What Running Our Own ChatGPT Ads Taught Us, and the Tool It Produced

ChatGPT ads became a self-serve product in Europe on August 31, 2026. A company registered in any of the 31 covered European markets can now open an OpenAI advertising account and buy placement without going through an agency, per Digiday's report on the European self-serve opening that same day. LightSpot did exactly that. The Ads tab, live in the app since September 6, 2026, came out of those campaigns.

This is a field report, not a launch post. ChatGPT Ads is less than a year old, its rules resemble no existing ad platform, and published advice goes stale in weeks. What follows is what changed our own practice, and what we built as a result.

Why buy placement when you sell organic visibility

LightSpot sells organic AI visibility, so buying a sponsored slot could read as an admission. It is not. Paid placement and organic citation are two separate surfaces inside the same answer, and one does not deliver the other. Testing that separation with real money was the point of the exercise.

The SE Ranking study published on August 10, 2026 analyzed 50,006 US commercial prompts collected on July 23, 2026 across 20 niches, and Search Engine Land's write-up of that study appeared the same day. Its most useful finding: in only 3.63% of served ads did the advertiser's own domain also appear among the sources cited in the answer above it (SE Ranking, July 2026 data). Paying does not buy the citation. In the overwhelming majority of cases, the brand paying and the brand ChatGPT cites are not the same brand.

A company that builds its product on that observation ought to test it with its own money. We ran the test on our own category questions, using the landing pages built for the ChatGPT ecommerce funnel. A landing page is simply the page an ad sends its clicks to. The placement itself, a labeled sponsored block below the answer, is set out in what ChatGPT advertising changes for brands.

Four things the campaigns taught us

Four lessons showed up often enough to be worth building into a product: cap the spend before activating, treat targeting as copywriting, fix the landing page before the ad, and run one campaign at a time. Each of them changed how the Ads tab behaves. Each is set out below with the source that confirms it.

Set the spend cap before you activate, not after

A ChatGPT Ads account will accept a campaign with no explicit daily ceiling. That is the most expensive trap on the platform, because spend does not behave like a classic search campaign. Search Engine Land's hands-on guide to running ChatGPT ads, published August 10, 2026, reports a 25 dollar minimum daily budget and costs per click between 2 and 5 dollars on US campaigns, with an underdelivery warning as soon as a max bid drops below 3 dollars (Search Engine Land, August 2026).

A click here costs what a click costs on an expensive channel, and a loose ceiling turns into spend before the first readable report. That is why the push screen shows the bid, the cap and the resulting number of clicks, and asks for an explicit acknowledgment that the daily budget is set at OpenAI.

Context hints carry the entire targeting

There are no keywords in ChatGPT Ads. Targeting happens at the ad group level through context hints, which are plain-language phrases describing the conversations where an ad belongs. OpenAI's help center page on creating ad groups calls them broad thematic signals rather than exact-match rules. It recommends splitting genuinely different products or needs into separate ad groups instead of pooling them into one list.

Media buyers find that guidance thin. Scott Ensign, chief strategy officer at the agency Butler/Till, told AdExchanger's September 3, 2026 report on what advertisers want from the platform that it "gives you essentially no instructions on how to do that other than telling you to not just enter keywords."

This is the part of the work that feels least like media buying and most like copywriting. A line such as "why does ChatGPT recommend my competitors and not me" describes a situation. "SEO audit" describes a category. The first one works, the second one dilutes.

The landing page outweighs the ad

The ad format leaves little room. OpenAI's help center page on creating ads caps the title at 50 characters with a recommended range of 16 to 24, and the body copy at 100 characters with a recommendation of 32 to 48. At that length, the gap between two well-written ads is narrow. The gap between two landing pages is not.

There is also a technical constraint many sites discover only when an ad gets rejected. OpenAI's advertiser guidance on allowing its web crawlers explains that the destination is validated by OAI-AdsBot, the robot that fetches and reviews an advertiser's landing page. A page it cannot reach is not eligible to run.

A strict robots.txt file, the text file at a site's root telling crawlers which paths they may fetch, is enough to block an entire campaign. So is a web application firewall or a bot-mitigation layer. Our agents already check AI crawler access during an AI visibility audit. The ads crawler deserves the same check, and it is cheaper to run before spending than after a rejection.

