Too many affiliate links? I put the toplist on trial

The short version: I operate comparison sites in two of the most heavily monetised corners of the internet, and I'd started to worry that our pages — long ranked lists, an affiliate button on every row — looked like sales brochures to Google and to AI engines. So I put the format on trial: Google's actual written policy record, the conversion-rate literature, a measured teardown of 18 pages that currently win the rankings we want, and AI citation data. The verdict surprised me. Nobody — not Google, not the AI engines, not the data — penalises the number of affiliate links on a page. What separates winners from casualties is whether each recommendation carries its own evidence, and whether the page resolves the reader's choice instead of just displaying it. Almost everything else the industry believes about list length turned out to be folklore.

If you run comparison content anywhere — software, credit cards, hosting, supplements — the anxiety is the same and so, I think, is the answer.

The fear, stated honestly

A typical "best X" page in a monetised vertical lists ten to twenty-five options. Each gets a review link and a visit button. Some pages render the list twice — once as a comparison table, once as cards. Count the calls-to-action and you can easily hit thirty on a single URL.

Stand back and squint, and it looks salesy. The instinctive fix is obvious: cut the list to one or two confident recommendations, drop the buttons into the prose, look less like a shop. I nearly started doing exactly that. Then I realised I was about to make a revenue-affecting structural decision off a vibe, and vibes are cheap to check.

How I ran the trial

Four lanes, run independently so they couldn't contaminate each other:

1. The policy record. Everything Google has actually published that touches affiliate pages: the spam policies, the Reviews System documentation, the quality-rater guidelines, the site-reputation-abuse policy, and on-record statements from Google's own people.
2. The conversion literature. What published data — not blog consensus — says about list length, choice overload, and format.
3. A measured teardown. I pulled the current top-ranking pages for ten commercial "best X" keywords across six markets and two verticals, fetched each one, and counted: list length, visible CTAs, formats, hero boxes, disclosures. Eighteen pages yielded full measurements. (Fifteen more blocked automated fetching — recorded as unmeasurable, not guessed.)
4. AI citation behaviour. What ChatGPT, Perplexity and Google's AI Overviews actually cite for these queries, cross-checked against my own citation-scan corpus.

Finding 1 — There is no link-count ceiling. Anywhere.

I expected to find at least a soft official signal against link-dense pages. There isn't one.

Google's link-spam policy says affiliate links are fine on one condition, and it's not a number: "It's not a violation of our policies to have such links as long as they are qualified with a rel="nofollow" or rel="sponsored" attribute value." Tag them properly and the count is your business.

The spam category that does exist is called thin affiliation, and it's about content, not links: pages "where the product descriptions and reviews are copied directly from the original merchant without any original content or added value." Google even lists the escape routes — original reviews, rigorous testing, ratings, comparisons. A three-item page of copied merchant blurbs violates it. A twenty-item page where every entry carries original, differentiated assessment doesn't.

The Reviews System documentation goes further and explicitly blesses the format: reviews can be "about a single thing, or head-to-head comparisons, or ranked lists of recommendations." The one demand it makes scales with list length, and this is the sentence I now treat as the actual law of the genre: when you call something the best, "include why you consider it the best, with first-hand supporting evidence." A twenty-item list owes twenty justified verdicts. That's the real cost of a long list — an evidence debt, not a policy risk.

Even the famous casualty supports this reading. HouseFresh, the air-purifier review site whose traffic collapse after the March 2024 core update became the genre's cautionary tale, didn't recover by stripping affiliate links. They recovered — to roughly four times their previous peak — by publishing more original testing evidence, not fewer buttons.

Finding 2 — The winners don't run shorter lists. They resolve the choice faster.

This was the teardown's job: not what should work, but what the pages currently winning actually do.

Median list length among winners: twenty items in the gambling vertical, ten in finance. Nobody at the top is running a two-item page. Visible CTAs clustered between eight and forty per page; one 71-item UK directory shipped roughly ninety-five and ranks fine. Even the pattern I was most suspicious of on my own sites — rendering the same list twice, table then cards — turned out to be standard practice: 72% of the winning pages do it.

