How Do You Actually Run ChatGPT Ads? We Distilled 137K AI Answers Into One SOP — Start by Seeing the "Two Shelves" (GEOly AI Playbook)
Hi, I’m Riven.
In our previous GEOly AI data report (where we pulled 84,665 ChatGPT ad cards), we made one thing clear: what this new shelf looks like. But everyone asked me the same follow-up: I get it now — so how do I actually run ads on it?
This is the answer. We took GEOly AI’s ChatGPT monitoring data — 136,999 answers, 5,479 ad samples, 521 advertisers (US + GPT-5.5) — and distilled it into a standard operating procedure you can hand straight to a client. Not an analysis report — an SOP. Today I’m making the core decision logic public.
First, the sample: these aren’t gut calls
The entire evidence base behind this SOP comes from GEOly AI’s real monitoring. Let’s put the methodology on the table first:
| Monitoring scope | Value |
|---|---|
| Platform / model | ChatGPT (GPT-5.5) |
| Answer sample size | 136,999 |
| Ad sample size | 5,479 |
| Advertisers covered | 521 |
| Market | United States |
Note: the evidence is a cross-sectional snapshot (~90% from a single day). It’s a real behavioral signal visible on the ChatGPT user side, not OpenAI’s back-end data. The ad ecosystem moves fast, which is exactly why the final stage of this SOP is always “continuous monitoring.”
The master key: ChatGPT has “two shelves,” and ads only sell on one
This is the first principle behind the whole playbook. Every action is derived from this one sentence:
ChatGPT answers shopping questions in two modes, and the business logic of those two modes is the opposite of each other.
| Research mode | Recall mode | |
|---|---|---|
| Typical query | ”best / top X”, informational | ”X vs Y”, “is it worth it”, “premium / luxury” |
| Share | ~67% of queries | bottom of the funnel, comparison/consideration |
| Browses the web? | Yes (40%+ triggered) | No (vs-queries only 18.6%) |
| Organic shopping cards | Many (best-type 66.8%) | Few (vs-type 12.7%) |
| Ad rate | ~0.9% (basically can’t buy in) | 8–22% (heavily monetized) |
| Who wins | Walmart / Target / Best Buy (free organic shelf) | mid-tier retail + DTC (paid ad cards) |
| Play | GEO (Reddit + reviews + retail feeds) | buy ad cards |
SOP rule #1: first determine which mode your target query belongs to, then decide the play.
- Research-mode questions (best/top, informational): don’t count on ads — the ad rate is just 0.9%. Win with GEO.
- Recall-mode questions (X vs Y / worth-it / premium): this is the ad battleground, ad rate 8–22% — and the gate is budget + model memory, not “whether the AI recommends you organically.”
In one line: GEO and ads are two different battlefields. Do both, but never confuse them.
Four counterintuitive truths the data locked in
These four run through the whole SOP, and each one upends your instinct about “AI advertising”:
① The grounding paradox: when ChatGPT shows an ad, it almost never browses the web. In answers that contain ads, only 1.0% browse the web, 1.0% cite sources, and 0.8% carry an organic shopping card — versus 40.6% / 40.4% / 57.6% in answers without ads. In other words, the moment ChatGPT serves an ad, it’s essentially answering “from memory,” not researching on the fly.
② Ads are decoupled from organic recommendation. After stripping out the trailing sponsored card, only about 2.7% of the ad-bearing answer body actually mentions the advertiser, and the median depth of first appearance is 95% (the very end). Put differently: 90%+ of the brands in ad cards are not the brands the AI recommends organically. The ad is the channel that inserts you into the last slot of the answer when the AI didn’t recommend you.
③ Giants win organic for free; challengers pay to buy ads. Walmart / Target / Best Buy dominate the organic shopping cards — but their ad spend is zero. The brands actually buying ads are Quince, Sephora, Chewy, e.l.f. — the challengers. Advertising is fundamentally a challenger’s weapon against an organic shelf monopolized by giants.
④ The first-mover window is still wide open. Overall ad penetration is just 4% (only 1 in every 25 answers carries an ad), and 84% of ad topics have just one advertiser, with giants entirely absent. Right now — claiming high-intent comparison queries at low cost — is the moment of weakest competition.
Turning “understood” into “executable”: a 6-stage SOP
We broke the play into 6 stages, each with a defined deliverable: Diagnose → Decide → Plan → Create → Launch → Monitor. Here’s each stage.
