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

The setup

We pulled a Google Search Console export for tabiji.ai. Sitewide CTR sat at 0.83%, masking a six-fold spread across page templates:

Our /compare/ and /popular-picks/ pages ranked in roughly the same band as /scams/ (positions 4–10) but converted at one-fifth the rate. Ranking wasn’t the problem. Searchers saw our listings on page one and clicked someone else.

We launched three controlled title tests — 20 pages on /scams/, 364 on /compare/, and 300 on /popular-picks/ — changing only the <title> tag and locking our evaluation window in advance.

What V3 won by

Round 1 on /scams/ pitted three title reframings against the control across 20 pages. Variant V3 won decisively:

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 a +1.93pp CTR lift in aggregate, improving every page in its cohort:

• 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)

Why it worked: those zero-CTR pages were already earning impressions for queries like “is Sofia safe for tourists” and “is Bucharest safe to visit.” Our original title used a standard listicle frame that ignored the searcher’s question. Putting the question directly in the title unlocked the clicks.

Meanwhile, our fear-hook variant — “Don’t Fall for These 6 Tourist Scams in Sofia” — leaned on loss aversion and an imperative verb. Classic copywriter instinct. It lost.

Titles that mirror the searcher’s natural query beat titles that describe the page’s contents. Most programmatic SEO pipelines generate content descriptions; searchers click query matches.

The 14× topic gap

Next, we segmented /popular-picks/ by venue category. The CTR spread was stark:

A 14× spread from top to bottom (bookshops at 1.83% vs. coffee shops at 0.13%) on the exact same template and domain authority. The four lowest-CTR topics generated huge impression volume and almost zero clicks. When we rank #5 for “best coffee shops in Tokyo,” we sit beneath Eater, TimeOut, and entrenched local blogs.

When we rank for “best bookshops in Almaty,” we are often the only dedicated guide on the SERP — and we capture the click.

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

Segmenting /compare/ by destination popularity showed the exact same curve:

At the same ranking position, niche-vs-niche pairs convert 3× better than big-vs-big pairs. Head-to-head matchups between famous destinations are crowded with TripAdvisor threads, Reddit megathreads, Lonely Planet, and Travel + Leisure. On niche pairs like Naoshima vs. Teshima, we own the intent.

Why this matters in the AI-search era

In 2026, head-term SEO is a losing trade. AI Overviews swallow the click on broad queries like “what is X” and “best X in Y.” When Gemini or ChatGPT synthesizes a one-paragraph answer at the top of the SERP, ranking #4 on a head term gets you scraped as a silent citation, not clicked (see our AEO playbook).

Long-tail niches are where the click survives. The more specific the tradeoff or local inventory, the less confident a generic AI summary looks, and the more likely the searcher clicks through to verify details.

Unit economics make the split even sharper. Enriched programmatic pages cost real money to build — Gemini API calls for synthesis, Nano Banana Pro for custom illustrations, MiniMax for audio, plus the verification passes detailed in Scaling with AI is Hard because AI is Lazy. Spending ~$25 to generate a head-term page that sits under an AI Overview at 0.13% CTR is a guaranteed loss. Spending that same $25 on “best bookshops in Almaty” at 1.83% CTR pays back.

• Head terms: High impressions, AI-Overview cannibalization, ~0.2% CTR. Negative ROI.
• Niches: Moderate impressions, weak SERP competition, ~1.5%+ CTR. 5–10× better unit economics.

It’s the SERP equivalent of Training Data Is the Moat: as AI commoditizes the broad middle, only specific, long-tail ground still yields margin.

Test on losers, hold out winners

One experimental rule saved us from muddying the data: run title experiments only on underperforming cohorts, and hold out your winners.

If you dump bookshops (1.83% CTR) and coffee shops (0.13% CTR) into a single sitewide A/B test, the aggregate average hides the signal. A title formula that lifts coffee shops to 0.4% while dragging bookshops down to 1.2% looks flat on paper while quietly damaging your best pages.

Isolate the losers, find the variant that fixes them, and leave the winners alone until you test them in their own cohort.

What’s still in flight

We have two larger rounds reading out at the end of May: 525 scam pages across 7 arms (scaling V3 against five new challengers) and 364 compare pages across 11 arms. I’ll publish the full readout once the window closes.

Until then, the 14× topic spread and 3× pair gap point to one rule: stop paying inference costs to fight for head terms. Build for the specific long tail where searchers still need to click.