But two patterns were nearly universal, and neither is about length:

• 94% open with a single #1 pick, highlighted above the fold, before any table. Even the 71-item directory spotlights one recommendation first.
• 67% show a three-to-five item shortlist — "quick picks", each tagged best for something — before the full list.

Read those together and the design philosophy of the whole genre snaps into focus. The winners don't reduce choice; they resolve it. The reader who wants an answer gets one in the first screen. The reader who wants the full market survey scrolls on and gets that too. The long list isn't the sales pitch — it's the reference section.

Finding 3 — The conversion "science" behind short lists is folklore

The strongest argument for cutting a toplist down was never Google — it was conversion psychology. Fewer choices, more decisions: everyone knows the jam study. Twenty-four jams on the table, 3% bought; six jams, 30% bought.

Here's what everyone skips: in 2010 a meta-analysis pooled fifty choice-overload experiments — five thousand participants — and found the average effect of more options on choice outcomes was essentially zero. Sometimes more options hurt. Sometimes they helped. The variable that actually moved outcomes was whether something resolved the choice — a clear recommendation, a default, a "best for you" signal. Which is precisely the structure the winning pages converged on. Practice quietly rediscovered what the research actually says, while the folklore kept citing the jam.

It gets worse for the received wisdom. The most load-bearing claim in affiliate circles — that comparison tables are the highest-earning element on a page — traces back, as far as I could follow it, to no published dataset at all. It's repeated most confidently by companies selling comparison-table plugins. And nowhere could I find a real split test of table versus inline links versus single-pick formats. The industry's most repeated conversion advice is unmeasured.

That's not a reason to despair; it's an opening. Anyone with a portfolio and click instrumentation can generate better data than the entire published literature currently offers. I'm now wiring position-level click tracking across my pages for exactly that reason.

Finding 4 — AI engines don't punish toplists. They cite them.

The newest version of the fear is that AI answers will treat monetised pages as radioactive. The observed behaviour is close to the opposite: when I look at what actually gets cited for commercial "best X" prompts — in my own scan corpus and in every third-party study I could find — long, structured, affiliate-monetised comparison pages are the dominant citation winners. The engines seem to like exactly what the format provides: ranked structure, concrete numbers, one page that maps a whole market.

Whether heavy CTA density specifically suppresses citation is — genuinely — unmeasured. I found no study on it and my own data can't isolate it yet. But two structural findings matter more. Cited pages share traits: specific figures over adjectives, structured tables, tight topical match to the question. And query shape has a ceiling built in: pages get cited for "review of X"-shaped questions at roughly three times the rate of "best X"-shaped ones. The unglamorous individual review pages hanging off your toplist may be more citable than the toplist itself. Deepen them.

What I'm actually changing

The trial ended with a build list, not a bonfire:

1. A #1 pick hero on every list page. The single strongest convergent pattern, and we didn't have it.
2. A three-item quick-picks row, each tagged best for a use case, before the full table. Resolve, then present.
3. Keep the long list — behind an evidence gate. An option earns a place (and a button) only when it carries a written verdict: why it sits where it sits, in original words, with something concrete behind it. Can't write the verdict? The item comes off. That's the only length discipline the evidence supports, and it trims lists honestly — from the quality side, not the fear side.
4. Keep the double render, but make the second one earn its place — the table carries specifications, the cards carry judgement. Two renders of identical buttons is noise; two layers of different information is architecture.
5. Instrument everything. Position-level click tracking, because the published literature is empty and whoever measures first, knows first.

What I'm explicitly not doing: cutting lists to one or two picks, stripping buttons to look humble, or treating "it feels salesy" as a strategy input. The fear was pointing at the wrong variable. It usually is.

Caveats, because they're the difference between data and content

Eighteen measured pages is a teardown, not a census — the biggest brands in both verticals block automated fetching, so the sample tilts toward sites that don't. Winner patterns are correlation; nobody A/B tested a hero box against its absence, including, yet, me. And AI citation behaviour is a weather report — the engines' source preferences have shifted before and will again.

But the direction of every lane agreed, and that's rare enough to act on: the risk was never the number of links. It's recommendations that can't justify themselves. Fix that, and the long list stops being a liability and becomes the moat — because writing twenty defensible verdicts is exactly the work the thin sites won't do.