Stage 1 · Diagnose: should you even run ads, and where do you stand?
The first move isn’t picking keywords — it’s classifying the client’s organic shelf position, which decides whether ads are “icing on the cake” or “your only lifeline”:
| Type | Organic shopping card | Role of ads | Examples |
|---|---|---|---|
| Reinforce | Already winning organic | Double down (icing) | Home Depot, Etsy, Quince |
| Compensate | Occasional, weak | Fill the gap (necessary) | Brilliant Earth, Select Blinds |
| Only shelf | Nearly/entirely invisible | The only way onto the AI shelf | Kashwére, affiliate sites, B2B |
Key data: 60% of advertisers are completely invisible on the organic shopping shelf (only 204 of 505 appear on organic cards). For these invisible brands, ChatGPT ads aren’t optional — they’re mandatory.
Stage 2 · Decide: keyword priority, sort by this table
Bucket the client’s real category queries into the “two modes,” then within recall mode rank by ad rate — this is your keyword list:
| Intent (query) | Ad rate | Value |
|---|---|---|
| X vs Y comparison | 22.2% | 🟢 Top battleground |
| Worth it? | 13.1% | 🟢 High value |
| premium / luxury | 8.1% | 🟢 High AOV fit |
| Other / informational | 4.6% | 🟡 Situational |
| budget | 3.8% | ⚪ Weak |
| best / top lists | 0.86% | 🔴 Can’t buy → switch to GEO |
One especially valuable finding: when a query itself carries consideration phrasing like premium / luxury / worth-it, the ad rate is 11.63%; without it, just 3.65% — a 3.2× gap. These words are monetization triggers in themselves — lock onto them first when choosing keywords.
Stage 3 · Plan: map topics × budget × roles into one media plan
Once the keyword list is set, the next step is to turn “where to spend, how much, and what leads what” into an executable plan:
- Topic selection: anchor on recall-mode, high-intent comparison terms, then layer zero-ad / single-advertiser whitespace as first-mover slots; go heavy on high-AOV consideration scenarios (weddings / jewelry / silk pajamas) — “wedding” queries run an ad rate of 8.6%, 2× the baseline.
- Budget allocation: concentrate on head comparison terms (the top 50 advertisers account for 66.8% of impressions — effective concentration is correct), and test long-tail whitespace with small spend (38% of advertisers ran just once — testing is normal). In the first-mover phase you can hold slots cheaply.
- Auction expectation: auctions are thin (84% of topics have a single advertiser), so bid pressure is low; but a few already-crowded categories (Essential Oils 17, Backdrops 15, Pants 14 advertisers) need higher budget or simply avoiding — run the ROI first.
- Delivery channel: ChatGPT sponsored cards currently run through Criteo; the client needs an eligible advertising entity + product feed + deep landing pages ready.
Stage 4 · Create: fight “last-slot insertion” — the logic is the opposite of Google Shopping
Because ad cards sit almost at the very end of the answer (median depth 95%), the AI has already finished recommending others above you. So the creative’s job is to pull attention back. Every rule is reverse-engineered from the data:
| Rule | Data basis | What to do |
|---|---|---|
| Don’t lead with price; lead with category/collection | Only 0.5% of creatives show price | Use category/collection lead lines in titles (“Cosmetics”, “Gift The Softest PJs”) |
| The creative must be strong enough to “intercept” | Sits at the end; others already recommended above | Strong brand identity + benefit + gift/seasonal/custom angle |
| Use listicle + gifting angles | Top words: best/top/2026, custom 140, gift 63 | Angle library: category nouns / listicle / custom-gift |
| Land on deep pages; never the homepage | Only 8.3% land on homepage, 22.8% on PDP, 19.3% on PLP | Go straight to product/listing pages |
| 1–2 hero creatives + long-tail testing | 67% of creatives appear once; the top hero was reused 64× | Concentrate on hero creatives, test the long tail with small spend |
Stage 5 · Launch: actually get the cards live
Everything you planned has to reach a shippable state. Three things, none optional:
- Access ready: confirm the advertising entity, Criteo / ChatGPT ad eligibility, and that the product feed and landing site are in place;
- Card spec: title (category/collection lead) + image (strong brand identity) + body (benefit / gifting angle), always landing on PDP / PLP;
- Coordinate with the shelf: if the client is also doing agentic commerce, the ad card and the shelf product card must stay consistent — the same product’s title, selling points and landing page should match across the “ad card” and the “AI shelf card,” so you don’t run two conflicting sets of data.
