Ads in ChatGPT are paid placements that render in a labeled, visually tinted container below or beside the assistant's answer — never inside the answer text itself. OpenAI shows them to logged-in Free and Go users aged 18+, selects them from conversation context rather than from the model's reasoning, and keeps Plus, Pro, Business, and Enterprise tiers ad-free.
That distinction has a practical consequence most tracking setups miss: a brand name on the screen no longer proves the model picked it. Sometimes the assistant recommended you. Sometimes a competitor paid for that line.
This piece covers where paid units render on each assistant, exactly how they are labeled, whether buying a slot changes organic answers, and a measurement protocol that keeps bought impressions out of your earned-visibility numbers.
We ran a 30-day panel to size the problem. In our tracking data, 69% of the ad units we logged promoted a brand that did not appear anywhere in the organic answer text — meaning a naive string match would have credited those brands with a recommendation they never earned.
What are ads in ChatGPT?
Ads in ChatGPT are paid placements shown in a separate, tinted container below or beside the generated answer. OpenAI selects them from conversation context, labels them "Sponsored" or "Ad," and serves them only to Free and Go tiers. They sit outside the answer text.
To a human reading the screen, the boundary is obvious — shaded box, advertiser favicon, tag string. To a scraper pulling rendered DOM, or an analyst eyeballing a screenshot at speed, the boundary is a CSS class. That gap is the whole reporting problem in miniature.
OpenAI announced it was testing ads in ChatGPT before opening the format up, and its help center documentation sets the current rules:
- Tiers: Free and Go only. Plus, Pro, Business, and Enterprise stay ad-free.
- Users: logged-in adults (18+). Logged-out sessions and minors are excluded.
- Excluded categories: no ad serving near health, mental health, politics, or other sensitive topics.
- Separation: ad selection runs after answer generation, not inside it.
The ad-free paid tier is the single most useful fact in this article. It gives you a control group you can run the same prompt against, in the same hour, to see whether the organic answer differs. Section 8 makes that a protocol.
Who pays for the slot, and what advertisers actually buy
Advertisers do not bid on an answer. They bid on a conversation context. OpenAI's stated model is that a targeting system reads the conversation for commercial signals, then selects an ad — the language model that wrote the answer has no role in choosing which advertiser appears beneath it.
Three things follow, and they matter for how you plan:
- You cannot buy an organic recommendation. The two systems are separated by design, and our paired testing (below) found no measurable crossover.
- You cannot block a competitor's ad on "your" prompt. There is no exclusivity on a conversation topic the way there is with a brand-term trademark policy in classic search.
- Ad eligibility follows commercial intent, not brand mention. A prompt that names you but reads as informational will usually draw no inventory at all.
Where paid slots appear on each assistant right now
Paid inventory in AI assistants is not one market. It is five or six separate products with different labels, different eligibility, and at least one platform that tried ads and reversed course. State of play as of July 2026:
| Assistant | Paid inventory status | Where it renders | Label observed | Who sees it |
|---|---|---|---|---|
| ChatGPT | Live since Feb 2026 | Tinted box below the answer; merchant/product cards on shopping intents | "Sponsored" at launch; a more prominent "Ad" tag on bottom-of-answer units | Free and Go tiers, 18+, US plus CA/AU/NZ expansion |
| Google AI Overviews | Live | Inside and below the AI answer block, including product carousels | "Sponsored" | Eligible commercial queries, served from existing Search/Shopping campaigns |
| Google AI Mode | Live | Below or within the response | "Sponsored" | Rolling out by market |
| Microsoft Copilot | Live | Beneath the response; Showroom Ads in limited pilot | "Sponsored" plus a Microsoft Advertising attribution | Copilot search users |
| Perplexity | Withdrawn Feb 2026 | Previously sponsored follow-up questions and sidebar units | "Sponsored" | No live ad inventory |
| Gemini app | Not commercial inventory | — | — | Google has disputed reports of a 2026 app rollout |
| Claude, Grok | No announced ad units at time of writing | — | — | — |
Two entries deserve a footnote.
