An ad in a list is not an ad in an answer

In the eight weeks between mid-December 2025 and mid-February 2026, the AI industry answered the same question three different ways.

On 16 December 2025, Meta began using conversations with Meta AI to personalise advertising across Facebook, Instagram, WhatsApp and Messenger — announced the previous October, with no opt-out for people who use the assistant, and with the change excluded in the EU, the UK and South Korea.

On 16 January 2026, OpenAI began testing advertising in ChatGPT for logged-in adults on its free and lower-priced tiers, rolling out from 9 February. The units are contextual, labelled, placed beneath the response, and matched using the current conversation, past chats and memory. Paid tiers stayed ad-free.

In that same February, Perplexity discontinued the sponsored placements it had been testing since 2024. Its executives gave the reason plainly: a user needs to believe this is the best possible answer.

Two companies coupled commercial interest to an answer engine. One decoupled it. Inside eight weeks. That is not an industry converging on a norm — it is an industry that has not decided, deciding in public, on live products.

It is worth being precise about what the disagreement is actually about, because it is not “are ads bad.”

A list and an answer are different objects

Search advertising solved its trust problem decades ago with a label and a position, and it worked. It worked for a structural reason that is easy to miss: a results page is a list of places to go. The unit of the product is a link. An advertisement in a list is one item among many, and a label tells you exactly which item to discount. You can route around it. The rest of the list is untouched.

An answer is not a list. It is a single synthesised object with no visible seams. There are no items to compare and nowhere to route to. When you place a commercial interest beside an answer and label it, you have told the reader that money is in the room. You have not told them — and structurally cannot tell them — where in the answer it went.

That is the whole disagreement. It is a design question before it is an ethical one.

The label is not free

The instinctive fix is disclosure, and disclosure is genuinely better than concealment. But there is a finding that complicates it, and it points the wrong way for anyone who assumed labelling settles the matter.

Research on AI disclosure in advertising finds that labelling content as AI-generated raises what the literature calls persuasion knowledge — the reader’s active awareness that someone is trying to persuade them — and that this decreases trust in both the advertisement and the organisation behind it, even when the content is otherwise identical.

Read that as a design result rather than a marketing one. The label works. It does exactly what it is supposed to do: it switches the reader from receiving to evaluating.

Now apply it to a product whose entire value is that the reader stopped evaluating. An answer engine is worth using because you granted it your reliance — because you no longer open five tabs and cross-check. A label that reactivates checking is not a neutral disclosure. It is a partial withdrawal of the thing the product sells.

Which is why Perplexity’s stated reason is more interesting than a moral objection. The argument was not that advertising is wrong. It was that even clearly labelled advertising risks making users suspicious of everything — that the contamination is not of the advertisement, but of the answer around it.

What Google did is the tell

If you want evidence that these really are two different surfaces, look at the company with more experience of search advertising than anyone.

At the time of writing, the Gemini app carries no advertising. Google’s AI Mode and AI Overviews in Search do — labelled sponsored results, with new conversational formats announced at Google Marketing Live 2026 pairing an ad with a Gemini-written “AI explainer,” also labelled.

One company, two surfaces, two positions. That is not indecision. It is a company that understands search advertising extremely well drawing a line between a search surface, where a sponsored unit sits in a list of destinations, and an assistant surface, where it would sit inside a relationship. Whatever the internal reasoning, the revealed judgement is that these are not the same product.

Three honest complications

Advertising pays for free access. An assistant funded only by subscription is an assistant for people who can afford one. OpenAI’s structure — ads on the free tiers, none on the paid — states fairly directly that the ad-free experience is the premium one, which is honest and also uncomfortable. The ad-free position is not costless; someone pays, and it is not the advertiser.

“No ads” is not “no incentives.” An assistant with no advertising can still steer toward its own first-party products, its own subscription tiers, its parent’s services or its partners. Removing the ad unit removes one visible interest and leaves every invisible one in place. A product with no ads and an undisclosed house preference is in a worse position than a labelled ad, not a better one.

Mitigations are real. Meta’s stated exclusion of religion, sexual orientation, political views, health, ethnic origin, philosophical belief and union membership from personalisation is not nothing. Neither is a label, or physical separation between a unit and an answer. The argument here is that these are floors rather than solutions.

What the split is actually about

Attention arrives one of two ways. Either it was taken — won in a contest against everything else competing for the same eye — or it was granted, handed over deliberately by someone who decided to stop checking.

Search runs almost entirely on taken attention, which is why the label works there. Every result is competing, the user knows it, and their guard is up by default. An advertisement is one more competitor in a contest the reader already knows they are in.

An answer engine runs on granted attention. Its value is precisely that the contest ended: the user stopped opening tabs, stopped weighing sources, stopped checking. That surrender is the product.

So the same mechanic behaves differently in the two places. In search, an advertisement spends attention that was never granted. In an answer, it spends something the reader handed over — and unlike a click, a grant cannot be partially withdrawn. You either trust the answer or you go back to opening five tabs.

The industry did not split in early 2026 over whether advertising is acceptable. It split over whether an answer engine is a search engine with better formatting, or a different kind of object that has to be financed differently. Two companies bet the first. One bet the second. Google, revealingly, bet both.

Three things to watch

The way to hold this without picking a team is to know what would settle it. All three are observable from outside.

One: does the answer change, or only the page? A unit beneath an answer is a placement. An answer that differs because of a commercial relationship is a different thing entirely, and the distinction is testable by anyone with two accounts and some patience.

Two: does disclosure arrive where the decision is made? A policy page is disclosure in the legal sense and invisible in the decisional one. The question is whether the interest is visible at the moment the recommendation lands — which is criterion 6 on the scorecard, and the difference between scoring 2 and scoring 4.

Three: does anyone publish a withdrawal number? Every company in this story publishes adoption. If an ad-supported tier is not costing trust, the cheapest way to demonstrate it would be to publish the rate at which people mute, disable, downgrade or leave after a unit appears — attributed to the change that preceded it. Nobody does.

The third is the one to watch, because it costs almost nothing to produce and nobody produces it. The metrics that get published are the ones that go up.

Sources

Dates and positions as reported at the time of writing, 31 August 2026. Products in this category change quickly; check the current state before relying on any of it.

• → Meta’s use of AI conversations for ad personalisation from 16 December 2025 — CNN Business, 1 October 2025
• → OpenAI’s advertising in ChatGPT, dates and mechanics — the company’s own page at openai.com, and secondary reporting for the rollout timeline
• → Perplexity discontinuing advertising, February 2026, with the executives’ stated reasoning — Search Engine Land, 18 February 2026
• → Ads in Google AI Mode and the formats announced at Google Marketing Live 2026 — Browser Media; on the Gemini app’s position, Tom’s Guide
• → On AI disclosure raising persuasion knowledge and lowering trust in the advertisement and the organisation — Journal of Interactive Advertising (2025)