Stage 6 · Monitor: treat monitoring as infrastructure, not a wrap-up
This is a young, volatile ad surface — the 8.5× inventory surge in mid-June is proof. So monitoring matters more than any static media plan. Track four things continuously:
- You vs competitors — which queries you’re winning ad slots on (held / lost / whitespace shifts);
- Whether you’re also recommended organically and which sources cite you (research-mode progress);
- Whether your organic shelf position is shifting (reinforce / compensate / only-shelf);
- Brand safety — whether your card shows up next to irrelevant queries (semantically broad categories leak in spillover advertisers, e.g. “Calendars” pulls in Asana / Monday).
Remember it as one picture: two battlefields, fought separately
| Battlefield | Entry | Powered by |
|---|---|---|
| Research mode (free answers / citation lists) | GEO | Reddit presence + editorial review coverage (TechRadar / Allure / Consumer Reports) + product feeds into retail shopping indexes |
| Recall mode (sponsored cards) | Buy ads | Budget + model memory + strong creative |
One number explains why you must do both: reddit.com was cited organically by ChatGPT 87,942 times (appearing in ~30% of answers), making it the #1 research-mode source — yet 64% of advertisers have never been cited organically.
Source authority can’t be bought — you earn it with content and word of mouth (GEO). Ad slots are bought with budget and creative. Do both; the key is to tell the scenarios apart.
A condensed checklist for cross-border sellers
If you only take one page away, it’s these ten:
- Classify your organic shelf: are you Reinforce / Compensate / Only-shelf? (decides the role of ads)
- Intent inventory: list your category’s recall-mode, high-monetization queries (X vs Y / worth-it / premium);
- Mode routing: research-mode words → GEO, recall-mode words → ads;
- Whitespace scan: find topics with zero or a single advertiser (first-mover slots — 84% of topics have just one);
- Keyword priority: vs ＞ worth-it ＞ premium ＞ high-monetization scenarios; avoid best/top;
- Budget: concentrate on head comparison terms + small whitespace tests;
- Creative: lead with category/collection not price; strong identity to beat last-slot insertion; gift/custom angles;
- Landing: PDP / PLP, never the homepage;
- Monitor: you vs competitors / organic recommendation / shelf position / brand safety — weekly & monthly;
- Coordinate: keep GEO (research mode) and ads (recall mode) on separate battlefields, but unify them into the AI-traffic / AI-revenue accounting.
Boundaries & honest disclaimers (must be said)
I don’t want to hand you something that looks certain but falls apart under scrutiny, so here are the limits of these conclusions:
- The evidence is a cross-sectional snapshot (~90% from a single day) — no trend promises, depends on continuous monitoring;
- US + GPT-5.5 only — other markets / models need separate testing;
- Each answer records only one primary card, so per-response advertiser/creative counts are a lower bound;
- “Ad rate” is a real ChatGPT behavioral signal; but the “volume distribution across intents” is shaped by our probe mix and does not represent real user search volume;
- KPI settlement uses the client’s ad back-end + GA4 real data; monitoring data is only for slot-holding and trend.
Closing: about this SOP, and GEOly AI
Last time we used GEOly AI to see what the new shelf looks like; this time we turned it into the standard moves to actually run ads. The logic is simple: in the Google era we learned SEO + SEM; in the AI-answer era, GEO + AI ads are the new default — they compete for the same answer-page real estate, one free and one paid, with completely different entry paths.
And this whole play only works because of one thing: you have to see what’s actually happening on that answer page. Who’s advertising, on which queries, what the creatives look like, whether you’re recommended organically — that’s exactly what GEOly AI does.
Want to see what your own category’s “two shelves” look like inside ChatGPT? Come run a diagnosis at GEOly AI. This ChatGPT ad window opened earlier than you think, and it’ll close faster than you think too — follow Riven, follow GEOly AI, and let’s claim this new shelf in AI answers together.
— Riven & the GEOly AI team
Data note: conclusions here are based on GEOly AI’s public monitoring of ChatGPT (GPT-5.5, US) — 136,999 answers / 5,479 ad samples / 521 advertisers. The evidence is a cross-sectional snapshot reflecting real behavioral signals visible on the user side; it does not represent OpenAI’s official ad data. Bidding and cost-related conclusions are inferred from impression patterns.