I shipped 24 comparison pages in 11 days. Here's the honest CTR.
In May 2026 I shipped 24 comparison pages for Capi, my Telegram money tracker: eight rivals, three languages, eleven days. Seven weeks later those pages have had 26 visits, at most three of them human, and zero signups, while the wider site logged its first referral from an AI answer engine. This is the full data, why every page stays up anyway, and what I would do differently.
I write posts like this because the marketing content I learned the most from was never the case study with the up-and-to-the-right chart. It was the rare post where a founder published the numbers that did not work. This is one of those. I build Capi, so I have an obvious interest in you eventually trying it, but nothing below depends on you believing me: the method is described well enough to rerun on your own site, and the numbers come straight from my self-hosted analytics database.
What did 24 comparison pages get me in seven weeks?
Twenty-nine pageviews and 26 unique sessions across all 24 pages, measured from launch on May 19 through July 10, 2026. Four of those views were me, checking my own work on ship night. One session the next day looked like a real reader. Twenty-two were a datacenter crawler. Two more might be human. Search engines referred exactly zero visitors to any comparison page.
Some context so you can judge the scale. Capi's site does about 200 pageviews a month in total, mostly to the homepage and a groceries budgeting post that quietly picks up search clicks. Against that baseline, 24 new pages producing effectively zero human traffic in seven weeks is not a rounding error. It is the result.
The pages themselves are the standard playbook executed carefully: a hub plus one page per rival, Capi vs YNAB being the template, then Monarch, Copilot Money, Rocket Money, Moneko, Cointry, Zenmoney and a Mint-replacement page, each in English, Portuguese and Spanish. Question-format headings, extractive answers, real comparison tables, five to six JSON-LD blocks per page. If the format has a citation ceiling, I did not leave much of it unused.
Why did I build comparison pages at all?
Because AI answer engines cite them. When someone asks ChatGPT or Perplexity for "alternatives to YNAB," the engines pull disproportionately from pages structured exactly like these: an X-vs-Y URL, a semantic comparison table, a direct answer in the first paragraph. I was copying a pattern that demonstrably works for others.
The specific inspiration was Meet Lea, a LinkedIn tool whose comparison hub gets cited by name in LLM answers for its whole category. The supporting research was public: SE Ranking's ChatGPT study found domains with profiles on review platforms are roughly three times more likely to be picked as sources, and a study of 18,000 verified ChatGPT citations found 44.2% of citations come from the first 30% of a page's content, which is why every page fronts its answer. The bet was reasonable. The bet is still open. What follows is just what the first seven weeks actually look like from inside, which the case studies never show you.
For the record, the 11 days cost less than the title implies. The English eight went live on May 19, Portuguese on May 29, Spanish on May 30: one chassis, one evening of real writing per batch, and a deploy script doing the rest. In practice it was three evenings spread across 11 days, which matters later when we talk about what those evenings should have bought instead.
Who actually visited the 24 pages?
Twenty-six sessions, and I can account for nearly every one. I was the first visitor, on ship night, from Brazil. One person from Thailand arrived the next day. Then came 23 consecutive days of zero visits, followed by a ten-day wave of 22 single-page sessions from Singapore that was almost certainly a crawler, and two possible humans from the United States and India.
| When | Who | Sessions | My read |
|---|---|---|---|
| May 19 | Brazil, desktop, 4 pages | 1 | Me, admiring my own work |
| May 20 | Thailand, desktop, 1 page | 1 | Probably a real human |
| May 21 to June 12 | Nobody | 0 | 23 days of silence |
| June 13 to 23 | Singapore, desktop, 1 page each | 22 | Datacenter crawler rendering JS |
| June 22 | United States, desktop, 1 page | 1 | Possibly human |
| July 10 | India, desktop, 1 page | 1 | Possibly human |
The Singapore wave deserves a closer look, because it is the most interesting thing in the table. Twenty-two sessions over ten days, every one a desktop browser from the same country, every one viewing exactly one comparison page and leaving, no referrer, methodically covering the cluster. Humans do not browse like that. Headless crawlers that execute JavaScript do, and my analytics only fires on rendered pages, so ordinary bots never show up in these numbers at all. Something with a rendering budget walked the entire cluster three and a half weeks after launch. I cannot prove which AI company's crawler it was, and I am not going to pretend otherwise. But if you build pages for machines, machine visits are the leading indicator you get.
So what is the honest CTR?