Perplexity began testing sponsored follow-up questions with named brands before pulling the format entirely, citing user trust. Ad inventory can vanish as fast as it appears — your classification rules need to survive that, which is why the protocol below classifies on container structure rather than on a hardcoded label list.
On Gemini, Adweek reported advertiser briefings pointing to a 2026 rollout, after which Google publicly disputed the claim. Treat the Gemini app as ad-free until you observe a labeled unit yourself. We cover the broader monetization arc in our breakdown of sponsored answers across AI search.
How ad labels differ from citation labels
Ad labels and citation labels look similar and mean opposite things. A citation says the model used this source. An ad label says this advertiser paid to sit next to the answer. Confusing them is the most common classification error we see in client dashboards.
| Signal | Ad container | Citation |
|---|---|---|
| Background | Tinted / shaded card | None — inline with answer |
| Boundary | Bounded block with padding | No container |
| Brand icon | Advertiser favicon | Domain chip or superscript number |
| Action | Click-through CTA | Source link, no CTA |
| Tag | "Sponsored" or "Ad" | Numbered reference or source rail |
| Position | Below or beside the answer body | Inline, or in a source list |
The visual cues are consistent enough to automate. The tag strings are the unstable part. Over our tracking window we logged three distinct label variants inside ChatGPT alone: a "Sponsored" string on inline merchant cards, a standalone "Ad" tag on bottom-of-answer units after OpenAI made the tag more prominent, and product carousels where the label sat at container level rather than on each card.
Practical rule: classify on container structure first, label string second. Label strings change on the platform's schedule; the tinted container has been stable across every rollout we have tracked.
Why sponsored slots quietly inflate AI visibility metrics
A sponsored slot inflates your metrics whenever your measurement counts brand strings on a page instead of brand mentions in an answer. Three distinct failure paths, and most teams have at least one live right now.
Path one: naive extraction. The tracker captures the full rendered response and regex-matches brand names. The ad unit is part of that capture. Every paid impression becomes a "mention."
Path two: panel identity drift. Whether ads appear at all depends on how the session is authenticated. Free-tier logged-in US sessions see ads; API runs, paid tiers, and other regions do not. If your vendor rotates identities without recording which one produced each response, your ad exposure rate moves for reasons that have nothing to do with your market.
Path three: screenshot review. Humans skimming evidence screenshots for a monthly deck routinely miss a tint boundary at thumbnail size. We have seen paid competitor placements pasted into slides as proof of an organic recommendation.
This is the same class of silent bias as when refusals, timeouts, and tool failures skew visibility metrics: the number still renders, it just no longer measures what its label claims.
Our 30-day panel: how much tracked inventory is actually paid
We ran a dedicated audit to size ad contamination rather than estimate it. Method below so you can reproduce it against your own prompt set.
Method. 1,240 tracked prompts across four verticals (B2B SaaS, DTC ecommerce, local services, fintech), run daily for 30 days from 15 June to 15 July 2026, across six assistants. ChatGPT runs executed on both a US Free-tier logged-in session and an ad-free paid tier — same prompt, same hour. That produced 37,200 ChatGPT Free-tier responses and a paired ad-free control for each.
What we found:
- 17.4% of Free-tier ChatGPT responses contained at least one ad unit. Well below the ~49% US figure circulating from third-party rollout trackers this spring — the gap is explained by prompt mix, not by measurement error.
- Commercial-intent prompts carried ads at 38.2%; informational and B2B prompts at 4.1%. Commercial prompts were 39% of our panel. Any published "share of responses with ads" number is really a statement about that publisher's prompt mix, which is why you should size this on your own prompts rather than borrow ours.
- Of the 6,473 ad units logged, 69% promoted a brand absent from the organic answer body. Only 31% showed a brand that had also earned an organic mention in the same response.
- For three tracked DTC brands, counting ad-slot text inflated measured AI share of voice by 6 to 11 percentage points. For the B2B SaaS accounts, inflation was under 1 point — their prompts rarely trigger inventory.
- Ad units clustered on turn one. 71% of ad-bearing conversations showed the unit on the first response; a follow-up turn on the same topic usually returned none.
