GA4 called “AI Search” with this source regex:

chatgpt|openai|perplexity|claude|anthropic|gemini|copilot|grok|you\.com|phind

Move this channel above Referral in priority order. Combine it with xSeek’s GA4 configuration guidance, LLMrefs for citation tracking, and Profound for AI revenue attribution.

Measurement principle for AI spend

Measure your organic AI referral baseline before buying media. That separates GEO momentum from paid lift and gives you a cleaner read on incrementality. Without the baseline, ChatGPT spend becomes a fox trail in fog: interesting, expensive and hard to follow.

Section 08

Pricing dynamics: premium intent, premium CPM

ChatGPT ads launched at roughly USD 60 CPM for early brand advertisers, according to Search Engine Land, January 2026.A user mid-thought, weighing options in plain language, is a very different commercial moment from a scroll-pause on Instagram.

ChatGPT (launch)
$60 CPM
ChatGPT (self-serve)
$25-60 CPM
Copilot (est.)
$20-40 CPM
Meta (lead gen)
$20-40 CPM
Meta (reach)
$10-15 CPM
Google Display
$4-12 CPM
Recommended CPC start
$3-5 CPC

Directional only.OpenAI has not published official CPM benchmarks, so treat every figure as a lantern, not a law.

Public CPA benchmarks are not ready yet.Start by using Google and Meta CPA benchmarks as a ceiling, then test down as attribution becomes cleaner.

Section 09

From $200k to self-serve: the 120-day sprint

The speed of the rollout is the story.By 5 May 2026, OpenAI had moved to self-serve access with no minimum spend.

  1. Nov 2022
    ChatGPT public launch
  2. Sep 2025
    ACP + Instant Checkout announced
  3. 15 Jan 2026
    OpenAI advertising approach published
  4. 31 Jan 2026
    $200k minimum confirmed
  5. 9 Feb 2026
    Ads go live (Free/Go US)
  6. Mar 2026
    Instant Checkout discontinued
  7. 5 May 2026
    Self-serve, zero minimum
  8. 28 May 2026
    CPA bidding launches
Perplexity: the contrarian data point

Search Engine Land confirmed in February 2026 that Perplexity AI was phasing out sponsored answers, citing the risk that ads could weaken trust in neutral answers. Perplexity chose subscription revenue over CPMs. OpenAI chose scale. Two paths, one warning: in AI search, trust is the territory.

Section 10

Regulation, risk, and readiness

OpenAI’s no-PII guarantee is a structural comfort point: advertisers receive aggregated metrics, not conversation content, names, emails or precise locations.Healthcare and many financial services are restricted.Build the citation trail before paid access opens, so the den is ready when the gate moves.

ChatGPT advertising regulatory landscape: GDPR, EU AI Act, OpenAI ad policies
Framework Current status Advertiser implication
GDPR No-PII sharing policy aligns with purpose limitation and data minimisation. Italian Garante EUR 15M fine (Dec 2024) annulled by Rome court March 2026, signalling need for more nuanced AI enforcement. Lower GDPR exposure than cookie-based platforms. Standard DPA applies. Monitor evolving enforcement.
EU AI Act (Article 50) Transparency obligations phase in mid-2026. Requires disclosure of AI-generated content and labelling of AI-assisted media. Draft Code of Practice published late 2025. Mandatory "Sponsored" labelling already in place. Build AI-generated creative disclosure into approval workflows.
OpenAI Ad Policies Prohibited: gambling, adult, illicit drugs, alcohol, tobacco, deceptive content, political. Restricted: financial services, healthcare, mental health. Eligible: household goods, local services, travel, entertainment, digital products, education. iGaming and healthcare currently blocked. FinTech requires case-by-case review. Music/entertainment eligible. Build GEO foundations now for categories not yet open.
For folkfox’s regulated verticals

For iGaming and healthcare brands currently blocked from direct paid ads: build the organic AI citation architecture now. When the category opens, brands with clean, cited, answer-ready pages will not be starting cold. Read why your best pages miss AI citations and what to do about it.

Section 11

What close observers are saying

The people closest to this shift are circling the same clearing: performance opportunity, trust risk and careful labelling.The winning brands will not chase every shiny signal. They will move quietly, measure cleanly and keep the user’s trust warm.

"Cost-per-action advertising is a logical next step for OpenAI's ads business, expanding the advertiser base and bringing the system closer to competitors like Meta and Google."
Claire HolubowskyjSenior Research Analyst, Enders Analysis • via Digiday, May 2026
"Any hint that AI shopping is quietly steering users toward paid influence could confirm a fear that the system isn't on our side."
Heidi WalduskyVP Brand and Integrated Marketing, Quad • 2026 AI Trust Survey
"Copilot users engage significantly more with ads: driving 73% higher click-through rates and 16% stronger conversion rates versus traditional search."
Kevin KleinHead of Advertiser Analytics, Microsoft Advertising • August 2025
"Ads do not influence the answers ChatGPT gives you. Answers are optimised based on what's most helpful to you. Ads are always separate and clearly labelled."
OpenAIOfficial advertising statement, January 2026
Section 12 · HowTo

Seven fox-smart moves for semantic targeting now

The folkfox pillar article on ChatGPT Paid Ads sets the strategic scene.Think of it as a fox run: find the scent, mark the trail, move with patience, pounce only when the proof is clean.