One campaign at a time, or you read nothing

This is the structural limit of the channel. OpenAI returns no prompt-level data: an advertiser sees impressions, clicks, spend and conversions, but never the conversations that triggered the ad. Search Engine Land's August 2026 guide also notes that reporting cannot be segmented below country level, even when targeting goes finer.

At that granularity, two campaigns launched together on adjacent messages produce an unreadable result. The rule that follows holds for any advertiser: one live campaign per hypothesis, and one variable changed between iterations.

What the Ads tab actually does

The Ads tab sits in the AI Visibility section of LightSpot, next to Who gets cited. It answers two questions for a site owner: which brands pay to appear on the questions that matter in their category, and what to put up against that. It records, it drafts, and it pushes paused.

See who pays. During the weekly probe, our agents record the sponsored cards actually rendered by ChatGPT on the customer's category questions. Each observation keeps the advertiser, the ad title, the landing page, how often it showed, and whether the customer is cited organically on that same question. The row that matters is the one where a rival pays and the customer appears nowhere.

Draft the campaign. From the selected questions, our agents write context hints, a title, body copy and a destination page picked from pages of the site already audited. The draft draws on the site audit and on the rival ads observed on those questions. Everything is editable, and the character counters mirror OpenAI's own limits.

Push it paused. Once the Ads Manager account is connected, the draft lands in the customer's account, paused, after a confirmation step. Performance readings then flow back into the tab, alongside the organic citation rate on those same questions.

The split below is the money boundary, written out setting by setting.

The limits we set, and why

Three limits are deliberate. The Ads tab never activates a campaign, never replaces a media buyer, and never claims the channel pays off. Each one exists because the alternative would either move the money boundary onto LightSpot or promise a result that no published data supports in 2026.

The tab never activates a campaign. No line of code calls activation, and that is not a temporary safeguard: the money boundary stays with the customer. A tool that writes, pushes and spends in the same place asks for a level of trust nobody should hand to a three-click integration.

It is also not a replacement for a media buyer. In AdExchanger's September 3, 2026 report on what advertisers want from the platform, practitioners describe a still-rudimentary platform, with no multi-account management for agencies and no competitive data to benchmark against. AdExchanger reports, citing Scott Ensign, that the platform is particularly challenging for agencies and large enterprises, with no way to manage multiple advertisers in a single place. We do not fix that gap. We partly fix the second one, since recording sponsored cards is the only way to see rival activity while OpenAI publishes no public ad library.

Finally, it makes no claim that the channel pays off. In Search Engine Land's hands-on guide to running ChatGPT ads, John Horn compares roughly 1,500 users on each side, Google Ads traffic against ChatGPT Ads traffic for the same US advertiser: 41 seconds of average engagement versus 17, and a conversion rate of 3.71% versus 0.21% (Search Engine Land, August 2026). One advertiser, one vertical, but a signal it would be dishonest to bury.

The channel is young, inventory is moving, and nobody has stable benchmarks yet.

What to take away

Five things carry over to any advertiser testing ChatGPT Ads in 2026, whatever the category. They concern the gap between paid and organic, the spend ceiling, how targeting is written, what the landing page has to allow, and how many campaigns can run at once without making the results unreadable.

• Paying does not buy the citation. Across 50,006 US commercial prompts analyzed by SE Ranking in July 2026, the advertiser's domain appeared among the sources cited above its own ad in only 3.63% of served ads. The two surfaces are separate work.
• Decide the spend cap before activation. An unbounded daily budget is the channel's first risk, ahead of ad quality, at 2 to 5 dollars per click on US campaigns (Search Engine Land, August 2026).
• Targeting is written, not selected. Context hints describe conversational situations, not product categories.
• Your landing page has to be reachable by OpenAI's crawler. A page OAI-AdsBot cannot fetch is a rejected campaign.
• Run one campaign at a time. With no prompt-level data, two simultaneous tests cannot be told apart.

The Ads tab is included in Eclat and Zenith at no extra cost. Accounts on the Free plan see a locked preview. Details are on the pricing page.