The last three findings are the operational ones. Sell physical products or run high commercial-intent prompts, and unaudited ad text can move your headline number by double digits. Sell enterprise software, and today's contamination risk is small — but the prompt set that matters to you is the one drifting toward commercial intent as buyers start asking assistants for shortlists rather than for explanations.
A provenance ladder for classifying every AI mention
Before you report a mention, you need to know how it got there. We use a five-tier provenance ladder — P0 to P4 — and every tracked mention gets exactly one tier. This framework turns "the brand appeared" into a defensible number.
| Tier | Mention type | Definition | Counts toward AI share of voice? |
|---|---|---|---|
| P0 | Earned mention | Brand named in the model-generated answer body, no ad unit for that brand on screen | Yes — this is the core metric |
| P1 | Earned citation | Brand's URL appears in the source list or citation rail | Yes, but as a separate citation metric |
| P2 | Paid unit | Brand appears only inside a labeled ad container | No — log as paid AI impressions |
| P3 | Blended | Brand appears in both the organic body and an ad unit in the same response | Yes, flagged; requires a control run |
| P4 | Ambiguous | Merchant card, feed-driven product module, or affiliate block with unclear labeling | Quarantine until classified by a human |
P3 is where judgement is required and where most disputes happen. A blended response is not proof that spend bought the organic mention, and it is not proof that it didn't. It is a prompt for the control run described next.
P4 exists because product modules are genuinely hard. Feed-driven shopping cards can be organic, paid, or a blend depending on surface, and the label often sits on the carousel rather than the card. Quarantine is cheaper than a wrong tier that ships to a board deck.
Run the ladder against competitors too, not just your own brand. Half the value is knowing whether the rival who "keeps showing up in ChatGPT" earned it or bought it — the same question behind building an AI-native competitive set.
How to separate paid from earned in your reporting
Seven-step protocol we run. Roughly a day to set up, then it runs unattended.
- Pin your session identity. Decide whether your panel measures the ad-exposed experience (Free-tier, logged-in, US) or the ad-free one (paid tier or API). Record the identity with every single response. Never mix them in one trend line.
- Capture structure, not just text. Store the response with container boundaries intact — DOM structure or a segmented capture. Plain-text extraction destroys the only reliable ad signal you have.
- Classify by container, then label. Flag any tinted, bounded, CTA-bearing container as a candidate ad unit. Confirm with the label string. Do not rely on the string alone.
- Assign a provenance tier. Apply P0–P4 to every brand occurrence, including competitors. One tier per occurrence.
- Run the paid-off control. For every response containing an ad unit, re-run the identical prompt within the same hour on an ad-free tier. Diff the organic brand set.
- Report two numbers, never one. Earned AI share of voice (P0 + P1) and paid AI presence (P2) go in separate rows. A combined figure is not a metric, it is an average of two unrelated things.
- Re-baseline monthly. Ad eligibility, label strings, and country coverage all change. Sample 50 responses by hand each month and check your classifier still agrees with a human.
Step 5 is the one teams skip, and it is the one that answers the question executives actually ask.
Does buying a slot change your organic recommendations?
No — not within measurement noise, on the prompts we tested. OpenAI's published position is that ad selection sits outside answer generation, and Perplexity said the same about its sponsored follow-ups while the format was live. Those are vendor statements, not audits, so we tested it.
Across 214 paired runs — each an ad-serving response matched to an ad-free control on the same prompt within the same hour — the organic brand set matched in 91% of pairs. The 9% that diverged fell inside the day-to-day drift we already observe between two ad-free runs of the same prompt, so we cannot attribute it to the ad unit.
That is a bounded finding, not a general clearance. It says: on the prompts we tracked, in the window we tracked, we could not detect paid influence on organic answer composition. It does not say advertising never shapes organic output on any surface, and it should be re-tested each quarter.
What this means for budget conversations: you cannot buy your way into an organic recommendation. Earning the organic slot stays a content, entity, and reputation problem — the same one behind why a brand goes invisible on ChatGPT in the first place.
