The 14× CTR Gap: Why Niche Beat Head on 1,200 Pages

The setup

We started with a Google Search Console export from tabiji.ai. Sitewide CTR was 0.83% — but that average hid a six-fold spread:

Compare and popular-picks were ranking at similar positions to scams (most pages 4–10) but converting at one-fifth to one-sixth the rate. Position wasn’t the problem. Searchers were seeing the pages in the SERP and choosing something else.

So we shipped three structured CTR experiments — 20 pages on /scams/, 364 on /compare/, 300 on /popular-picks/ — with one intervention each (title only, everything else held constant) and decision rules locked in advance.

What V3 won by

The /scams/ round-1 cohort was 20 pages across 4 title variants. Three reframings against the existing baseline. The winner was V3:

Old: 6 Tourist Scams in Sofia (2026) — Real Stories & How to Avoid Them | tabiji.ai
V3: Is Sofia Safe for Tourists? 6 Scams to Avoid (2026) | tabiji.ai

V3 delivered +1.93pp CTR lift in aggregate, with every measured page improving:

• Kathmandu: 3.85% → 7.61% (+3.76pp on 184 impressions)
• Sofia: 0.00% → 2.07% (+2.07pp on 290 impressions)
• Bucharest: 0.46% → 1.09% (+0.63pp on 367 impressions)

The hypothesis turned out load-bearing: those zero-CTR pages were already pulling impressions for queries like “is Sofia safe for tourists” and “is Bucharest safe to visit”. The original title — a generic listicle frame — didn’t match the question form at all. Once the title answered the actual question, the impressions converted.

The losing variant — “Don’t Fall for These 6 Tourist Scams in Sofia” — used loss aversion plus an imperative verb. It’s the kind of framing marketing writers reach for. It lost.

Portable lesson:

Titles that match the natural query phrasing beat titles that describe content. Searchers are typing questions; pages that answer those questions in the title get clicked.

This sounds obvious. It’s not what most listicle generators produce — they optimize for content description, not query match.

The 14× topic gap

After V3, we segmented the popular-picks section by topic. The CTR spread was vertiginous:

14× spread, top to bottom. Same site, same ranking algorithm, same template. The four worst topics were also among the highest by impression volume — massive impression haul, almost no clicks. Pages ranking for “best coffee shops in {city}” were getting beaten in the click by every authority listicle competitor: Eater, TimeOut, the city guide blogs.

Bookshops in Almaty? We’re often the only quality content on the SERP, and CTR shows it. Coffee shops in Tokyo? Five louder competitors, all of them with more domain authority.

Niche-vs-niche beats big-vs-big at the same rank

The /compare/ section repeated the same structural pattern. Before designing variants, we segmented by destination fame:

Big-vs-big pairs lose 3× harder than niche-vs-niche pairs at the same ranking position. Famous-destination queries are owned by TripAdvisor, Reddit, Lonely Planet, Travel + Leisure. Tabiji ranks but loses the click. Niche pairs win because we’re often the only quality content the searcher can find.

Why this matters in the AI-search era

If you’re shipping content in 2026, the head-term game is structurally lost. AI Overviews ate the click for “what is X” and “best X in Y”. When ChatGPT or Gemini gives a one-paragraph answer with three citations, the head-term page that ranks #4 doesn’t get clicked — it gets summarized. Authority sites win those scraps; everyone else gets nothing. The AEO playbook we wrote earlier covers this surface in more depth.

What’s left is niches. The longer-tail the search, the less likely an AI Overview has a confident answer to summarize, and the more likely the user clicks through to read the actual page.

The production-cost side makes this worse for head terms and better for niches. AI content costs tokens. Tabiji’s data-enriched pages cost real money to generate — Gemini for synthesis, Nano Banana Pro for art, MiniMax for music, plus the cleanup overhead documented in Scaling with AI is Hard because AI is Lazy. If you spend ~$25 per page and ship it onto a head term where AI Overviews own the click, you’ve bought a permanent loser. Spend the same $25 on “best bookshops in Almaty” and the page actually converts.

The leverage equation has flipped:

• Head terms: big impressions, AI-owned click, low CTR. Negative ROI on AI-generated content.
• Niches: smaller impressions, less AI-Overview competition, high CTR. Positive ROI — sometimes 5–10× better.

This is the same structural argument as Training Data is the Moat: as AI eats the consumption surface, the differentiated long-tail is the only place that still earns a click — and the only place a brand can plant a flag the AI layer can’t immediately commoditize.

Test on losers, hold out winners

The single most important methodology choice we made: scope tests to underperforming subsets where titles are clearly failing. Don’t test variants across the whole section.

Put bookshops (1.83% CTR) and coffee-shops (0.13% CTR) in the same cohort and you’ll measure an aggregate. The aggregate hides everything. A “winning” variant might help coffee-shops while hurting bookshops — and on net, you’ve made the section worse.

Test where you’re losing. Let the winners keep winning. Decide separately whether to extend any variant to the winners.

What’s still in flight

Two larger experiments are live and review at the end of May: 525 scam pages across 7 arms (scaling V3 plus five fresh variants), and 364 compare pages across 11 arms (testing seven different framing families). I’ll publish the deltas once they’re in.

But the V3 result, the 14× topic-level spread, and the niche-vs-head pair structure are already enough to ship the conclusion: in 2026, niche is the only surface where AI-generated content still pencils out. Stop writing for head terms. Write for the long tail nobody else is good at.