Cognatu
Insights

The Ad Slot That Doesn't Exist

A conversation has no banner. Here is the format we think belongs there instead, and the constraint that makes it work.

By Shirish Garg  ·  September 8, 2026  ·  5 min read

Somewhere in an AI app this morning, a conversation about a leaking washing machine ended.

The answer was good. The model worked out it was probably the door seal, walked through the fix, and put the part at about twenty pounds. The person typed thanks, which they did not need to do, and closed the app.

Hold that last moment. A solved problem, a person who now needs a specific twenty-pound part, and an app that just did the work free of charge. In any other medium this is the most valuable second on the internet. In this one, nothing happens.

There is nowhere for anything to happen. Chat has no ad slot. Every attempt to create one so far has been a transplant from an older medium, and the body keeps rejecting it.

A guest from another medium

The banner shipped in 1994 and has been following us around ever since, from the page to the feed to the video pre-roll, always the same move: find the empty space, fill it. Readers answered with a skill of their own. By the late nineties researchers had a name for it, banner blindness, the learned ability to not see the thing that was paid to be seen1.

A conversation offers no empty space at all. It is a private exchange between two parties, and anything inserted into the middle of it is interrupting someone.

The industry is inserting anyway. Perplexity sells sponsored follow-up questions that sit inside the exchange2. Google places ads inside AI Overviews, in the body of the answer itself3. I understand the pull: during the conversation is where the attention is. But being present in a conversation is not the same as being welcome in it, and this medium punishes the difference harder than any page ever did.

One note, after, labelled, approved

We are proposing a convention for a medium that does not have one yet. Nobody has earned the right to pronounce here, us included, so here is the proposal stated plainly, with the reasoning attached and the exits marked.

One sponsored note. After the conversation ends. Labelled as an ad. Approved, creative by creative, by the publisher whose app it appears in.

For the washing machine: the model's last message stands untouched, and beneath it a small bordered card reads Sponsored, a named brand, the door seal for that model, eighteen pounds, fitting guide included. One card. Nothing else moved.

A conversation column ending in the model's final answer about a washing machine door seal. Below the final message, separated by a clear border, sits a single card labelled Sponsored: a brand name, the exact replacement part, a price, and a link to a fitting guide. Nothing inside the conversation itself is changed or highlighted.
The whole format. The border is load-bearing.

Why after, and not during

The moment an ad can appear inside an answer, every answer acquires a second possible motive. You do not need to believe the model is actually steering toward a sponsor. The doubt alone is the damage, and it lands on every answer, including the ones nobody paid for.

After the final turn, the arithmetic changes. The advice was complete before the ad existed, so there was nothing upstream to bend. An ad that arrives once the help is finished is a suggestion. The identical ad arriving during the help is a conflict of interest.

There is a harder version of this rule inside the system: whatever matches the ad must never reach backwards into the conversation. The model that answers you cannot know a sponsor is waiting, or the after is cosmetic. We think that firewall is the actual product, and the card is just its visible end.

A dark vertical wall labelled the firewall divides the picture. On the left, a conversation: a question and an answer from a model that knows no sponsor exists. One blue arrow passes left to right through a single gap in the wall, carrying only what the exchange was about, a category and nothing more, into a matcher that produces one labelled sponsored note. A return arrow from the matcher toward the conversation is stopped dead at the wall with a cross: nothing reaches back, and the model never knows an ad is waiting.
The card is the visible end. This wall is the product.

Why one, and not three

Attention after a solved problem is real but small. A postscript's worth. One note fits inside it; three notes is a webpage, and people already trained themselves on webpages1.

Scarcity is also what gives the slot a price. The token post argued that as intelligence gets cheap, intent is what stays scarce. A single slot per conversation is that scarcity made physical. Sell it once and it is worth bidding for.

Why labelled, and not blended

The FTC has required clear disclosure on native advertising since 20154, but compliance is the weak argument. A conversation works because the reader's guard is down. An ad announced as an ad gets judged on whether it is useful. One that has to be discovered gets judged on the deception, and so does the app that allowed it. Disguise does not borrow trust. It spends it, all of it, at once.

Why the publisher signs every creative

Because the app owns the relationship, not the network. A network that can inject arbitrary creatives into your product is the toll booth wearing a media kit. Publisher approval is slow, and it caps how fast anything scales. Both true. The alternative is an app betting its entire trust surface on a stranger's copywriting.

The cap is the point

Add the constraints up and the inventory is tiny: one note, only after the exchange, only where commercial intent actually surfaced. Measured in impressions, a rounding error next to display.

We are not planning to fix that. The banner's whole history is inventory expanding until price and attention collapsed together, formats getting louder as they got cheaper. This format runs the film backwards: few slots with real intent behind them. Fewer, worth more.

Two columns compared. On the left, the display model: an endless page studded with many ad slots, each cheap, most ignored. On the right, the conversational model: one conversation producing at most one labelled note, and only when the exchange itself surfaced commercial intent. The left column is dense and grey; the right is nearly empty with a single blue card.
The formats price differently because they count differently.

What would change our mind

Written down, so we can be held to it.

If clearly labelled after-notes get ignored at banner rates, the attention argument fails and the format is merely polite. If users of real apps read even one labelled note as a betrayal, the trust argument fails. And if the firewall proves impossible to keep, if any system that matches ads to conversations ends up bending conversations toward the money, then the honest conclusion is that conversations cannot carry advertising at all, and we would rather find that out than argue with it.

The constraint I am least sure of is one rather than zero for certain conversations. Health, children, grief. I do not have a principled line there yet, only a list of exchanges under which I would not sell a note.

What nobody has, us included, is pricing. There is no comp for what a conversation is worth, and every candidate anchor breaks in a different place. That is the next essay, and I can tell you now it will not end in a number.

Cognatu is building exactly the format described above, so treat me as an interested party making a proposal, not a neutral describing an inevitability.

Conventions outlive their inventors

The banner outlived nearly every publisher that ran it. Formats are the longest-lived decisions a medium makes, and they get made early, by whoever ships first and calls it normal.

Conversational advertising is a few years old at most. Its conventions are being set right now, in product launches rather than standards bodies, and mostly by companies whose incentive is to blur the line this post is drawing.

The first ad format of the conversational era is being decided now. We would like it to be one you can look at straight.

The on-device part is the whole argument

See what leaves the device

Our demo runs a real conversation, works out what it is about, and matches an ad without the text ever leaving the machine. Open your network tab and watch what crosses. The card it shows at the end is the format this post is arguing for. No account needed.

Watch it run Or read how the SDK works →

Next: An AI Operating System, or a Toll Booth. Who runs the layer that decides which model answers, and whose machine it should run on.

Sources

  1. Nielsen Norman Group, Banner Blindness: Old and New Findings. Two decades of eye-tracking showing readers reliably skip content in known ad positions and shapes.
  2. Perplexity, Why we're experimenting with advertising. Sponsored follow-up questions placed inside the answer flow.
  3. Google, Ads in AI Overviews. Search and Shopping ads shown within the AI-generated answer itself.
  4. FTC, Native Advertising: A Guide for Businesses. Disclosure must be clear and prominent enough that consumers recognise paid content as advertising.

Cognatu builds privacy-first infrastructure for AI apps · cognatu.com