Two products. One playbook. One tripled its share of AI answers. The other stayed at zero.
The difference was measurable on day one — before writing a single article.
Everyone is selling "GEO" (Generative Engine Optimization) right now. Almost nobody selling it has measured whether it works — because measuring it honestly means risking the answer "it doesn't."
I measured it. On my own money, my own products, for three months. Same operator, same content pipeline, same publishing cadence, same article formats. Two different products. Here's what happened.
The setup
Product A (ShareRec) — a screen-recorder-with-instant-link tool. Niche: screen recording / Loom alternatives.
Product B (RainVoice) — an AI phone receptionist for service businesses. Niche: AI receptionists.
Method: an automated pipeline generating citation-optimized articles, every one fact-audited by hand before publish.
Measurement: a tracker that asks ChatGPT and Perplexity 10 buyer-intent questions per niche and records every URL each engine cites. Run monthly since May 2026.
The metric that matters is answer-slot share: of all the citation slots the AI engines hand out across your niche's buyer questions, how many does your domain win?
The results
Tracker run
ShareRec share
RainVoice share
May 3
3% (5/157)
0%
May 26
6%
1%
Jun 15
7%
1%
Jul 20
8%
1%
Jul 29
13% (19/141)
0%
Three months in, ShareRec is the single most-cited domain in its own niche — ahead of every competitor, review site, and YouTube video. 20 of its 36 articles have earned at least one AI citation. Perplexity doesn't just link it; it quotes the product's feature list verbatim.
RainVoice — same pipeline, same care, 21 live articles — never cracked 1%, and decayed back to zero.
The one metric that predicted it
Count the distinct domains competing for citation slots in each niche's AI answers:
ShareRec's niche: 69 competing domains. Fragmented — small tools, indie blogs, a few review sites.
RainVoice's niche: 114 competing domains, led by Reddit threads, funded startups, and enterprise vendors, with dozens of near-identical products all publishing the same content.
In a fragmented niche, a consistent publisher compounds: 3% → 13% in three months. In a saturated one, the same effort earns 0% forever. And the saturation number was knowable before a single article was written.
What didn't matter (this surprised me)
Content format. Of the 20 pages that earned citations: 7 "alternatives" listicles, 5 use-case pages, 3 how-tos, 2 feature pages, 1 head-to-head comparison. Every format works.
Word count. Cited pages span 1,500–2,800 words. A substance floor, not a magic length.
Page-level targeting. In half of all citation events (32 of 62), the engine cited a different page than the one built for that exact query. You don't rank a page — you build a topical cluster and the engine picks its favorite.
What did matter
Niche winnability — the competing-domain count. Fragmented niches are winnable; saturated ones are not.
Cluster depth — 36 interlinked articles. Engines cite the domain that looks like the specialist.
Consistency — share compounded month over month; no single breakthrough article.
Factual accuracy — every claim hand-verified against competitors' live pages. AI engines learn not to cite pages with wrong facts.
The honest caveats
Citations are noisy. The two engines disagree on the same page 36% of the time. One measurement means nothing; the trend is the signal.
This is n=2 products and 2 engines (ChatGPT + Perplexity). A strongly validated hypothesis, not a law of nature.
Nobody can promise a specific citation. Anyone who does is lying. What's provable: in a winnable niche, consistent accurate publishing grows your share — and winnability is measurable up front.
Want these two numbers for your niche?
Your current answer-slot share, and the competing-domain count that tells you where you can win — before you spend a dollar on content.