Is affiliate marketing dead now AI answers the question?

No — but the payout moved. Sites that rented their rankings from a search engine are losing. Sources that a model cites are not. The difference is first-hand information: prices you checked yourself, the times you say don't buy this, and a number with a source next to it. That is hard to generate and easy to verify.

This article is about which half of affiliate marketing died, and what the surviving half needs.

What actually changed?

For twenty years the deal was this: you wrote something useful, a search engine ranked it, and a share of ten blue links became visits. You did not own the traffic. You rented it, and the rent was content.

An answer engine changes the shape of that. Ten results collapse into one answer with a handful of citations underneath. The number of pages that get shown does not shrink a bit — it shrinks a lot, and it concentrates.

Which means the middle disappeared. Not the top, not the bottom: the middle. All those pages that summarised what was already written elsewhere — "10 best AI tools", the same paragraph about each, an affiliate link — existed because a search engine needed something to rank. A model does that job instantly, for free, without an affiliate link. There is no version of the internet where those pages come back.

Why do some sites lose and others don't?

Because a model can rephrase anything, but it cannot check anything.

It cannot open a billing page and see that the price changed last month. It cannot use a tool for six weeks and notice it gets slow. It cannot tell you that the second-best option is the one you should buy because the best one is a nightmare to cancel. All of that is observation, and observation has to come from somewhere.

That somewhere is what gets cited. A model building an answer wants a source with a specific, checkable claim attached to it — a price, a date, a measurement, a stated limitation. Generic copy provides nothing to attach a citation to.

This is also why the commission stopped being the problem people assume. The pages that survive are the ones that say "skip this one". Hedging honestly — this is good for X and bad for Y — is exactly the shape of a sentence that gets quoted, because it is the shape of a sentence that is true.

Does anyone still click through?

Fewer people. Honestly, I cannot tell you by how much, and neither can most of the people publishing numbers about it.

What is structurally true: when the answer is complete, the click does not need to happen. When the answer is partial — a price the reader wants to verify, a comparison they want to see laid out, a decision where the stakes are real — it still does. So volume falls and intent rises, and those two move in opposite directions on your revenue.

What that means practically: a thousand visitors from a list of ten results and a hundred from an AI citation are not comparable numbers, and treating them as the same metric will lead you to the wrong conclusion about what is working.

What kind of page still gets clicked?

Four kinds, and they all sit late in a decision rather than early.

Notice what is missing: the broad introductory article. "What is the best AI tool" is now answered above the results, and there is nothing left for you at the bottom of that page.

Notice also that all four need maintenance. A pricing page that is two years old is worse than no pricing page, because it is confidently wrong — and being confidently wrong once is how a source stops being cited.

How do you actually get cited?

Two halves: being fetchable, and being quotable. The first is technical and takes an afternoon. The second is editorial and never stops.

Fetchable means an AI crawler can get your words:

• Serve real HTML. Most AI crawlers do not run JavaScript, so a page that assembles itself in the browser is, to them, an empty page.
• Let them in. GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and the rest read robots.txt — decide deliberately, then write it down.
• Test it rather than assume it. Fetch your own page with a crawler's user agent and count the words you get back. I do this on this site, and the answer should be "all of them".

Quotable means there is something to lift:

• The answer in the first 40–60 words, before the anecdote. This is the part that becomes the passage.
• Headings that are the question someone typed, not a clever line.
• Numbers in a table.
• A source and a date next to any figure that is not yours.
• A visible date for when the page was last checked.
• An about page that says who you are and how you earn, because a model weighing whether to trust a source is looking for exactly that.

None of it guarantees a citation. All of it is required for one.

Does this change which programmes are worth joining?

It changes what you check before joining. Two things matter more than the headline rate:

The attribution window and how it is measured. Programmes vary enormously and change their terms without much notice. Read the current terms of each one yourself rather than trusting any article, this one included — a rate quoted in a blog post is a rate that was true on some unstated date.

Whether the product survives the visit. Traffic that arrives from an AI citation is further along and more decisive, which cuts both ways: fewer tyre-kickers, and less room for a product that does not deliver. Recommending something mediocre used to cost you a refund. Now it costs you the citation.

Is it still worth starting?

Only if you have something first-hand to put in it. That is the whole test.

If your plan is to restate what is already online, you are competing with a system that does that instantly and for free, and you will lose — not slowly, and not unfairly. If your plan is built on tools you actually use, prices you actually checked, and recommendations you are willing to defend to someone who disagrees, that is a different business and it is not crowded.

The uncomfortable part is that this version takes longer to pay. A summary site could be built in a weekend. A source takes months, because being a source means having checked things, and checking things takes time.

What I do not know yet

This site is the experiment, and it is early. It is built the way this article describes — server-rendered pages, AI crawlers allowed, the answer at the top, prices with dates, 27 of the 68 tools on the list paying me nothing. The search console has only just started collecting.

So I cannot yet tell you what it produced. When there are numbers, they will go on this site with the date attached, including if they are bad. That is the only version of this article worth writing, and it is also the point: an article that admits what it has not measured is the kind a model can safely quote.