Access Ads Manager

Register with OpenAI or an approved partner such as StackAdapt, Pacvue or Criteo. Self-serve Ads Manager supports CPM, CPC and CPA with no minimum spend from 5 May 2026.

Install the pixel first

Put the OpenAI JavaScript conversion pixel and Conversions API on product, sign-up and confirmation pages before spending. CPA optimisation needs the trail markers in place.

Build intent clusters

Use LLMrefs or xSeek to see how users describe the need in ChatGPT. Group campaigns by conversational themes, not keyword scraps.

Write native creative

Write Sponsored Recommendation copy as if it were paragraph six of the answer: specific, useful, 60 to 100 words, and one clear next step. No slogans. No fog. The ad that reads like advice wins.

Set directional bids

Start CPC at USD 3 to 5. Plan CPMs of USD 25 to 60. Treat early campaigns as structured tests for assisted conversions and branded search lift, not neat little last-click ROAS boxes.

Build GEO in parallel

Organic AI citation is the trust trail that makes a paid impression land warmer. If ChatGPT already cites your brand, the ad feels less like an ambush and more like the next logical step.

Fix attribution first

Apply the GA4 AI Search regex. Add UTMs. Connect Profound for AI revenue attribution. Then, and only then, start talking about incrementality.

Want folkfox to build the trail?

folkfox designs AI visibility architectures, PPC strategies, and GEO content frameworks for brands across FinTech, healthcare, music and iGaming.Less frantic forest noise. More clean, clever, commercially useful movement.

FinTech & SaaS

FinTech buyers already in the den

FinTech decision-makers are using ChatGPT and Copilot for long, layered research journeys. Paid adjacency is already live, and the early trail is warm.

See our FinTech approach →
iGaming

iGaming waits, GEO hunts

Consumer iGaming ads are currently prohibited. Build organic AI citation authority now so the trust trail exists when the category opens.

See our iGaming approach →
Healthcare

Healthcare needs careful citation

Healthcare advertising is restricted. Informational GEO, careful schema and compliant AI citations are the safest routes through the thicket.

See our healthcare approach →
Sources and references
  1. [STD] OpenAI. "Ads in ChatGPT." Help Center. Jan 2026. help.openai.com
  2. [STD] OpenAI. "Our approach to advertising." Jan 2026. openai.com
  3. [STD] OpenAI. "New ways to buy ChatGPT ads." May 2026. openai.com
  4. [STD] OpenAI. "Ad policies." May 2026. openai.com
  5. [IND] Digiday. "OpenAI turns on CPA ads." May 2026. digiday.com
  6. [NEWS] TechCrunch. "ChatGPT reaches 900M WAU." Feb 2026. techcrunch.com
  7. [IND] Pennock. "ChatGPT Ads 2026 Guide." pennock.co. CPM $25-60, CPC $3-5 benchmarks.
  8. [ACA] Grbovic et al. "Scalable Semantic Matching." arXiv 2016. arxiv.org
  9. [IND] MMA Global. "SIFT." 2025. mmaglobal.com
  10. [IND] Microsoft Advertising. "73% higher CTRs." Aug 2025. ads.microsoft.com
  11. [IND] Quad / Harris Poll. "AI sponsored trust." Apr 2026. quad.com
  12. [IND] Stripe. "Instant Checkout in ChatGPT." Sep 2025. stripe.com
  13. [IND] Paz.ai. "Instant Checkout discontinued." Mar 2026. paz.ai
  14. [STD] EU AI Act. "Article 50 transparency." artificialintelligenceact.eu
  15. [NEWS] Search Engine Land. "Perplexity stops ads." Feb 2026. searchengineland.com
  16. [DB] xSeek. "Measure AI Search." Apr 2026. xseek.io

Fresh tracks

Latest from the Paid Search trail

    Chart data 1
    ChatGPT (launch)$60 CPM
    ChatGPT (self-serve)$25-60 CPM
    Copilot (est.)$20-40 CPM
    Meta (lead gen)$20-40 CPM
    Meta (reach)$10-15 CPM
    Google Display$4-12 CPM
    Recommended CPC start$3-5 CPC
    Keep reading

    Read more on this topic

    The den

    Where to?

    Pricing

    Choose a section. Enter opens it, Escape continues reading.

    Cookie preferences

    folkfox uses data the way we use strategy: only when it earns its place.