There is no honest CTR to report, because there are almost no impressions to divide by. Over the last 30 days the entire site, not just the comparison cluster, received nine referrals from search and AI engines: two from Google, four from Bing, one from DuckDuckGo, one from Yahoo, and one from Microsoft Copilot. The comparison pages' own referrer column is empty. Any ratio computed from numbers this small is noise wearing a percentage sign.
That Copilot referral is my favorite number in this entire dataset. One human asked an AI assistant something about money trackers, the assistant answered with a link to my site, and the human clicked it. That is the entire strategy working end to end, exactly once. It is simultaneously proof the mechanism exists and a reminder of the scale I am actually at: one. I wrote about this gap between what founders imply and what dashboards show in confessions of a finance app builder, and this cluster is the cleanest example I own.
Why keep 24 pages that humans do not read?
Because they cost nothing to keep and their audience was never this month's humans. The pages are static HTML on a server I already pay for: zero maintenance, zero marginal cost. They are retrieval infrastructure for answer engines, and the early signals, a full crawler pass in June and a first AI referral in week seven, are the ones this strategy predicts before any human traffic arrives.
I want to be careful here, because "it will work later" is what every failed content strategy says right up until the end. So let me separate what I know from what I hope. I know the pages got crawled and rendered. I know one AI engine has already referred one human. I know the format matches what the citation research rewards. What I hope, and cannot yet verify, is that the pages are being quoted inside answers I cannot see, and that citations compound as the engines refresh their indexes. The 3x aggregator statistic is a correlation from someone else's dataset, not a promise about mine. If by November the cluster still has zero search impressions and no further AI referrals, the correct move is to admit the bet failed, and I will publish that follow-up too.
Meanwhile the pages do a second job: they give the rest of the site somewhere to point. Every blog post in the cluster, including the pillar ranking, links into the comparison pages, so each new post adds another path in. The referrer data cannot yet show a reader taking one of those paths, and I am not going to claim one. But the paths are cheap, and they compound with every post.
What would I do differently?
Distribution first, pages second. My site had, and still has, more product than audience by two orders of magnitude: dozens of carefully structured pages and roughly seven visitors a day. Building 24 more pages did nothing to change that ratio, because pages are not a channel. If I were starting again, every page would ship paired with one act of distribution aimed at actual humans.
Concretely, that means the boring things I postponed while enjoying the craft of templating: honest posts in communities where budget-app people already argue, directory listings, replying to every "YNAB alternative?" thread with something useful, and essays like this one, which exist precisely because a post about failure travels further than a comparison table. The daily logging habit that powers Capi itself taught me the same lesson from the other direction: what I track every day matters more than what I set up once. Tracking works when the loop is fed daily. Audiences, it turns out, work the same way.
The cluster was not a waste. It is a warehouse built a season before the road that reaches it. My mistake was filing those three evenings under growth when they belonged under inventory.
The tracker behind all these pages.
Capi lives in Telegram: log expenses in any currency by voice, text or receipt photo, and see your whole month in one place.
Free for 30 transactions a month, no card required.
Questions founders ask about comparison pages
Do comparison pages still work for SEO in 2026?
For classic search, slowly and only with authority: a new domain should expect months of near-zero impressions, which matches my data. The stronger 2026 case is answer engines. Research on ChatGPT citations shows comparison-shaped content with real tables gets cited far more than prose, and my site's first recorded AI referral, a click from Microsoft Copilot, arrived within the first two months.
How long until AI engines cite a new comparison page?
In my data, crawling started about three and a half weeks after launch: 22 single-page datacenter sessions rendered the pages between June 13 and 23. A human arrived from an AI engine referral somewhere in the weeks after that; my analytics can only date the click to a 30-day window. Actual quoted citations inside AI answers are harder to observe from the outside, and I have not verified one yet.
Should a small startup build comparison pages?
Yes, but only after you have at least one channel where humans already find you. The pages are cheap: one template, one evening per batch, static HTML with zero maintenance cost. They are retrieval infrastructure, not distribution. If nobody links to them and no engine has indexed you, they will sit unread, exactly as mine did for 23 straight days.
How much does Capi cost?
Capi is free for up to 30 transactions a month in any currency. Capi Core is US$9.90 a month or US$69.90 a year and removes the cap, adds statement uploads and insights. Capi Together, the couples plan, is US$99 a year for the whole household. Every plan handles multiple currencies natively, converted into the main currency you pick.