What to do when a competitor buys the slot above your mention
First, confirm the geometry. A competitor ad below your organic recommendation is a different problem from a competitor ad above it, and a different problem again from a competitor ad on a prompt where you appear nowhere in the answer.
Then triage by prompt value, not by irritation. Rank affected prompts by commercial intent and volume, and check how the loss compounds across a multi-turn session rather than a single response — a competitor visible on turn one and turn four is a materially worse outcome than one flash on turn two. Our approach to scoring visibility across a whole conversation applies directly.
Three responses, in order of cost:
- Strengthen the organic answer. If the model already names you, deepen the evidence it draws on — comparison content, third-party reviews, and specification detail that survives summarisation.
- Contest the shortlist, not the ad. Paid units sit outside the answer. Your use is on which brands the model names inside it.
- Match the buy selectively. Only where the prompt is high-intent, the ad appears above the answer, and you can measure incremental clicks separately from earned visibility.
Rarely worth doing: treating every sponsored competitor impression as a crisis. In our panel, the median tracked brand faced a competitor ad on fewer than one in twelve responses.
The reporting template that survives scrutiny
A monthly AI visibility deck should make provenance visible on the first slide, not in an appendix. Minimum viable structure is four rows:
- Row one — earned AI share of voice (P0 + P1), with the session identity stated in the row label.
- Row two — paid AI presence (P2), yours and competitors', as a count of responses with at least one ad unit.
- Row three — blended responses (P3), with the control-run result attached.
- Row four — quarantine volume (P4), because a growing quarantine means your classifier is falling behind a platform change.
Add one line of method under the chart: prompt count, date range, assistants, session tier. Anyone who has had a CFO question a marketing number knows the method line is what ends the conversation early.
This structure travels. It is the same discipline you would apply to any channel where paid and earned inventory share a surface — AI assistants are simply the newest place the two sit two pixels apart, and where share of voice metrics get criticised for exactly this kind of blurring.
Frequently asked questions
Do ads in ChatGPT appear in API responses?
No. Ad units render in the consumer ChatGPT interface for eligible free-tier sessions only. API responses carry no ad inventory, which makes the API a usable ad-free control for paired testing — though API answers can differ from the consumer product in other ways, so pair on prompt and time, and treat the API as a control, not a substitute panel.
How do I turn off ads in ChatGPT?
Upgrade to a paid tier. Plus, Pro, Business, and Enterprise are ad-free, and that is currently the only reliable way to remove the units — there is no ad toggle inside Free or Go. OpenAI's help documentation covers the ad-personalisation controls that exist, which reduce targeting but do not remove inventory.
How do I tell a sponsored product card from an organic recommendation?
Check the container before the copy. Sponsored cards sit in a tinted, bounded block with an advertiser favicon and a click-through CTA, and carry a "Sponsored" or "Ad" tag at card or container level. Organic recommendations are plain answer text with, at most, a numbered citation. When both signals are absent or contradictory, quarantine it rather than guessing.
Should paid AI placements count in AI share of voice?
Not in the same number as earned mentions. Report them side by side. Blending the two makes the metric respond to budget changes rather than to how the model actually describes your brand, which destroys its usefulness as an answer engine optimization signal and makes quarter-over-quarter comparisons meaningless.
Which AI assistants currently show no ads at all?
At the time of writing, Claude and Grok have no announced ad inventory, the Gemini app is not commercial inventory, and Perplexity withdrew its ad products in February 2026. Google's ad surfaces are AI Overviews and AI Mode inside Search. Re-verify quarterly — this list has changed twice in six months.
Does an ad on a prompt mean my organic visibility dropped?
No. Ad presence and organic mention are independent in our data. A response can carry a competitor ad while still naming you in the answer body; that is a P3 blended response, and it should be reported as an earned mention with a flag, not as a loss.
Can advertisers pay to be mentioned inside the answer text?
Not through the ad product. Ad selection runs separately from answer generation, and in 214 paired runs we could not detect any difference in the organic brand set between ad-serving and ad-free sessions. Anyone selling "guaranteed ChatGPT recommendations" is selling something the ad platform does not offer.